<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Naeemah Small on Medium]]></title>
        <description><![CDATA[Stories by Naeemah Small on Medium]]></description>
        <link>https://medium.com/@naeemahaz?source=rss-e449fc8bdc50------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*g--2I19FI-69oHEp_lpziQ.jpeg</url>
            <title>Stories by Naeemah Small on Medium</title>
            <link>https://medium.com/@naeemahaz?source=rss-e449fc8bdc50------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 13:38:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@naeemahaz/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-5e40a45de757?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e40a45de757</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fiction]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[fraud-detection]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 02:00:06 GMT</pubDate>
            <atom:updated>2026-11-12T02:00:06.003Z</atom:updated>
            <content:encoded><![CDATA[<h3>The People v. Walsh: Three Doors, Forty-One Dead Neurons, and a Token Everybody Knew About — Episode 8</h3><blockquote>The shadow model had 41 dead neurons out of 192. The quant’s original prototype had zero. Dead ReLUs are a fingerprint. They tell you <em>who trained it</em> — and who didn’t. The prosecutor can prove who built the scheme. She can’t charge him with the death — the felony-murder statute doesn’t reach wire fraud. So she needs the hands. The man who ran the wires that night says he left at 11:39 and doesn’t know what a learning rate is. One of those is a lie.</blockquote><blockquote><strong>Previously</strong></blockquote><blockquote>Log loss showed ORACLE’s beliefs severed from outcomes on the west side. An F1 score showed the fraud detector blind to exactly the fraud the shadow model wrote. Walsh’s “helpful” directory listing was missing one file. Kane said: <em>Nathan said it was a pilot.</em> Walsh was arrested for wire fraud and laundering — and told the detective that at 11:40 p.m. he was on stage, on a livestream, timestamped. <a href="https://medium.com/@naeemahaz/the-oracle-files-ef10bcff9a1f">https://medium.com/@naeemahaz/the-oracle-files-ef10bcff9a1f</a></blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1245278245%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1245278245%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2224344508-641f04e6d71807b7d78d1cb7c5230f44086a67dd671cdf31bd1e8c3b0de53be7-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/79b9159a1ca32c02c914095a04ae8a76/href">https://medium.com/media/79b9159a1ca32c02c914095a04ae8a76/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GAyGv7SHrvwfM7HrSGLKhQ.png" /></figure><p>Harlan Pike watches the Desert Chamber livestream in his office at 6 a.m. with the sound off, because he does not need to hear his client’s speech about financial inclusion; he needs the clock in the corner.</p><p><em>11:39:12.</em> Walsh walks to the podium. <em>11:41:07.</em> He takes the award. <em>11:52:40.</em> He steps down, shakes the hand of a city councilman, and is swallowed by a crowd of two hundred people in cocktail dress, all of whom can be subpoenaed and will say the same thing.</p><p>Pike pauses it on Walsh’s face, mid-smile.</p><p>“He gave the bank his token for month-end wires,” he tells the associate taking notes. “Every month, for a year. Everybody at Meridian knew. So somebody was in that server room with Nathan Walsh’s token at 11:38 while Nathan Walsh was on a stage three miles away.”</p><p>He un-pauses. The smile moves.</p><p>“Find me who. Find me the ops guy. The prosecution has a fraud. We’re going to give them a killer — and it isn’t going to be the man who was holding a trophy.”</p><h3>Day 6 — Celeste Varga</h3><p>Assistant County Attorney Celeste Varga has never lost a wire-fraud case and has never tried a murder, and she tells me both of those things in the first four minutes, in the order she thinks I need them.</p><p>“The fraud is clean,” she says. Her conference room has one wall of evidence and no windows. “Clean enough that Pike will try to plead it down to something with a fine. The homicide is the problem. I don’t have a person in the room. I have a token.”</p><p>The console log — ops-console-01, the air-gapped machine in the corner of Hall C that releases wires — shows a session opened with Nathan Walsh&#39;s token at <strong>11:38 p.m.</strong> It released five wires between 11:38 and 11:39. It stayed open until 11:47. Dana badged into Hall C at 11:40. Her rebuild was killed at 11:45:10 from that same console session — which means the session was <em>idle</em> for six minutes with a woman in the room, and then it wasn&#39;t.</p><p>“Walsh was on a stage at 11:39,” Varga says. “Pike has the clock. So either the token was shared — which Pike will say — or Walsh is a magician.”</p><p>“The token was shared,” I say. “Wrona told me. Month-end. ‘Everybody knows.’”</p><p>“Then I need to know what Tomasz Wrona was doing between 11:38 and 11:47. And before I can ask him under oath, I need to show the grand jury the shadow model was <em>built on purpose</em> and <em>built by someone</em> — because Pike’s whole theory is going to be that Walsh was a busy executive who inherited a mess.” She points at the wall. “Walk me through the machine. All of it. If I can’t explain it to twelve people in a jury room, it didn’t happen.”</p><p>I take her through the three doors.</p><h3>Evidence Locker 16 — Softmax</h3><h4>The equation</h4><p>I load scorer_v2.pt on an air-gapped laptop — you don&#39;t run somebody else&#39;s weights on a machine with a network card — and trace one declined west-side application through it. At the last layer the network produces three raw numbers, <strong>logits</strong>, one per door:</p><pre>decline: 0.1     standard: 1.0     premium: 2.0</pre><p>They aren’t probabilities. <strong>Softmax</strong> makes them probabilities — the multi-class cousin of Episode 2’s sigmoid.</p><pre>softmax(z)ᵢ = e^(zᵢ) / Σⱼ e^(zⱼ)</pre><h4>The number</h4><pre>e^0.1 = 1.105   e^1.0 = 2.718   e^2.0 = 7.389   sum = 11.212<br>P(decline) = 0.099   P(standard) = 0.242   P(premium) = 0.659</pre><p>A <strong>66% “premium” verdict</strong> for an applicant ORACLE had just declined at 0.62. Double the logits — “temperature” — and it becomes 86%. The machine was built to be decisive.</p><h4>What it proves</h4><p>I find the file the shadow was trained on, in the same undocumented bucket: labels_final_v3_REAL.csv, which is how you know a human made it. The &quot;premium&quot; label was assigned <em>by hand</em> to 1,200 applications. Every one of the 1,200 later wired a program fee to Nevada. The network wasn&#39;t taught to recognize creditworthiness. It was taught to recognize <em>the kind of applicant who would cooperate.</em></p><p>Kane’s branch originated 700 of those 1,200 rows. Someone <strong>above</strong> the branch picked the other 500.</p><p>“Who can write to that bucket?” Varga asks.</p><p>“Theo. Wrona’s service account. Walsh.”</p><p>“Three names.” She writes them on the wall. “Pike will say the first two. Give me a reason it’s the third.”</p><h4>Try it</h4><pre>import numpy as np<br>z = np.array([0.1,1.0,2.0]); p = np.exp(z-z.max()); p /= p.sum()<br>print(np.round(p,3))   # [0.099 0.242 0.659]</pre><h3>Evidence Locker 17 — ReLU</h3><h4>The equation</h4><p>Between the input and the three doors, scorer_v2.pt has three hidden layers of 64 units. Every unit computes a weighted sum, then passes it through the activation function that made deep learning possible:</p><pre>ReLU(x) = max(0, x)</pre><p>Negative in, zero out. Positive in, unchanged. Its derivative is 0 or 1 — so a gradient passing through ten ReLU layers is multiplied by 1, where ten sigmoid layers would multiply it by at most 0.2⁵¹⁰ ≈ 10⁻⁶ and kill it. That one fact is why networks went from three layers deep to a hundred.</p><h4>The number — and the fingerprint</h4><p>ReLU has a failure mode. A unit pushed far enough negative during training outputs zero for <em>every</em> input, and because its derivative is zero there, it can never recover. A <strong>dead ReLU.</strong></p><p>I histogram the activations of all 192 hidden units across 10,000 applications. <strong>Forty-one</strong> units output zero for every single input. Dead.</p><p>Forty-one out of 192 is sloppy training. I run the same histogram on Theo’s weekend prototype from his personal repo: <strong>zero</strong> dead units. He used a leaky variant and a learning-rate warm-up — the habits of someone who trains carefully because he was taught to.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jF7NL-PiMoyeGFigwiNIyg.png" /></figure><h4>What it proves</h4><p>Whoever took Theo’s architecture and retrained it on 1,200 hand-labelled “cooperators” did not have Theo’s habits. They had a deadline, a laptop, and no one looking. Theo <em>built</em> this network. Theo didn’t <em>train</em> this one.</p><p>“That rules Theo out of the training,” Varga says. “It doesn’t rule Wrona in.”</p><p>“No. But Wrona’s an ops engineer. He builds pipelines, not models. The person who retrained this knew <em>enough</em> machine learning to do it and <em>not enough</em> to do it well.” I look at the wall. “Walsh was an engineer before he was a CTO. He has a conference talk about training neural nets from 2016. It’s not very good.”</p><p>Varga writes <em>dead units → not Theo</em> on the wall, and <em>not an ops engineer either</em> under it, and underlines the second line twice.</p><h4>Try it</h4><pre>x = np.array([-2.0,-0.5,0.0,0.7,3.0])<br>print(np.maximum(0,x), (x&gt;0).astype(float))</pre><h3>Evidence Locker 18 — Entropy</h3><h4>The equation</h4><p>How predictable are the wires out of RK Capital? <strong>Shannon entropy</strong> measures the surprise in a probability distribution, in bits — the average number of yes/no questions you’d need to guess an outcome.</p><pre>H(P) = −Σ pᵢ · log₂(pᵢ)</pre><h4>The number</h4><p>If wires were spread evenly over the four accounts, P = [¼, ¼, ¼, ¼] and H = <strong>2.00 bits.</strong> Two questions per wire.</p><p>The actual split, over fourteen months: P = [0.70, 0.10, 0.10, 0.10].</p><pre>H = −(0.70·log₂0.70 + 3·0.10·log₂0.10) = 1.357 bits</pre><h4>What it proves</h4><p>The scheme <em>leaks.</em> Guess “account one” and you’re right seven times in ten. Money that is truly laundered looks like noise. This looks like a <em>preference</em> — and a preference belongs to a person. It will matter, in the next episode, that this exact split appears in a spreadsheet on Nathan Walsh’s drive, solved as an optimization problem, two days before the threshold was moved.</p><p>Varga tapes the entropy number to the wall next to the spreadsheet. “That,” she says, “is a man’s handwriting on a wire.”</p><h4>Try it</h4><pre>H = lambda p: -np.sum(p*np.log2(p))<br>print(round(H(np.array([.7,.1,.1,.1])),3))   # 1.357</pre><h3>Day 7 — the deposition</h3><p>Theo Lindqvist, under oath, in a room with a court reporter and a glass of water he doesn’t touch:</p><p>“Five months ago I saw the DAG. nightly_reconcile. I recognized my architecture in the job&#39;s config. I went to Nathan.&quot; He breathes. &quot;He said the prototype was in my personal repo, with my name on every commit, and if anyone ever looked, the author of the model was going to be <em>me.</em> He said he was protecting me by keeping it quiet. He said Marisol was running a pilot and I should let the grown-ups handle it.&quot;</p><p>“And you did.”</p><p>“I did.” He finally picks up the water. “Dana asked me in July if I’d ever built anything for the west side. I said no.”</p><p>Varga lets that sit on the record for exactly as long as it deserves.</p><p>Theo is not charged with anything. He is fired that afternoon, by a board that needs a name to fire, and reinstated nine days later, by the same board, after Varga’s office sends a letter.</p><h3>Day 7, 4:00 p.m. — the pharmacy</h3><p>The Nevada filing for RK Capital Partners was paid with a prepaid debit card. The card was purchased, sixteen months ago, at a pharmacy on Camelback Road in Arcadia, four blocks from Nathan Walsh’s house. The pharmacy keeps ninety days of video. The card issuer keeps the activation record forever, and the activation was done from an IP address that Walsh’s home internet provider has, under subpoena, confirmed was his.</p><p>“M. Kane,” Varga says, “typed from Nathan Walsh’s living room.”</p><p>The fraud is now, in her words, <em>airtight.</em> Wire fraud. Laundering. Identity theft — Kane’s. Tampering — the directory listing. Obstruction.</p><p>“And the murder?” Reyes says, from the doorway.</p><p>“The felony-murder statute lists its predicate felonies,” Varga says. “Wire fraud isn’t one of them. I spent a weekend trying to make it one. I can’t. If someone killed Dana Okafor during the commission of this fraud, I need <em>that person.</em> Not the architect. The hands.”</p><h3>Day 8 — the hands</h3><p>Reyes interviews Tomasz Wrona for the second time. I watch through the glass.</p><p>“I ran the wires at 11:38,” he says. His hoodie is grey; it has the Meridian logo on the chest. “Nathan gives me his token for month-end. Everybody knows. He was at the dinner. I released five wires, I closed the ledger, and I left.”</p><p>“The hall door has you in at 11:12,” Reyes says. “Nobody else after her. Badge-out?”</p><p>“The hall reader only logs <em>in.</em> Everybody knows that too. You put in a ticket, nobody fixes it.” He looks at the table. “I left the cage at 11:39. I was out the hall door before she was in it. I didn’t see her.”</p><p>“The console session stayed open until 11:47.”</p><p>“I don’t always lock it. It’s air-gapped. There’s a badge door.” A pause. “I should have.”</p><p>“Did you set the learning rate on her rebuild to 2.5, Mr. Wrona?”</p><p>“I don’t know what a learning rate is.”</p><p>Reyes writes that down. Through the glass, I write down that it is the first sentence he has said that I’m sure is a lie — because every ops engineer who has ever paged a training job knows exactly what a learning rate is; it’s the thing that makes the job fail at 3 a.m.</p><p>That evening Reyes gets the Medical Examiner’s final report. Blunt-force trauma, consistent with the edge of a server rack. But also: the angle. The wound is on the back of the head, <em>high</em>, and the body was found face-up. The ME’s note, in the careful language of someone who testifies for a living: <em>injury pattern consistent with a backward fall, possibly while turning; the decedent’s badge lanyard was found broken; fibers recovered from beneath the fingernails of the right hand are grey cotton, consistent with sweatshirt fleece.</em></p><p>Reyes reads it twice and puts it down.</p><p>“Grey cotton,” she says.</p><p>“Half of Meridian’s ops team wears that hoodie,” I say. “It’s swag. They gave out two hundred at the holiday party.”</p><p>“I know.” She looks through the glass at the empty interview room, at the chair Tomasz Wrona sat in. “That’s exactly the kind of thing a man would know, too, when he picked what to wear.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/796/1*dVAwImWRYwLoXdPdB1F4Ng.png" /></figure><p><strong>Episode 9:</strong> The Widest Line: The CTO Solved His Own Money Laundering as a Calculus Problem. A Lagrangian in a spreadsheet. A support vector with the victim’s last name. And a console log the investigation never pulled — read into the record by the defense. He didn’t guess how to split the money. Minimize detection cost subject to the total adding up. Lagrange multipliers. Solution: 70/10/10/10. The wires followed it for fourteen months to within a percentage point. He labelled λ “risk per $” in the margin.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1246089225%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1246089225%3Fshare%3Dcopy%26fl%3Dsv%26fe%3Dci&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2225393357-c46fa718dda1b1c463f0a7f82faf3252a9ee61627ee55745b9a158163fb5ab30-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/e0fba12748fc0939ee84dce24a55d37c/href">https://medium.com/media/e0fba12748fc0939ee84dce24a55d37c/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e40a45de757" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-ef10bcff9a1f?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef10bcff9a1f</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[fiction]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 00:01:11 GMT</pubDate>
            <atom:updated>2026-11-10T00:01:11.994Z</atom:updated>
            <content:encoded><![CDATA[<h3>Ninety-Six Percent: The Fraud Detector That Was Accurate at Nothing -Episode 7</h3><blockquote>Our fraud detection performs at 96% accuracy.” The compliance officer read that sentence to the board for eight quarters. Nobody asked: 96% accurate <em>at what?</em> Accuracy flattered the board for two years. One SQL query and an F1 score ended it. Then the CTO was arrested — and produced an alibi. Log loss · The F1 score <strong>Case file:</strong> 25–0313-H — Day 5</blockquote><blockquote><strong>Previously</strong></blockquote><blockquote>Multiple regression isolated the signature: $6,667 extra per west-side loan at equal income. The paradox — a stricter threshold producing bigger loans — had an answer: a shadow model, scorer_v2.pt, run every night at 2:10 a.m. by a job owned by Tomasz Wrona, in a bucket only Theo, Wrona&#39;s service account, and Nathan Walsh can write to. Reyes disclosed her conflict in writing on day one. Wrona says he&#39;s never opened the job. He says he&#39;s always there. <a href="https://medium.com/@naeemahaz/the-oracle-files-9226bd04685d">https://medium.com/@naeemahaz/the-oracle-files-9226bd04685d</a></blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1245275604%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1245275604%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2224340550-4573ddb84f6f153965c295637ec834fbfbd36903d2ccc1c08bd699bb2733abee-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/35bb72296f82b1959a23abc48f15cf14/href">https://medium.com/media/35bb72296f82b1959a23abc48f15cf14/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l3zqeU-vygFiM94j-XCNzg.png" /></figure><p>Gerald Fenn, Chief Compliance Officer of Meridian Trust, has read the same sentence to the board eight quarters in a row, and he reads it again at 8:00 a.m. on the fifth day, because the deck was finalized before anyone died and nobody has told him to change it.</p><p>“Our third-party fraud detection platform continues to perform at ninety-six percent accuracy across the retail lending book.”</p><p>Vivian Strand nods. The general counsel writes it down. Nathan Walsh, at the far end of the table, looks at his phone, then looks up, then says, “Gerald, could we get the breakdown by branch for the record? Given the week.”</p><p>Gerald Fenn has never been asked for the breakdown by branch. He does not have it. He says he’ll have it by end of day.</p><p>He never asked the vendor what the ninety-six percent was ninety-six percent <em>at.</em> It did not occur to him that there was a second question.</p><h3>9:30 a.m. — the zipper</h3><p>I load the last 400 west-side applications that passed through ORACLE <em>and then</em> through the shadow, and line each one up against what actually happened: did the borrower repay?</p><p>If ORACLE’s probabilities are honest, they should track reality. A loan it rated 90% should pay back about nine times in ten. If something downstream has been overriding its outputs — approving what it declined, inflating what it limited — then its confidence and the outcomes will have come apart like a zipper.</p><p>Accuracy won’t show this. Accuracy only sees right or wrong. I need the metric that sees <em>how sure the model was</em> at the moment it was wrong. Because that’s where tampering lives: in the gap between a belief and a result.</p><h3>Evidence Locker 14 — Log loss</h3><h4><strong>The equation</strong></h4><p><strong>Log loss</strong> (binary cross-entropy) charges a model, for each prediction, the negative log of the probability it assigned to what actually happened.</p><pre>LogLoss = −(1/n) Σ [ yᵢ·ln(pᵢ) + (1 − yᵢ)·ln(1 − pᵢ) ]</pre><p>Hedge and be wrong, you pay a little. Be <em>certain</em> and be wrong, you pay enormously. It is a proper scoring rule: the only way to minimize it is to tell the truth about your own uncertainty. It is also exactly what logistic regression minimizes while training — Episode 1’s gradient descent was walking down this surface the whole time.</p><h4>The number</h4><p>Four applications, truth y and prediction p: (1, 0.90), (1, 0.80), (0, 0.30), (1, 0.60).</p><pre>ln 0.90 + ln 0.80 + ln 0.70 + ln 0.60 = −1.196<br>LogLoss = 1.196 / 4 ≈ 0.299</pre><p>Now suppose the fourth borrower had <em>defaulted</em> but the model said 0.99. That one term becomes ln(0.01) = −4.6, and the average jumps from 0.30 to about <strong>1.32.</strong> A single confident mistake costs more than a dozen hedged ones.</p><p>Reference point: a coin flip for everything scores ln 2 ≈ 0.693. Anything above that is worse than guessing.</p><p>ORACLE’s log loss on Central applications: <strong>0.31.</strong> On the west-side applications that went through the shadow: <strong>0.91.</strong></p><h4>What it proves</h4><p>Worse than a coin flip. ORACLE’s beliefs and the outcomes have been pulled apart. The model isn’t broken — on Central it’s excellent. Something <em>between</em> its prediction and the disbursement is changing the answer, and the shape of the damage says the change is not random. It is in one direction, in one place.</p><h4>Try it</h4><pre>import numpy as np<br>y = np.array([1,1,0,1]); p = np.clip(np.array([.9,.8,.3,.6]), 1e-15, 1-1e-15)<br>print(round(float(-np.mean(y*np.log(p)+(1-y)*np.log(1-p))), 3))   # 0.299</pre><h4>11:00 a.m. — at what</h4><p>Meridian has a fraud detector. Vendor product, quarterly license, a dashboard with a big green tile that says <strong>96% accuracy.</strong> Gerald Fenn has read that tile to the board for two years. Nobody asked what it was 96% accurate <em>at.</em></p><p>I pull the confusion matrix for the last 1,000 west-side loans. It takes a SQL query and four minutes. Nobody had run it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/672/1*sus741-dx8TorTuMfDKW2w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8Wc49uEWl33tsqlT7AgIlA.png" /></figure><h3>Evidence Locker 15 — The F1 score</h3><h4>The equation</h4><p><strong>Precision:</strong> of the loans flagged as fraud, how many were. <strong>Recall:</strong> of the actual fraud, how much was flagged. <strong>F1:</strong> their harmonic mean — high only when <em>both</em> are.</p><pre>Precision = TP / (TP + FP)<br>Recall    = TP / (TP + FN)<br>F1        = 2·P·R / (P + R)</pre><h4>The number</h4><pre>Precision = 40 / 50 = 0.800<br>Recall    = 40 / 70 = 0.571<br>F1        = 2 · 0.800 · 0.571 / 1.371 ≈ 0.667</pre><p>The detector misses <strong>43% of fraud.</strong> That is the number the board should have heard. Ninety-six percent was mostly the 920 honest loans it did nothing to.</p><p>Why the harmonic mean? Take a detector with precision 1.0 and recall 0.02 — it flags one case, correctly, and nothing else. The arithmetic mean says 0.51. F1 says 0.039. The harmonic mean refuses to let one good number hide one terrible one.</p><h4>What it proves</h4><p>I look at the thirty misses. Every one of them is a west-side loan re-scored to “premium” by scorer_v2.pt. The shadow isn&#39;t just inflating loans. It is generating exactly the fraud the detector cannot see — because the detector was trained on data from before the shadow existed, and has never been retrained, because the tile was green.</p><h4>Try it</h4><pre>TP,FP,FN = 40,10,30; P=TP/(TP+FP); R=TP/(TP+FN)<br>print(round(2*P*R/(P+R),3))   # 0.667</pre><h3>1:15 p.m. — the small thing</h3><p>Then I see the small thing. The kind of thing I’d have missed on the first night, before I learned to distrust anything that’s <em>handed</em> to me.</p><p>Walsh sent Reyes a full listing of the storage bucket on day one, unasked — <em>I hope this helps</em> — from his personal address, the same as the badge printout. I put it beside the live listing I pulled under warrant.</p><p>His copy has every file but one.</p><p>It doesn’t have scorer_v2.pt.</p><p>He didn’t forget it. You don’t forget the one file that is the whole crime. He <em>edited a directory listing</em> and sent it to a homicide detective as a favor, and he did it in Helvetica.</p><p>Reyes reads the two listings side by side for a long time. “He’s been helping,” she says, “the way a man helps you look for your keys when he’s got them in his pocket.”</p><p>Her warrant covers every shared drive at Meridian. I search Walsh’s for one word: <em>allocation.</em></p><p>One hit. A spreadsheet. optimal_allocation.xlsx. Last modified fourteen months ago, two days before the threshold commit. One sheet. Four rows. A formula bar that reads like a confession written in Greek letters.</p><p>I don’t understand it yet. I will, in Episode 9. For now I understand this: it is a spreadsheet about how to split a sum of money across four accounts, and the split in its bold row is 70% to the first and 10% to each of the others, and I have seen that split before — in the wires out of RK Capital Partners.</p><h3>3:00 p.m. — the lender, again</h3><p>Kane receives us at the branch a second time. The fig tree is still alive. She is not.</p><p>Reyes puts two pages on the desk: the $6,667 — the number Kane asked for — and the directory listing with the missing file, Walsh’s name at the top of the email.</p><p>Kane reads both without moving anything but her eyes.</p><p>“Nathan said it was a pilot,” she says finally. Her voice has gone somewhere flatter. “A premium tier for under-banked ZIPs. The board had voted to close us. He said this would save the branch. Good PR. He asked me to send him applicants who’d <em>cooperate</em> — his word — with a partner lender in Nevada. He said Theo built the scoring. He said the model was <em>fairer</em> than ORACLE.”</p><p>“And the program fee?”</p><p>“He said it was the partner’s cost of capital.” She looks at the window. “I believed that for ten months. Then I saw where it went. I didn’t go to anyone. I wrote Dana a memo without my name on it and told her to stop, because I thought I was protecting her from a <em>career</em> problem. I deleted a letter to her the morning after she died. You can have my laptop. The letter’s in the recovery sector. I didn’t know how to delete it properly. I’m a lender.”</p><p>“The Nevada filing,” Reyes says. “M. Kane.”</p><p>“I have never been to Henderson,” Kane says. “You know who has access to my signature on four hundred loan documents. He didn’t need me to type my name. He needed it to be <em>typed.</em>”</p><p>She doesn’t ask for a lawyer. She asks whether Dana’s family has been told, and when Reyes says yes, she nods like a person signing something.</p><h3>6:40 p.m. — the arrest</h3><p>Reyes doesn’t arrest Nathan Walsh at Meridian. She arrests him in the parking lot of a restaurant in Arcadia, in front of his wife, because she wants him off balance and because, she tells me later, “he’s had four days of choosing the room. I wanted one.”</p><p>Wire fraud. Conspiracy. Money laundering. Tampering with evidence. Obstruction. She reads them in a voice I’ve never heard from her — not flat, almost tender — and he listens with the face of a man hearing a slightly disappointing quarterly result.</p><p>“I’d like to call my attorney,” he says.</p><p>“You’d like to call Harlan Pike,” Reyes says. “I know. He’s already called <em>me.</em>”</p><p>And then, because she can’t not: “Where were you at 11:40 p.m. on the thirteenth, Mr. Walsh?”</p><p>Walsh looks at her with something close to pity.</p><p>“On stage,” he says. “At the Desert Chamber gala. Accepting the Innovation in Financial Inclusion award.” He actually smiles. “Two hundred witnesses, Detective. A photographer. A livestream. I gave a speech about the west side. It’s on the Chamber’s website. It’s <em>timestamped.</em>”</p><p>The cuffs go on anyway. But I watch Reyes’s face when he says it, and I see the thing she’s been building for five days — the straight line from the shadow model to the man in Hall C at 11:40 — bend.</p><p>Walsh was in the fraud. Walsh was not in the room.</p><p>So who was?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/670/1*IY8Kj14QiosiKopFgo62mw.png" /></figure><p><strong>Episode 8: </strong>The shadow model had 41 dead neurons out of 192. The quant’s original prototype had zero. Dead ReLUs are a fingerprint. They tell you <em>who trained it</em> — and who didn’t.</p><p>The prosecutor can prove who built the scheme. She can’t charge him with the death — the felony-murder statute doesn’t reach wire fraud. So she needs the hands. The man who ran the wires that night says he left at 11:39 and doesn’t know what a learning rate is. One of those is a lie.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1245278245%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1245278245%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2224344508-641f04e6d71807b7d78d1cb7c5230f44086a67dd671cdf31bd1e8c3b0de53be7-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/79b9159a1ca32c02c914095a04ae8a76/href">https://medium.com/media/79b9159a1ca32c02c914095a04ae8a76/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef10bcff9a1f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-9226bd04685d?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/9226bd04685d</guid>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[fiction]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 19:21:09 GMT</pubDate>
            <atom:updated>2026-11-02T19:21:09.785Z</atom:updated>
            <content:encoded><![CDATA[<h3>Holding Everything Else Still: The $6,667 That Only Exists Between Two Models — Episode 6</h3><blockquote>At equal income, a west-side applicant walked out with $6,667 more than a central one. That number does not exist in the data. It exists in the <em>difference</em> between two regressions. It’s the same math a salary-equity audit uses. It’s the same math a fair-lending examiner uses. Is detective’s mother is in the fraud? The threshold was raised to make loans <em>harder</em> — so why did they get bigger? There’s a second model. It runs every night. And the man who keeps it running says he’s never opened it.R² · Multiple linear regression <strong>Case file:</strong> 25–0313-H — <strong>Day 4</strong></blockquote><blockquote><strong>Previously</strong></blockquote><blockquote>The regression line through the loan book fits 94% of the story and is wrong in exactly one place: the west side. Kane denied the Nevada company and told the auditor to ask Walsh what the board voted. Walsh told the auditor before being asked. And row 212 of the 341 “premium” loans is a refinance on Buckeye Road in the name of <em>Reyes, L. </em><a href="https://medium.com/@naeemahaz/the-oracle-files-de3410a35b8a">https://medium.com/@naeemahaz/the-oracle-files-de3410a35b8a</a></blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1242503362%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1242503362%3Fshare%3Dcopy%26fl%3Dsv%26fe%3Dci&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2220770986-4b3345ed9b480347f2140315b1d1758e36bf3289f5a77ce0debe75fe89204c4d-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/25e406106817a06ef24e6bdfd0b62186/href">https://medium.com/media/25e406106817a06ef24e6bdfd0b62186/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h9em_JBh2MFC3LE4IL1T7g.png" /></figure><p>Tomasz Wrona makes his daughter’s lunch at 5:10 a.m. because the mornings are the only part of the day he is certain he will be home for.</p><p>Peanut butter. The crusts off. A note, because she’s seven and she likes the notes, and because he has been told by a counselor at her school that routines matter for children who have moved countries twice.</p><p>On the refrigerator, under a magnet shaped like a saguaro, is a letter he has read so many times the fold has gone soft. <em>Notice of Receipt — Form I-140, Immigrant Petition for Alien Worker. Petitioner: Meridian Trust Bank. Authorized signatory: Nathan Walsh, Chief Technology Officer.</em></p><p>His phone lights up on the counter. <em>N. Walsh.</em></p><blockquote>You good? Take tomorrow off if you need. Family first.</blockquote><p>He types <em>I’m ok.</em> He looks at it. He deletes <em>I’m</em> and sends <em>ok</em>, and puts the phone face down, and finishes the sandwich, and draws a very small saguaro on the note.</p><h4>7:40 a.m. — two coffees</h4><p>“There’s something I need to tell you about the loan book.”</p><p>Reyes puts the coffee in front of me and sits, and I realize I have been holding my laptop closed with one hand the way you hold a door.</p><p>“Row 212,” I say.</p><p>She doesn’t flinch. “Row 212. My mother. Eight months ago a man with a Meridian lanyard came to the shop and told her she’d been <em>re-scored</em> for a premium refinance. She was so proud she called me from the parking lot. I told her to read everything twice. She wired the fee.”</p><p>“When did you find it?”</p><p>“The first night. The 4 a.m. loan book. I searched for her name before I did anything else, because I’m not stupid, and because I knew what I’d find.” She slides a folded sheet across the table. A memo on Phoenix PD letterhead, dated the morning after Dana died, 6:15 a.m. <em>Disclosure of potential conflict — Det. I. Reyes — Case 25–0313-H.</em> Her lieutenant’s initials at the bottom. A second detective — <em>Sgt. R. Delgado</em> — assigned to review every warrant she requests.</p><p>“I’m on this case because I asked to be, and I’m still on it because I put this on paper before I called you. Delgado reads everything. Including this conversation, probably.” She picks up her coffee. “And yes. I didn’t call the judge yesterday after the Nevada filing. Marisol Kane saved my mother’s shop in 2009. I owed her one conversation before a warrant. That’s the whole of it. You can decide what it makes me.”</p><p>I decide, for now, that it makes her the only person in this case who has told me something true before I asked.</p><p>“Okay,” I say.</p><p>“Okay.” She nods at the laptop. “Now. Kane wants a number. Give her one she can’t fold her hands at.”</p><h4>Evidence Locker 12 — R²</h4><p><strong>The equation</strong></p><p><strong>R²</strong> is the fraction of the variance in y that the model explains, compared to just guessing the average.</p><pre>R² = 1 − SS_res / SS_tot</pre><p>SS_tot is how wrong you’d be with no model at all — the spread around the mean. SS_res is how wrong you are with the line.</p><p><strong>The number</strong></p><p>From Episode 5: SS_res = 17.6. SS_tot — the squared deviations of the five loans from their mean of 23.6 — sums to 309.2.</p><pre>R² = 1 − 17.6 / 309.2 = 0.943</pre><p>Income explains <strong>94.3%</strong> of the variance in loan size. (For a single input, R² is just the square of Episode 3’s correlation. The pieces of this toolkit keep turning out to be each other.)</p><p><strong>What it does not prove</strong></p><p>Ninety-four percent is the number that goes in the executive summary. It is also the number that lulls people into not looking at the residuals. A high R² with a <em>patterned</em> residual is a model that is right on average and wrong in a specific place.</p><p>That specific place is where the crime lives.</p><p><strong>Try it</strong></p><pre>print(1 - ((y-yhat)**2).sum() / ((y-y.mean())**2).sum())   # 0.943</pre><h4>9:15 a.m. — the second column</h4><p>The residual pattern wasn’t noise. Central under the line, West over it. Income isn’t the only thing driving loan size. The <em>branch</em> is. So I add it to the model, and I shut up, and I let the ledger talk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6kC7z5W_T8zYf4-WKeifnw.png" /></figure><h3>Evidence Locker 13 — Multiple linear regression</h3><p><strong>The equation</strong></p><p>Add inputs; the line becomes a plane.</p><pre>y = β₀ + β₁·x₁ + β₂·x₂ + … + ε</pre><p>Each βⱼ is the effect of its input <strong>holding every other input fixed.</strong> That phrase — <em>all else equal</em> — is the entire power of the method. It is how a salary-equity audit isolates a gender gap. It is how a fair-lending examiner isolates a ZIP code. It is the mathematical form of the word <em>fairness.</em></p><p>Same solver as OLS. X just has more columns.</p><p><strong>The number</strong></p><p>Add a column: x₂ = 1 for West, 0 for Central. Solve the normal equations (by computer; the 3×3 inverse is tedious, not hard):</p><pre>β₀ = −0.80      β₁ = 0.34      β₂ = 6.67</pre><p><strong>Loan ≈ −0.8 + 0.34·income + 6.67·West.</strong></p><ul><li>Each $1,000 of income now adds <strong>$340</strong> — it was $540 when income was secretly doing the branch’s work too. Omit a real driver and the others absorb it. That is how an honest-looking slope lies.</li><li>Being on the West side adds <strong>$6,667</strong> to the loan — <em>at the same income.</em></li></ul><p><strong>The fit: </strong>MSE drops from 3.52 to 0.85. R² rises from 0.943 to <strong>0.986</strong>. And the residuals no longer have a pattern. The ledger has nothing left to say, because it has said it.</p><p><strong>Scale it up. </strong>$6,667 per loan, times the 341 loans in the Episode 1 shoulder: roughly <strong>$2.3 million</strong> of loan principal that Meridian’s own policy would never have approved. Fifteen percent of it wired to Nevada: about $340,000 in “program fees” on those 341 alone. There are 1,200 premium loans in total.</p><h4>What it proves</h4><p>Kane’s number. At equal income, a west-side applicant walks out with about $6,700 more than a Central one — after accounting for everything Kane said to account for. That is not local risk adjustment. Local risk adjustment makes loans <em>smaller.</em></p><p><strong>Try it</strong></p><pre>X = np.column_stack([np.ones(5), [40,50,60,70,80], [0,0,1,1,1]])<br>print(np.round(np.linalg.lstsq(X, y, rcond=None)[0], 3))   # [-0.8  0.34  6.667]</pre><p><strong>What goes wrong, so you can defend it</strong></p><p>Multicollinearity (two inputs that move together — check the VIF). Omitted variables (leave out branch, income lies). Interactions (maybe income matters <em>more</em> out west — add income·west). The dummy trap. Overfitting. Each has a name because each has been used by a defense expert to make a regression disappear. Know them before Harlan Pike does.</p><h4>2:00 p.m. — the paradox</h4><p>I show Reyes the two equations side by side. The line with one input. The plane with two. The $6,667 that only appears when you hold income still — money that exists only in the <em>difference</em> between two models, which is the strangest place I have ever found money.</p><p>“That’s the skim,” she says. “But ORACLE was supposed to <em>stop</em> this. The west-side threshold was <em>raised.</em> Zero point six-two. That makes it <em>harder</em> to get a loan out there. Not easier.”</p><p>“I know.”</p><p>“So how does a stricter model produce bigger loans?”</p><p>We both look at the ledger. Three hundred and twelve pages, and the one thing it’s telling us doesn’t make sense — which in my experience means it’s telling us something else.</p><p>“Unless,” Reyes says slowly, “there’s a second model.”</p><p>I’ve been awake fifty-six hours. When she says it, I get cold in a way that has nothing to do with Hall C.</p><h4>4:30 p.m. — what calls what</h4><p>The second model isn’t in the repository. It isn’t in the model registry either, the one Walsh showed a conference slide about — <em>full lineage, every artifact versioned.</em> Everything is versioned except the thing that matters. I’ve stopped being surprised by that. I’ve started counting on it.</p><p>I find it the way I find most things that were hidden badly: by looking at what calls what.</p><p>An Airflow DAG named <strong>nightly_reconcile</strong> — a name chosen to be boring, which is a tell in itself — runs at <strong>2:10 a.m.</strong> It loads a file called <strong>scorer_v2.pt</strong> from a storage bucket nobody documented, with an access policy nobody reviewed. It has been running for fourteen months. Four hundred and twenty-some nights, at 2:10 a.m., while the dashboard stayed green.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y98MTkYNksxT2JnHR-U13Q.png" /></figure><p>And a quiet rule in the wrapper: any west-side application ORACLE declines at 0.62 gets <em>re-scored</em> by scorer_v2.pt, and if the second model says &quot;premium,&quot; the loan goes out at the premium ceiling — $88,000 to $92,000.</p><p>That’s the shoulder on the histogram. ORACLE was made <em>stricter</em> so the shadow could be <em>generous.</em> The scar was manufactured one loan at a time, at 2:10 a.m., by a cron job.</p><p>I pull the access policy on the bucket. Three principals can write to it.</p><p><strong>tlindqvist</strong>.<strong> </strong><strong>svc-ops-reconcile</strong> — a service account. Owner of record: <strong>T. Wrona.</strong> <strong>nwalsh</strong>.</p><p>“Who’s Wrona?” Reyes says.</p><p>“The man in the hoodie,” I say. “From the first night. In the doorway.”</p><p>She writes it down. It’s the first time his name has been written down in this case, four days in, and I think about that — about how a man can be in the room where a body is found and not get a line in anyone’s notebook for four days, because he was holding a coffee cup and looking at the floor and nobody introduced him.</p><h4>2:10 a.m. — Day 5</h4><p>I go to Hall C at 2:10 a.m. because that is when the shadow runs, and I want to see it run.</p><p>The cage door is open. The console at the end of the row — <strong>ops-console-01</strong>, the air-gapped machine that signed the session which killed Dana&#39;s rebuild — is lit. Tomasz Wrona is sitting in front of it with his hoodie up and a thermos, watching a progress bar.</p><p>He doesn’t startle. He’s heard the badge reader. He looks at me the way he looked at the floor on the first night: like a man who has learned that looking at things makes them yours.</p><p>“It runs every night,” he says. “I just make sure it finishes.”</p><p>“What is it?”</p><p>“Reconcile job.” He shrugs inside the hoodie. “Nathan’s. He set it up before I had the account. I inherited it. If it fails, I get paged. So I come in.” He nods at the progress bar: <strong>rescore_west — 412 / 540</strong>. &quot;I don&#39;t know what it does. I know what it <em>needs.</em> GPU, the bucket, forty minutes.&quot;</p><p>“You’ve never opened it.”</p><p>“I’m ops.” He says it without self-pity, as a job description. “I don’t open things. I keep them running.”</p><p>The progress bar finishes. <strong>write_premium_approvals — success.</strong> Somewhere, a letter with a gold seal is being queued for a printer.</p><p>“Mr. Wrona,” I say. “The night Dana died. Were you here?”</p><p>He is quiet for exactly one breath too long. I know that pause. I heard it on the phone at 1:52 a.m. on the first night.</p><p>“I’m always here,” he says.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/816/1*RkAfLbbbU2kgiUF40xs4aQ.png" /></figure><p><strong>Episode 7 — </strong>Our fraud detection performs at 96% accuracy.” The compliance officer read that sentence to the board for eight quarters. Nobody asked: 96% accurate <em>at what? </em>One confusion matrix. F1 = 0.667. It misses 43% of fraud — and every miss is a loan the shadow model wrote. <strong>The CTO is arrested in a restaurant parking lot.</strong> And then he tells the detective where he was at 11:40. On stage. On a livestream. Timestamped.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9226bd04685d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-de3410a35b8a?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/de3410a35b8a</guid>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[fiction]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 22:38:17 GMT</pubDate>
            <atom:updated>2026-11-02T19:22:55.202Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Ledger: Drawing the Honest Line Through 14,000 Loans — Episode 5</h3><blockquote>A regression line fits 94% of the story. The crime lives in the other 6% — and the detective’s mother is in it. Linear regression · Ordinary least squares · Mean squared error <strong>Case file:</strong> 25–0313-H — Day 3</blockquote><p><strong>Previously</strong></p><blockquote>Naive Bayes found a phrase — <em>program fee</em> — and a company: RK Capital Partners LLC, Henderson, Nevada. Dana’s brother had been re-scored from a $40,000 decline to a $91,000 “premium” loan, with 15% wired to Nevada. The night Dana died, five wires left RK Capital when the model says one. And the manager of record on the Nevada filing is <em>M. Kane. </em><a href="https://medium.com/@naeemahaz/the-oracle-files-c35519a04f56">https://medium.com/@naeemahaz/the-oracle-files-c35519a04f56</a></blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1240481831%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1240481831%3Ffl%3Dtl%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2218289712-af687fdbeaf949f0c058631e5b3f15f00cb3fe3c98b0bb7345552257c8342ed2-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/95af88a4e42a9bfb699efbed7146463b/href">https://medium.com/media/95af88a4e42a9bfb699efbed7146463b/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T6HZUESJv3qWExDLiY0Tfg.png" /></figure><p>Vivian Strand, chair of the board of Meridian Trust, has an omelet at the Biltmore every Thursday at 7 a.m. with the bank’s general counsel, and has not missed one in nine years, and does not intend to start because an analyst has died.</p><p>“The examiners are here in six weeks,” she says. “I want the west-side exit on the agenda regardless.”</p><p>The general counsel, who is a careful man, says that the timing may look unfortunate.</p><p>“The timing is what it is. We voted fourteen months ago. Four ZIP codes, eighteen months to profitability or we close the branch. Nathan said he had a plan to hit the number.” She cuts the omelet into precise squares. “Nathan’s plan was his business. The number was mine. The number got hit.”</p><p>She looks out at the golf course, which is green in a way nothing in Phoenix is green without someone paying for it.</p><p>“If the number got hit the wrong way,” she says, “I want to know before the examiners do. And I want the branch closed either way.”</p><h4>9:30 a.m. — 312 pages</h4><p>The loan book prints to 312 pages. Reyes prints it anyway. She’s the kind of detective who wants to hold the thing, and honestly, so am I; a PDF can be edited by anyone with a Friday afternoon, and paper has to be <em>burned.</em></p><p>I spread five rows on the table under a light that buzzes at 60 hertz — five loans, chosen because they’re typical, in thousands of dollars.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/820/1*rZAa5rNNN7iGonkaH102Ng.png" /></figure><p>“So richer people get bigger loans,” Reyes says. “That’s not a crime.”</p><p>“No. But <em>how much</em> bigger is a policy. A policy has a number. And if the number’s different on the west side, that’s not a crime either.” I pick up the pencil. “It’s a signature.”</p><p>I’m going to draw a line through this ledger. The ledger is going to tell me where it was bent. And you — you’re going to check my arithmetic, because I’ve been awake for fifty hours and that’s when I start believing things.</p><h4>Evidence Locker 9 — Linear regression</h4><p><strong>The equation</strong></p><p><strong>Simple linear regression</strong> fits a straight line through data: one input, one output, plus noise.</p><pre>y = β₀ + β₁·x + ε</pre><p>β₁ is the slope — how much loan you get per thousand dollars of income. β₀ is the intercept. ε is everything the line can’t explain. The slope has a closed form: the covariance of x and y over the variance of x.</p><pre>β₁ = Σ(xᵢ − x̄)(yᵢ − ȳ) / Σ(xᵢ − x̄)²        β₀ = ȳ − β₁·x̄</pre><p><strong>The number</strong></p><p>Means: x̄ = 60, ȳ = 23.6. The products of the deviations sum to 540; the squared x-deviations sum to 1,000.</p><pre>β₁ = 540 / 1000 = 0.54<br>β₀ = 23.6 − 0.54·60 = −8.8</pre><p><strong>Loan ≈ −8.8 + 0.54 × income.</strong> Every extra $1,000 of income buys $540 more loan.</p><p>Now the part nobody looks at. Predict each loan from the line and subtract:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*BHjiEdhbNNmV9KQ9LoOugg.png" /></figure><p>The line fits. But look at the <em>signs.</em> Both Central loans sit below the line. Two of the three West loans sit well above it. A single line is averaging two different policies.</p><p><strong>What it proves</strong></p><p>Not yet what Kane asked for. But it has told me <em>where to look</em>: not at the line, at what the line leaves behind.</p><p><strong>Try it</strong></p><pre>import numpy as np<br>x = np.array([40,50,60,70,80.]); y = np.array([12,17,26,31,32.])<br>b1 = ((x-x.mean())*(y-y.mean())).sum()/((x-x.mean())**2).sum()<br>print(b1, y.mean()-b1*x.mean())   # 0.54 -8.8</pre><h4>11:00 a.m. — why that line</h4><p>Reyes asks the question a defense attorney is going to ask, in the same flat voice she’d use for it. “Why <em>that</em> line? You could draw a hundred lines through five points.”</p><p>She’s right. I need a principle that picks exactly one, and a way of computing it that survives fourteen thousand rows and forty features without a pencil — because the pencil version is for her, and the matrix version is for the jury’s expert, and they have to agree to the third decimal.</p><h4>Evidence Locker 10 — Ordinary least squares</h4><p><strong>The equation</strong></p><p><strong>OLS</strong> picks the line that minimizes the sum of squared residuals. Squared, not absolute: squaring makes the objective smooth, punishes big misses far more than small ones, and — the deep reason — makes OLS the <strong>maximum-likelihood estimate</strong> when errors are normal. Episode 4’s MLE, meeting Episode 1’s bell.</p><p>In matrix form, with X holding a column of ones and the features:</p><pre>β̂ = (XᵀX)⁻¹ Xᵀ y</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/778/1*wiZ26AOTNQegR_kthTsUZQ.png" /></figure><pre>β₀ = (19000·118 − 300·7620) / 5000 = −8.8<br>β₁ = (−300·118 + 5·7620)   / 5000 =  0.54</pre><p>Same line as the pencil. Now derived from a principle, in a form that handles forty features as easily as one. And under the Gauss–Markov theorem it is the <em>best linear unbiased estimator</em> — no other linear method has lower variance. That is the sentence you say to the defense attorney.</p><p><strong>What it proves</strong></p><p>That I didn’t draw the line that told the story I wanted. The mathematics drew it. There was only one.</p><p><strong>Try it</strong></p><pre>X = np.column_stack([np.ones(5), x])<br>print(np.linalg.solve(X.T@X, X.T@y))   # [-8.8  0.54]</pre><h4>1:30 p.m. — the lender</h4><p>Marisol Kane agrees to talk at her branch, not the station. Fifty, immaculate, a lobby with a fig tree that’s actually alive. She receives us the way she must receive examiners: coffee poured, door open, hands folded, nothing on the desk but a legal pad she never writes on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C8iApu-hlQlViqrdHJ-FSw.png" /></figure><p>“Every lender in this state adjusts for local risk,” she says. “The west side has different collateral, different employers, different everything. If you’re going to tell me my loans are bigger, you show me by <em>how much</em> after you account for that.” Her eyes go to me, not Reyes. She knows which of us signed the audit. “Show me a number. Or don’t come back.”</p><p>Reyes lets that sit, then puts the Nevada filing on the desk between them, with the fig tree framed in the doorway behind it.</p><p>“RK Capital Partners,” Reyes says. “Manager of record: M. Kane.”</p><p>Kane looks at it for exactly as long as it takes to read it. Something happens in her face that I can’t name — it isn’t surprise and it isn’t fear, it’s something older, like recognition.</p><p>“I have never heard of that company,” she says. “I have never been to Henderson, Nevada. I have been to Las Vegas once, for a conference, in 2016, and I left early.” She folds her hands again. “You will find that’s true. You will also find that proving it is going to take you longer than it took someone to type my name into a form.”</p><p>“Who would type your name into a form, Ms. Kane?”</p><p>A long pause. The fig tree does nothing.</p><p>“Ask Nathan Walsh,” she says, “what the board voted fourteen months ago. And then ask him what he told me the pilot was <em>for.</em>”</p><p>She stands. The interview is over, and she has given us exactly one sentence, and I will spend the rest of the day realizing it was the most expensive sentence anyone has said to me yet.</p><h4>Evidence Locker 11 — Mean squared error</h4><p><strong>The equation</strong></p><p>I have a line. Reyes wants to know if it’s any good. “Any good” <em>is</em> a number; it just needs a name.</p><pre>MSE = (1/n) Σ (yᵢ − ŷᵢ)²          RMSE = √MSE</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/740/1*T7Oh4sHTNKvHG12degbcfQ.png" /></figure><pre>MSE  = 17.6 / 5 = 3.52<br>RMSE = √3.52 ≈ 1.88</pre><p>The line is typically off by about <strong>$1,880</strong> per loan. Against loans of $12,000–$32,000, that’s decent — and yet three of the five errors come from West, and the biggest ones are all in the same direction.</p><p><strong>What it proves</strong></p><p>That the line is <em>right on average and wrong in a specific place.</em> Which is exactly what a crime looks like when it’s been averaged.</p><p><strong>Why squared</strong></p><p>Big misses cost disproportionately — a $10k error costs 100× a $1k error. The function is smooth, so gradient descent can walk down it. And under normal errors, minimizing it is maximum likelihood. One formula, three justifications, no shrug.</p><p><strong>Try it</strong></p><pre>yhat = -8.8 + 0.54*x<br>print(((y-yhat)**2).mean()**0.5)   # ≈ 1.88</pre><h4>6:45 p.m. — the shoulder</h4><p>Walsh is at the station when we get back. He has brought dinner, because of course he has. He is sitting in the waiting area with a bag from a place that does good sandwiches, and he stands when he sees Reyes, and for a moment he looks exactly like what he says he is: a man whose employee died on his watch, trying to be useful.</p><p>“Vivian’s asking for an update,” he says. “The board. I told her I’d only tell her what you’ve cleared me to say.”</p><p>“Tell her nothing,” Reyes says, and takes a sandwich, and goes into the interview room, and does not invite him in.</p><p>He turns to me. Lowers his voice.</p><p>“The board voted to shut the west-side branch,” he says. “Fourteen months ago. Four ZIP codes, eighteen months to turn a profit or they close it. I fought it. I pitched a premium pilot — a way to lend <em>more</em> out there, responsibly, with better scoring. Marisol ran it. If something went wrong inside it — “ He stops. Rubs his face. “I should have watched it closer. That’s on me. I know that’s on me.”</p><p>He is telling me, before I ask, the exact thing Kane told me to ask him. He is telling it as a confession of negligence, which is the kind of confession a careful man makes when he’d like you to stop looking for a different one.</p><p>“Thank you,” I say.</p><p>“Anything,” he says. And then, gently: “You look like you should go home.”</p><h4>11:20 p.m. — the thing I find alone</h4><p>I don’t go home.</p><p>I pull the 341 loans in the $88,000–$92,000 shoulder — the scar from Episode 1 — and lay them next to the Nevada wires. Every one of them is a “premium” re-score. Every one has a program-fee wire. Every one is west side.</p><p>I sort them by date.</p><p>Eight months ago, row 212: a refinance. A small business. $89,500. Premium tier. Program fee wired to RK Capital Partners, $13,425.</p><p>The business address is on Buckeye Road.</p><p>The borrower’s name is <strong>Reyes, L.</strong></p><p>I sit with that for a long time. I think about a detective who has a judge on speed dial and a court order at 4 a.m. and a mother with a shop that Meridian’s west side saved once already. I think about how she told me that story unprompted, in an interview room, the way Walsh tells me things unprompted. I think about the phone call she didn’t make this afternoon, with her thumb over a judge’s name, after Kane’s initial showed up on a Nevada form.</p><p>I think about the fact that she brought me into this case knowing my name was on the audit.</p><p>And I think about the fact that there is a loan with <em>her</em> name on it in the exact 341 rows that are the crime — and that she has had this ledger, printed, in her hands, for two days.</p><p>I don’t call her. I don’t call anyone. Just like she said.</p><p>At 7:40 the next morning Reyes walks into the interview room with two coffees, puts one in front of me, sits down, and says:</p><p>“There’s something I need to tell you about the loan book.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/1*9sc16GY_goMLW5e7kYRN_Q.png" /></figure><p><strong>Episode 6: </strong>Holding Everything Else Still: The $6,667 That Only Exists Between Two Models. At equal income, a west-side applicant walked out with $6,667 more than a central one. That number does not exist in the data. It exists in the <em>difference</em> between two regressions. It’s the same math a salary-equity audit uses. It’s the same math a fair-lending examiner uses. Is detective’s mother is in the fraud, and she put it on paper before anyone asked. The threshold was raised to make loans <em>harder</em> — so why did they get bigger? <a href="https://medium.com/@naeemahaz/the-oracle-files-9226bd04685d">https://medium.com/@naeemahaz/the-oracle-files-9226bd04685d</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1242503362%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1242503362%3Fshare%3Dcopy%26fl%3Dsv%26fe%3Dci&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2220770986-4b3345ed9b480347f2140315b1d1758e36bf3289f5a77ce0debe75fe89204c4d-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/25e406106817a06ef24e6bdfd0b62186/href">https://medium.com/media/25e406106817a06ef24e6bdfd0b62186/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de3410a35b8a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Electric Car Predictions 2027: Will EVs Get Cheaper? A Simple Machine-Learning Forecast]]></title>
            <link>https://medium.com/@naeemahaz/electric-car-predictions-2027-will-evs-get-cheaper-a-simple-machine-learning-forecast-635d6976b631?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/635d6976b631</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 05:58:46 GMT</pubDate>
            <atom:updated>2026-10-28T05:58:46.985Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*A9brViNyyLdnjgUpG78fqQ.png" /></figure><blockquote>What could happen to electric car prices, batteries, charging, and sales next year? Follow a relatable story, explore a real forecasting experiment, and see the Python code behind the predictions.</blockquote><p><strong>Datasets are located:</strong> <a href="https://www.kaggle.com/datasets/subzeroheart/electric-car-predictions-2027-prices-and-charging"><em>https://www.kaggle.com/datasets/subzeroheart/electric-car-predictions-2027-prices-and-charging</em></a></p><p>Imagine walking into a dealership a year from now.</p><p>You are not looking for a car that feels like a spaceship. You want something comfortable, dependable, and affordable.</p><p>An electric car catches your eye.</p><p>The price looks reasonable. The interior feels familiar. Then the questions begin.</p><p>Where will I charge it? Will it save me money? Should I buy now or wait for better batteries?</p><p>Those questions — not just futuristic technology — will help shape the electric car industry in 2027.</p><p>My central prediction is <strong>practicality</strong>: the next stage of electric car competition will increasingly focus on making ownership easier to understand and easier to afford.</p><p>To explore that future, this article combines industry research with a small machine-learning experiment. The experiment projects approximately 28 million global electric car sales in 2027. However, that number is an exploratory trend projection, not a guarantee or an official industry forecast.</p><p>Let’s follow the story behind it — and walk through the actual code.</p><h3>Meet Maya: The Buyer Behind the Numbers</h3><p>Maya’s car needs another expensive repair.</p><p>She commutes to work, takes her children to school, and occasionally drives several hours to visit family. She is interested in an electric car, but she does not consider herself a technology person.</p><p>At the dealership, a salesperson starts explaining battery chemistry and charging power.</p><p>Maya interrupts.</p><p>“Can you tell me what owning it would actually be like?”</p><p>That is the right question.</p><p>Maya is fictional, but her situation gives us a useful way to examine the industry. Instead of asking only what technology is possible, we will ask what could change for someone buying and using a car.</p><h3>First, What Are We Predicting?</h3><p>This article looks ahead from October 2026 to calendar year 2027.</p><p>The numerical forecast covers global sales of two categories:</p><ul><li>Battery-electric vehicles, or BEVs, which run entirely on battery power.</li><li>Plug-in hybrid electric vehicles, or PHEVs, which combine a rechargeable battery with a combustion engine.</li></ul><p>The IEA’s electric car sales figures include both categories. Conventional hybrids that cannot be plugged in are outside this forecasting exercise. Consequently, “28 million electric cars” does not mean “28 million cars that never use gasoline.”iea+1</p><p>There is already considerable momentum. The International Energy Agency expects approximately 23 million electric cars to be sold globally in 2026, representing 28% of new car sales.iea</p><p>The question is what happens next.</p><h3>How Machine Learning Helps — and Where It Stops</h3><p>Think of a forecasting model as a person watching a moving bicycle.</p><p>By observing its previous position and direction, that person can estimate where it might go next.</p><p>But the rider could brake. A road could close. A storm could arrive.</p><p>A sales forecasting model has the same limitation: it can identify a historical pattern without knowing what will interrupt it.</p><p>For this article, we use two simple regression models:</p><ol><li>Linear regression, which finds a straight-line trend through past sales.</li><li>Ridge regression, which fits a similar relationship but penalizes larger coefficients.</li></ol><p>These are supervised-learning methods: we provide known inputs and outcomes, then ask the model to learn a relationship.</p><p>Our input is the year. Our outcome is annual electric car sales.</p><p>This is intentionally simple. The model does not directly understand battery prices, government policies, or charging access.</p><h4>Step 1: Gather the Historical Sales Data</h4><p>The experiment uses six rounded observations from IEA reporting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*Abd41-9ST-2KSFIkbYt0fQ.png" /></figure><p>The underlying reports describe approximately 3 million sales in 2020, 6.6 million in 2021, more than 10 million in 2022, nearly 14 million in 2023, more than 17 million in 2024, and around 21 million in 2025. Rounding makes the example easy to reproduce but reduces precision.iea+4</p><p>We deliberately do not enter the 2026 forecast as though it were an observed full-year result.</p><h4>Step 2: Prepare the Python Environment</h4><p>Readers who only want the industry predictions can skip the code and continue to the results.</p><p>Readers who want to reproduce the experiment need Python with NumPy, pandas, and scikit-learn installed.</p><p>The installation command is:</p><pre>pip install numpy pandas scikit-learn</pre><h4>Step 3: Run the Complete Forecasting Code</h4><pre>from pathlib import Path<br><br>import numpy as np<br>import pandas as pd<br>from sklearn.linear_model import LinearRegression, Ridge<br><br><br># 1. Enter rounded historical IEA sales figures.<br># Units: millions of cars.<br># Includes battery-electric cars and plug-in hybrids.<br>years = np.array([2020, 2021, 2022, 2023, 2024, 2025])<br><br>sales = np.array([<br>    3.0,<br>    6.6,<br>    10.0,<br>    14.0,<br>    17.0,<br>    21.0,<br>])<br><br># 2. Use years since 2020 as the input feature.<br># This keeps the input easier to interpret.<br>X = (years - 2020).reshape(-1, 1)<br><br># 3. Define two candidate models.<br># Ridge uses a fixed penalty; this example does not tune it.<br>model_factories = {<br>    &quot;Linear regression&quot;: lambda: LinearRegression(),<br>    &quot;Ridge regression&quot;: lambda: Ridge(alpha=1.0),<br>}<br><br># 4. Test on later years using only earlier observations.<br>#<br># Train 2020–2022 -&gt; predict 2023<br># Train 2020–2023 -&gt; predict 2024<br># Train 2020–2024 -&gt; predict 2025<br>test_results = []<br><br>for test_index in range(3, len(years)):<br>    X_train = X[:test_index]<br>    y_train = sales[:test_index]<br><br>    X_test = X[test_index:test_index + 1]<br>    actual = float(sales[test_index])<br><br>    for model_name, make_model in model_factories.items():<br>        model = make_model()<br>        model.fit(X_train, y_train)<br><br>        prediction = float(model.predict(X_test)[0])<br><br>        test_results.append({<br>            &quot;year&quot;: int(years[test_index]),<br>            &quot;model&quot;: model_name,<br>            &quot;actual_millions&quot;: actual,<br>            &quot;prediction_millions&quot;: prediction,<br>            &quot;absolute_error_millions&quot;: abs(<br>                actual - prediction<br>            ),<br>        })<br><br>backtest = pd.DataFrame(test_results)<br><br># 5. Compare mean absolute error.<br># Lower means smaller average misses in these tests.<br>scores = (<br>    backtest<br>    .groupby(&quot;model&quot;)[&quot;absolute_error_millions&quot;]<br>    .mean()<br>    .sort_values()<br>)<br><br>selected_model_name = scores.index[0]<br><br># 6. Refit the selected model using all historical data.<br>selected_model = model_factories[selected_model_name]()<br>selected_model.fit(X, sales)<br><br># 7. Predict calendar-year sales for 2026 and 2027.<br>future_years = np.array([2026, 2027])<br>future_X = (future_years - 2020).reshape(-1, 1)<br><br>predictions = selected_model.predict(future_X)<br><br>forecast = pd.DataFrame({<br>    &quot;year&quot;: future_years,<br>    &quot;forecast_millions&quot;: predictions,<br>})<br><br># 8. Compare the 2026 model projection with the IEA outlook.<br># This is a forecast comparison, not a comparison with<br># observed full-year 2026 sales.<br>iea_2026_outlook = 23.0<br><br>model_2026 = float(<br>    forecast.loc[<br>        forecast[&quot;year&quot;] == 2026,<br>        &quot;forecast_millions&quot;<br>    ].iloc[0]<br>)<br><br>difference = model_2026 - iea_2026_outlook<br><br># 9. Save reproducible tables.<br>output_dir = Path(&quot;ev_forecast_output&quot;)<br>output_dir.mkdir(exist_ok=True)<br><br>training_data = pd.DataFrame({<br>    &quot;year&quot;: years,<br>    &quot;sales_millions_rounded&quot;: sales,<br>})<br><br>training_data.to_csv(<br>    output_dir / &quot;training_data.csv&quot;,<br>    index=False,<br>)<br><br>backtest.to_csv(<br>    output_dir / &quot;backtest_results.csv&quot;,<br>    index=False,<br>)<br><br>forecast.to_csv(<br>    output_dir / &quot;sales_forecast.csv&quot;,<br>    index=False,<br>)<br><br># 10. Show the results.<br>print(&quot;\nBACKTEST RESULTS&quot;)<br>print(backtest.round(3).to_string(index=False))<br><br>print(&quot;\nAVERAGE ABSOLUTE ERROR, MILLIONS&quot;)<br>print(scores.round(3).to_string())<br><br>print(f&quot;\nSelected model: {selected_model_name}&quot;)<br><br>print(&quot;\nEXPERIMENTAL FORECAST&quot;)<br>print(forecast.round(2).to_string(index=False))<br><br>print(<br>    &quot;\n2026 model projection minus IEA outlook: &quot;<br>    f&quot;{difference:.2f} million cars&quot;<br>)<br><br>print(<br>    &quot;\nCaution: six rounded annual observations are &quot;<br>    &quot;not enough for a dependable industry forecast.&quot;<br>)</pre><p><strong>What the Code Does in Everyday Language</strong></p><p>The script follows a straightforward sequence:</p><ol><li>Give the computer the historical sales figures.</li><li>Ask two models to learn the relationship between time and sales.</li><li>Test each model on later years it did not see during training.</li><li>Compare how far each prediction missed.</li><li>Choose the better-performing model in those tests.</li><li>Refit it using the full historical series.</li><li>Extend its trend into 2026 and 2027.</li></ol><p>The testing order matters. We do not shuffle years randomly, because a forecast of the future should not be trained on information from that future.</p><h4>Step 4: Check Whether the Models Performed Well</h4><p>The rolling tests produced these results:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/840/1*WnoPfv5PGaXDf24PwNXvYw.png" /></figure><p>The average absolute errors were approximately:</p><ul><li>Linear regression: 0.41 million cars.</li><li>Ridge regression: 1.68 million cars.</li></ul><p>Linear regression performed better in this particular comparison.</p><p>That does not establish that linear regression is generally the best way to forecast electric car sales. It only tells us which of these two configurations performed better on three tests using this rounded dataset.</p><p>There is another limitation: we used the same small collection of tests to select the model and describe its performance. We did not retain an additional untouched test period.</p><h4>Step 5: Read the Forecast Without Being Misled</h4><p>After training on all six observations, the selected model produced:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/1*qhmt-ou34RvDdtoi3YzeHw.png" /></figure><p>In simple terms, the line adds about 3.58 million annual sales for every additional year.</p><p>However, the model’s 2026 projection is around 1.45 million higher than the IEA’s approximately 23 million outlook. That is an important warning: the straight-line trend may be too optimistic for the near term.iea</p><p>The 2027 result is also two years beyond the last observed annual input, not a one-step forecast from completed 2026 data.</p><p>The responsible takeaway is:</p><blockquote>The historical series supports an expansionary trend, but 28 million sales is an exploratory projection — not a reliable promise.</blockquote><p>The model does not generate the battery, price, or charging predictions below. Those are qualitative judgments informed by industry reporting.</p><h4>Prediction 1: The Fight for Affordable Electric Cars Will Intensify</h4><p>Back at the dealership, Maya asks about the monthly payment.</p><p>She is not interested in paying extra for acceleration she will rarely use.</p><p>My prediction is that more competition in 2027 will revolve around affordable models and practical configurations rather than premium specifications alone.</p><p>The IEA reports that declining battery manufacturing costs and more affordable models have improved EV price competitiveness over the past decade. However, larger vehicles and bigger batteries have limited how much of those savings reaches buyers.</p><p><strong>Why Battery Savings Do Not Automatically Reach Your Wallet</strong></p><p>Imagine a bakery gets cheaper flour.</p><p>It could lower bread prices. It could make larger loaves. Or it could keep prices unchanged.</p><p>Automakers face a similar choice when battery costs fall.</p><p>A less expensive battery can support a cheaper car, a larger battery, or a different profit margin. Consequently, falling battery costs do not guarantee that every model’s sticker price will fall.</p><p>Regional differences will remain important. In China, the IEA reports that 70% of battery-electric cars sold in 2025 were cheaper than the average conventional car. That does not mean comparable affordability has already arrived in every market or vehicle category.iea</p><p>For Maya, progress may mean something modest but meaningful: an electric car that finally fits the shortlist.</p><h4>Prediction 2: Established Battery Technology Will Matter More Than Breakthrough Headlines</h4><p>Maya reads a headline about a revolutionary battery.</p><p>Should she wait?</p><p>My prediction is that technologies already manufactured at scale will have more influence on ordinary buyers in 2027 than highly publicized breakthroughs.</p><p>One important example is lithium iron phosphate, commonly called LFP. It accounted for more than 55% of globally deployed EV batteries in 2025, according to the IEA.iea</p><p>The IEA also reports that LFP packs were, on average, more than 40% cheaper per kilowatt-hour than nickel manganese cobalt alternatives in 2025. However, differences in applications and energy-density requirements partly influence that comparison; it does not establish an equivalent saving for every comparable vehicle.iea</p><p><strong>What About Solid-State Batteries?</strong></p><p>Solid-state batteries use a solid electrolyte rather than the liquid electrolyte found in conventional lithium-ion batteries.</p><p>Their potential attracts substantial interest, but the IEA says their advantages still need demonstration at scale under real-world conditions.iea</p><p>My forecast is therefore not that solid-state batteries will transform every dealership in 2027.</p><p>It is that battery headlines will keep moving faster than mass-market purchasing options.</p><p>For Maya, the useful question remains: what does the battery in an available car deliver for its price?</p><h4>Prediction 3: Charging Will Expand, but Convenience Will Remain Uneven</h4><p>Now imagine Maya taking a long trip.</p><p>A national announcement about thousands of new chargers is encouraging. But her experience depends on the charging options along her actual route.</p><p>The infrastructure is growing. Nearly 1.8 million public charging points were added worldwide in 2025, increasing the global stock by more than 33%, according to the IEA.</p><p>My prediction is continued improvement through 2027, but not equal convenience everywhere.</p><p><strong>A Simple Guide to Two Confusing Terms</strong></p><p>Think of a battery as a water tank:</p><ul><li>kWh describes how much energy the tank can hold.</li><li>kW describes how quickly energy flows into or out of it.</li></ul><p>A charger’s advertised power is not the same thing as the charging performance every vehicle can achieve. The IEA notes that cars capable of using chargers above 250 kW still represent less than 5% of the electric car stock.iea+1</p><p>The electricity network also matters. The IEA warns that grid-capacity constraints could become more pronounced in some regions as EV adoption and charging speeds increase.iea</p><p>The everyday test for 2027 will therefore be less about spectacular charging claims and more about whether normal trips become easier.</p><h4>Prediction 4: Where You Live Will Still Shape the Economics</h4><p>Maya lives in an apartment.</p><p>Her coworker has a driveway and a charger.</p><p>Even if they buy identical cars, they may face very different charging routines.</p><p>The IEA identifies home charging as the preferred option for drivers who can access it because of its relative convenience and affordability.iea</p><p>My prediction is that reliable charging at home — or another regularly available location — will remain a major dividing line in the ownership experience.</p><p><strong>Will an Electric Car Save Money?</strong></p><p>The answer depends on more than the vehicle’s badge.</p><p>The IEA reports annual refueling savings of approximately $550–$1,000 across major battery-electric markets since 2020, with differences between regions and years. However, it also finds that relying exclusively on public fast charging can produce higher running costs than a gasoline car under the markets and assumptions examined.</p><p>For Maya, that means three separate questions:</p><ol><li>What does the car cost to buy?</li><li>What will it cost to charge using her actual routine?</li><li>What will the complete ownership cost be?</li></ol><p>A lower energy bill alone does not settle the third question.</p><h4>Prediction 5: Global Growth Will Hide Very Different Local Stories</h4><p>A worldwide chart can make the EV transition look smooth.</p><p>Individual markets are not smooth.</p><p>The IEA reports that global electric car sales fell 8% year over year in the first quarter of 2026 following policy changes in China and the United States, even while its full-year outlook still anticipated growth.iea</p><p>That illustrates an important distinction: a temporary decline and a growing annual market can coexist.</p><p>My prediction is that 2027 will continue to produce sharply different regional headlines.</p><p>Charging availability and policy support are among the constraints the IEA identifies in its outlook.iea</p><p>One country may see growing interest while another experiences a slowdown.</p><p>This is also why our global model cannot tell us whether a particular manufacturer will succeed or whether a particular local dealership will sell more cars.</p><p><strong>What the Model Cannot Tell Us</strong></p><p>The code makes the process visible, but transparency is not the same as accuracy.</p><p>The experiment has significant limitations:</p><ul><li>Only six historical observations.</li><li>Rounded rather than full-precision sales inputs.</li><li>Only one explanatory variable: time.</li><li>No separate regional models.</li><li>No direct inputs for prices, financing, policies, or infrastructure.</li><li>Only three backtests.</li><li>No untouched final test period.</li><li>No validated prediction interval.</li></ul><p>A more substantial project would add monthly or quarterly data, distinguish BEVs from PHEVs, test regional differences, and compare multiple approaches against simple baselines.</p><p>Most importantly, it would keep asking whether the historical relationship still applies.</p><p>Machine learning is useful when it makes assumptions testable. It becomes misleading when a precise-looking output conceals weak evidence.</p><p><strong>The Real Electric Car Story of 2027</strong></p><p>Maya closes the brochure.</p><p>She does not need a perfect prediction of the automotive future.</p><p>She needs to understand whether an electric car fits her life.</p><p>The industry’s challenge is to answer that question convincingly.</p><p>The forecasting experiment suggests that recent historical growth could extend to roughly 28 million global electric car sales in 2027. But the model’s optimistic comparison with the IEA’s 2026 outlook is a reason for caution, not celebration.iea</p><p>The more durable story is about ordinary decisions: an affordable payment, an understandable charging routine, and a car that does what its owner needs.</p><p>The electric car industry’s next big breakthrough may not be making ownership feel futuristic.</p><p>It may be making it feel normal.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=635d6976b631" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-c35519a04f56?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/c35519a04f56</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[fiction]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[writing]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 01:24:08 GMT</pubDate>
            <atom:updated>2026-10-28T22:44:41.563Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Two Hundred Emails, One Verdict Each: How a 1960s Algorithm Found a Nevada Shell Company — Episode 4</strong></p><blockquote><strong>Evidence introduced:</strong> Naive Bayes read eleven thousand emails before lunch. Maximum likelihood said the night she died was a 0.4% event. And a government form in Carson City has her name on it. <strong>Case file:</strong> <strong>25–0313-H — Day 2, afternoon, into Day 3</strong></blockquote><blockquote><strong>Previously</strong></blockquote><blockquote>The correlation that pointed at Theo was month-end close. The cosine similarity put the anonymous memo in Marisol Kane’s hand. Theo admitted he built a “premium-tier scorer” for Kane fourteen months ago and never asked what happened to it. Reyes’s mother’s shop was saved by a loan Kane signed. And Nathan Walsh keeps helping. — <a href="https://medium.com/@naeemahaz/the-oracle-files-205d65cfed9e">https://medium.com/@naeemahaz/the-oracle-files-205d65cfed9e</a></blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1240035532%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1240035532%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2217744201-3352c7662f9aa5a088d3ab0ab0c9c1a9ee5552c4233e9cfb7a7b46868e874ba9-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/f3690998a8206e5fe9d3c7e855cdb4ad/href">https://medium.com/media/f3690998a8206e5fe9d3c7e855cdb4ad/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jEmrf6CyISq5F1Jsaf1R2Q.png" /></figure><p>Josefina Ortega has worked the teller line at Meridian’s west-side branch for four years and can tell a wire that’s going to be a problem from across the lobby.</p><p>Mr. Alvarado is seventy-one and has brought his loan folder in a plastic grocery bag, which is how you know it’s the most important thing he owns. Ninety thousand dollars, premium tier, approved last month. He’s very proud. He shows her the letter with the gold seal.</p><p>“I just need to send the program fee,” he says. “The fifteen percent. The partner lender. The gentleman said it has to go this week or the rate changes.”</p><p>Josefina reads the wire instructions he’s been given. Thirteen thousand five hundred dollars. Beneficiary: <em>RK Capital Partners LLC.</em> Henderson, Nevada. A routing number she’s never seen on a west-side wire.</p><p>“Mr. Alvarado,” she says carefully, “who is the gentleman?”</p><p>He doesn’t know his name. He had a Meridian lanyard. He came to the house.</p><p>She processes the wire, because that is her job, and because the man is standing there with his folder in a grocery bag and an interest rate that will change. Then she writes <em>RK Capital Partners</em> on a sticky note and puts it in her drawer, under the stapler, where she keeps things she is going to ask about and never does.</p><p>She has eleven of those sticky notes.</p><h3>4:30 p.m. — the mountain</h3><p>Kane’s writing matches the memo. But a cosine is a resemblance, not a case. Reyes wants every email in the company that touches the “restructure” flagged — thousands — before her meeting with the County Attorney’s office tomorrow at four, and she wants to be able to explain to a prosecutor <em>why</em> each one was flagged without saying the words <em>neural</em> or <em>embedding.</em></p><p>So: no transformer. No fine-tuned anything. I need a classifier that trains in seconds, explains itself line by line, and has been doing exactly this job since the 1960s. There’s one. It’s older than the internet and it still works.</p><p>We hand-label two hundred emails at a table with a bad fluorescent tube. Reyes reads. I type 1 or 0. Forty are about the scheme. A hundred and sixty are not. It takes two hours and I learn that Reyes reads out loud with the flat, unhurried voice of someone who has read a lot of other people’s mail.</p><h3>Evidence Locker 7 — Naive Bayes</h3><p><strong>The equation</strong></p><p>Bayes’ theorem, with one brave assumption: that the words in an email are independent of each other given what the email is about. (They aren’t. It works anyway.)</p><pre>P(class | words) ∝ P(class) · Π P(wordᵢ | class)</pre><ul><li><strong>P(class)</strong> — the <strong>prior.</strong> How common the class is before you read a word.</li><li><strong>P(wordᵢ | class)</strong> — the <strong>likelihood</strong> of each word, from counts.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/1*uVenW-trVUO1ZKWZA8SFeA.png" /></figure><p>An 83% verdict from two words and a prior. Over the whole corpus, with Laplace smoothing so an unseen word can’t zero out a product, and log-probabilities so a hundred small numbers don’t underflow, it takes four seconds to score eleven thousand emails.</p><p><strong>What it proves</strong></p><p>Two hundred and twelve emails flagged. Reyes reads every one. By midnight she has a phrase that repeats: <em>program fee.</em> And a destination that repeats: <strong>RK Capital Partners LLC, Henderson, Nevada.</strong></p><p><strong>Try it</strong></p><pre>from sklearn.naive_bayes import MultinomialNB<br>from sklearn.feature_extraction.text import CountVectorizer<br>X = CountVectorizer().fit_transform(emails)   # list of str<br>clf = MultinomialNB(alpha=1.0).fit(X, labels) # alpha=1 → Laplace smoothing</pre><p><strong>Why this, in front of a prosecutor</strong></p><p>Because every word’s contribution is visible. “This email was flagged because it contains <em>restructure</em>, which appears in 75% of scheme emails and 10% of clean ones.” A jury can follow that. A jury cannot follow a 768-dimensional embedding, and shouldn’t be asked to.</p><h3>Day 3, 7:00 a.m. — the fee</h3><p>The pattern, once you’ve seen it, is impossible to unsee.</p><p>A west-side applicant is declined by ORACLE — the 0.62 threshold. Days later, the same applicant is approved anyway, at $88,000 to $92,000, “premium tier.” Then a “partner lender” sends wire instructions: fifteen percent of the loan, as a <em>program fee</em>, to RK Capital Partners in Nevada. Thirteen thousand dollars, give or take, per loan.</p><p>Then the money moves again. Every few nights a wire leaves RK Capital for one of four accounts — never the same night twice, never a round number. Somebody read the same fraud-detection playbook I did, and used it as a checklist.</p><p>Reyes wants a rate. How often does this happen? Not a vibe. A number she can put in front of a prosecutor with the reasoning visible.</p><p>There is a principle for that. It underlies almost every estimate in statistics and every loss function in machine learning, and it fits on an index card.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AYM8Rfs2CXMXlVLRhmc5Pg.png" /></figure><h3>Evidence Locker 8 — Maximum likelihood estimation</h3><h4>The equation</h4><p><strong>Maximum likelihood</strong> picks the parameters that make what you actually observed <em>least surprising.</em> Write down the probability of the data as a function of the unknown; find the value that maximizes it.</p><pre>ℓ(θ) = Σ log f(xᵢ | θ)        θ̂ = argmax ℓ(θ)</pre><p><strong>Number one — the overrides</strong></p><p>Of 15 west-side applications that ORACLE declined and a human then overrode to “premium,” 12 were approved. Treat each as a coin flip with unknown probability p:</p><pre>ℓ(p) = 12 ln p + 3 ln(1 − p)<br>dℓ/dp = 12/p − 3/(1 − p) = 0  →  p̂ = 12/15 = 0.80</pre><p>The most believable override-approval rate is 80%. Of course it is — the MLE for a coin is the sample proportion. The point is you <em>derived</em> it, so when the defense asks why that number, you have a reason instead of a shrug.</p><p><strong>Number two — the wires</strong></p><p>Wires out of RK Capital per night, over eight nights: 0, 1, 0, 2, 1, 0, 3, 1. Rare events in fixed intervals → a Poisson process with rate λ.</p><pre>ℓ(λ) = (Σxᵢ) ln λ − nλ + const<br>dℓ/dλ = 0  →  λ̂ = Σxᵢ / n = 8/8 = 1.0 wire per night</pre><p>Under Poisson(1.0), the chance of three or more wires in a night is about 8%.</p><p>The night Dana died, there were <strong>five.</strong></p><p>Probability under the model: <strong>0.4%.</strong></p><p><strong>What it proves</strong></p><p>The night Dana Okafor walked into Hall C was not a normal night for RK Capital Partners. Someone was moving money — a month’s worth — in a hurry. And whoever releases those wires was, at 11:38 p.m., awake, logged in, and in a position to notice a badge swipe at 11:40.</p><p><strong>Try it</strong></p><pre>import numpy as np<br>from scipy.stats import poisson<br>w = np.array([0,1,0,2,1,0,3,1]); lam = w.mean()<br>print(lam, round(1 - poisson.cdf(4, lam), 4))   # 1.0, 0.0037</pre><p><strong>Why MLE is everywhere</strong></p><p>Every time you minimize a loss, you are almost certainly maximizing a likelihood. Least squares is MLE with normal errors. Log loss is MLE for a coin. Softmax cross-entropy is MLE for a die. The engine from Episode 1 — gradient descent — is usually walking down a negative log-likelihood. You’ve been doing this all along. Now you know its name.</p><h3>11:00 a.m. — Marcus</h3><p>Marcus Okafor comes to the station without being asked, with a manila folder he hasn’t opened since Reyes knocked on his door.</p><p>Inside: a loan offer. Meridian Trust. <em>Premium tier.</em> Ninety-one thousand dollars. A gold seal. A letter congratulating him. Dated three days before Dana died.</p><p>And a second sheet. Wire instructions. <em>Partner program fee — 15%.</em> RK Capital Partners LLC. Henderson, Nevada.</p><p>“She said don’t sign anything from Meridian this week,” he says. He’s not crying. He’s past it, into the flat place. “I’d already signed. I signed on the Tuesday. I was going to tell her Sunday.”</p><p>He puts the folder on the table and pushes it toward me with one finger, the way you push away food.</p><p>“I didn’t apply for a premium anything. I applied for forty thousand. For the truck. They turned me down. Then a guy came to the house.”</p><p>“A guy.”</p><p>“Meridian lanyard. Nice. Said my application had been <em>re-scored.</em> Said I qualified for more.” He finally looks up. “Was that her model? The thing she built? Did my sister’s bank re-score me?”</p><p>“No,” I say. “Not hers.”</p><p>I don’t say <em>not yet.</em> I don’t say <em>that’s the question.</em> I say the one true thing I have, which is that his sister found his name in the data, and that is why she was in that room.</p><p>He nods like he already knew that. He probably did. He turns his phone over in his hand, the one that was off, and doesn’t say anything, and neither do I.</p><h3>3:50 p.m. — Henderson, Nevada</h3><p>Reyes’s subpoena to the Nevada Secretary of State comes back in nine hours, which is fast, which means somebody in Carson City also has a judge they can call.</p><p>RK Capital Partners LLC. Formed sixteen months ago. Registered agent: a mailbox at a shipping store in Henderson, the kind of address that is a building with a hundred and forty businesses in one wall of brass boxes.</p><p>Manager of record, on the Articles of Organization:</p><p><strong>M. Kane.</strong></p><p>Reyes doesn’t say anything. She photographs it. She photographs everything; I’ve started to find it comforting. Then she picks up her phone, and I see her thumb hover over a contact, and I see her put the phone down without calling, and I understand that the contact was a judge and the call was a warrant and that she has just decided, for reasons she isn’t going to share with me, to wait.</p><p>“Her name,” I say.</p><p>“Her <em>initial</em>,” Reyes says. “And a surname. Typed into a form on a website by someone with a prepaid card.”</p><p>“Or by Marisol Kane.”</p><p>“Or by Marisol Kane.” She looks at the filing for a long time. “You know what I’ve learned in Financial Crimes? When a name shows up exactly where you’d expect it, there’s usually a reason it was put there. Sometimes the reason is that it’s true.”</p><p>She stands up.</p><p>“I’m going to go talk to her. At the branch. Where the fig tree is.” She picks up the memo — <em>stop pulling threads</em> — and folds it into her notebook. “And you’re going to bring me a number. Kane told me on the phone this morning, before I’d asked her anything, that every lender in this state adjusts for local risk. She said if I’m going to tell her the west-side loans are bigger, I’d better show her <em>by how much</em> after accounting for that.”</p><p>“That’s a dismissal,” I say.</p><p>“That’s a specification.” She’s already at the door. “They’re the same sentence, said by different people. Bring me the number.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/738/1*dUIJ-zW44dbm3su5lh-ZEA.png" /></figure><p><strong>Episode 5: </strong>The Ledger: Drawing the Honest Line Through 14,000 Loans -A regression line fits 94% of the story. The crime lives in the other 6% — and the detective’s mother is in it. <strong>Read:</strong> <a href="https://medium.com/@naeemahaz/the-oracle-files-de3410a35b8a">https://medium.com/@naeemahaz/the-oracle-files-de3410a35b8a</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1240481831%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1240481831%3Ffl%3Dtl%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2218289712-af687fdbeaf949f0c058631e5b3f15f00cb3fe3c98b0bb7345552257c8342ed2-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/95af88a4e42a9bfb699efbed7146463b/href">https://medium.com/media/95af88a4e42a9bfb699efbed7146463b/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c35519a04f56" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-205d65cfed9e?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/205d65cfed9e</guid>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 22:51:54 GMT</pubDate>
            <atom:updated>2026-10-20T22:51:54.855Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Together Is Not Because: The Correlation That Almost Convicted an Innocent Man — Episode 3</strong></p><blockquote>r = 0.84 says he was there every night. The calendar says so was everyone. And a memo with no name on it just got one.</blockquote><blockquote>Previously</blockquote><blockquote>Theo’s login signed both the threshold change and the session that killed Dana’s rebuild — but his credentials live in a vault half of Operations can read, and one edit happened three hours after he went home. Dana mailed herself a USB drive twelve minutes before she died. On it: an anonymous memo warning her to <em>stop pulling threads on the west-side restructure. </em><a href="https://medium.com/@naeemahaz/the-oracle-files-4f5f58b18843">https://medium.com/@naeemahaz/the-oracle-files-4f5f58b18843</a></blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1237783564%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1237783564%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2214681532-b451ca8621908720718cd530eac0c7fbcbd25951a89a9b5265fe149e6c8f0b3f-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/c542269ba76f498462abe3bb78ca7529/href">https://medium.com/media/c542269ba76f498462abe3bb78ca7529/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DTO84SvP9khcvJTSCkG0FQ.png" /></figure><p>Marisol Kane waters the fig tree in her branch lobby at 6:05 a.m., before the tellers arrive, because it has been alive for nine years and she would like it to stay that way.</p><p>The lobby is dark except for the ATM glow. Her phone is on the reception desk, face up, and the local news alert is on it: <em>Meridian Trust employee found dead at Camelback data center; police investigating.</em> No name yet. She knows the name.</p><p>She finishes watering the tree. She dries her hands on a paper towel. Then she opens her email, goes to Drafts, and finds the one she has rewritten eleven times in three weeks and never sent — the one that begins <em>Dana, I should have told you this in person</em> — and she deletes it, and then she deletes it from Trash, and then she stands with both hands flat on the desk for a while and looks at the fig tree.</p><p>When the first teller comes in at 7:15, Marisol Kane is at her desk with the door open and nothing on the blotter but a legal pad she never writes on.</p><h3>8:00 a.m. — the legal pad</h3><p>I lay the two columns side by side on a legal pad, because I want Reyes to watch me do it slowly. Five nights from Walsh’s printout. The minutes Theo’s badge says he was in the building. The number of config edits that night.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/1*kS3R336LiYLS04vVQ-WFlA.png" /></figure><p>When one goes up, the other tends to. Reyes sees guilt. I see a number that hasn’t been computed yet, which means it hasn’t earned the right to mean anything.</p><h3>Evidence Locker 5 — Correlation</h3><h3>The equation</h3><p><strong>Pearson’s r</strong> measures how strongly two things move together, on a scale from −1 (perfectly opposite) through 0 (unrelated) to +1 (perfectly together).</p><pre>r = Σ(xᵢ − x̄)(yᵢ − ȳ) / √( Σ(xᵢ − x̄)² · Σ(yᵢ − ȳ)² )</pre><p>The top is the covariance — how the two move together. The bottom divides out each one’s own spread, so r is unit-free. It is the correlation of the two <strong>z-scores</strong> from Episode 2.</p><h3>The number</h3><p>Means: x̄ = 30, ȳ = 5. The deviations, products and squares sum to 160, 1,000 and 36.</p><pre>r = 160 / √(1000 · 36) ≈ 0.843</pre><p>Strong. Positive. About 71% of the variance in config edits is “explained” by Theo’s presence — <em>in the linear sense, and only in the linear sense.</em></p><h3>What it does not prove</h3><p>Correlation says two things move together. It says nothing about <em>why.</em> Three stories fit r = 0.84 equally well:</p><ol><li>Theo makes the edits. (Causation.)</li><li>Someone makes the edits <em>when Theo is there</em>, so the logs point at him. (A third thing — the saboteur’s schedule — drives both.)</li><li>Both are driven by something boring: month-end close keeps everyone late and triggers legitimate config changes.</li></ol><p>I check. The five nights on Walsh’s printout are the last five business days of the month. Half the company was in the building. And the edits to the <em>threshold block specifically</em> — not the routine changes — happened on nights Theo was <strong>home.</strong></p><p>The correlation was real. The story stapled to it was wrong — and the staple was Walsh’s printout.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0MZ63UbpNEJG0psZ_PnZ_Q.png" /></figure><h3>Try it</h3><pre>import numpy as np<br>x = np.array([10,20,30,40,50.]); y = np.array([2,4,5,4,10.])<br>print(round(np.corrcoef(x,y)[0,1], 3))   # 0.843</pre><h3>Remember</h3><p>Ice-cream sales correlate with drownings. Summer causes both. Nobody arrests the ice cream.</p><h3>10:30 a.m. — the memo</h3><p>Reyes’s warrant gives me eighteen months of email from each person who had reason to be in Hall C. Terabytes, mostly noise. I’m not going to read it. I’m going to ask it one question: <em>whose writing does the anonymous memo point the same direction as?</em></p><p>Because people don’t just reuse words. They reuse <em>proportions</em> of words. The way one person says <em>restructure</em> when another would say <em>reorg</em>, <em>threads</em> where another says <em>loose ends</em>, <em>west-side</em> with the hyphen. You can scrub a header. You can’t scrub a habit.</p><h3>Evidence Locker 6 — Cosine similarity</h3><h3>The equation</h3><p>Turn each document into a vector — a count of how often it uses each word. <strong>Cosine similarity</strong> measures the angle between two of those vectors. 1 means they point the same way. 0 means they have nothing in common. It ignores <em>length</em> — a long email and a short memo about the same thing score as similar — and cares only about <em>direction.</em></p><pre>cos(A, B) = (A · B) / (‖A‖ · ‖B‖)</pre><h3>The number</h3><p>Four telltale words. The memo, and two suspects’ email corpora, scaled down so you can check me by hand:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/794/1*YWM4Knn0HYyRciWGuOqnXQ.png" /></figure><p>The memo and <strong>Marisol Kane’s</strong> emails point almost exactly the same direction. Theo’s point somewhere else entirely. He writes about backtests. He has never once written <em>west-side</em> with a hyphen.</p><p>In a real investigation you don’t use four hand-picked words; you use TF-IDF vectors over the whole vocabulary, or sentence embeddings, and compute cosines in 768 dimensions. The formula does not change. Neither does the answer. I ran it both ways.</p><h3>What it proves</h3><p>That Kane wrote the warning. Not who she was warning Dana <em>about.</em></p><h3>Try it</h3><pre>import numpy as np<br>cos = lambda a,b: a@b/(np.linalg.norm(a)*np.linalg.norm(b))<br>M,K,T = np.array([3,1,0,2.]), np.array([2,1,0,1.]), np.array([0,2,3,0.])<br>print(round(cos(M,K),3), round(cos(M,T),3))   # 0.982 0.148</pre><h3>1:15 p.m. — the thing Theo was holding</h3><p>I show Theo the cosine scores because he deserves to know the memo isn’t his, and because I want to watch his face when he reads Kane’s name next to 0.98.</p><p>He goes very quiet. Then he says the thing he has been holding in both hands all week.</p><p>“Last year — about fourteen months ago — Marisol asked me for a favor.” He says it to the table. “A ‘premium-tier scorer.’ A small neural net to flag applicants for a higher loan ceiling. Under-banked ZIPs, that was the pitch. Good PR. I built it in a weekend. In a personal repo, on my own laptop. I sent her a notebook. I never heard about it again.”</p><p>“Was it deployed?”</p><p>“No.” Too fast. Then, slower: “I don’t <em>think</em> so. I never saw it in prod. I never looked.”</p><p>“Why not?”</p><p>He doesn’t answer that. He says “I’m sorry” instead, four times, which I don’t know what to do with. Innocent people apologize. So do the other kind.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AkilVrXeY0utsIgCnwbM6w.png" /></figure><h3>3:40 p.m.</h3><p>I walk into the interview room and put two things in front of Reyes: a correlation that meant nothing, and a cosine that pointed at Marisol Kane.</p><p>“Theo’s credentials were used,” I say. “Theo wasn’t.”</p><p>She reads the numbers for a long time. Reyes reads the way I imagine she interrogates: without blinking, and without letting you know which line she’s on.</p><p>“That clears him of the threshold,” she says. “It doesn’t clear him.”</p><p>She’s right. She knows why.</p><p>“A quant who builds a scoring model for the west-side lender,” she says, “whose login moves the west-side threshold. Whose login kills the rebuild. Tell me how that’s not a story.”</p><p>“It’s a story,” I say. “It’s just not a <em>number</em> yet.”</p><p>Her phone buzzes. She reads it, and her face does something I haven’t seen it do: it softens, for half a second, and then closes.</p><p>“Walsh,” she says. “He’s <em>freezing Kane’s branch access too.</em> As a precaution. He says — I’m reading — ‘she asked for a model review last year and I should have taken it more seriously. I hope this helps.’”</p><p>I hope this helps. Three days, two printouts, one unprompted email, and every single one of them has helped. Every single one has pointed <em>somewhere.</em></p><p>“He’s very helpful,” I say.</p><p>“He is,” Reyes says. She puts the phone face down. “My mother’s shop is on Buckeye Road. Three weeks from closing, in 2009, every bank in this city said no. Meridian’s west-side branch said yes. Thirty thousand dollars.” She doesn’t look up. “The lending manager who signed it was Marisol Kane.”</p><p>I don’t say anything. There isn’t a number for that either.</p><p>“So I’d like to know,” Reyes says, “exactly what that branch turned into. And exactly who turned it.”</p><p>She picks up the memo and reads it one more time. Then she says the thing I’ve been waiting for somebody to say.</p><p>“If Kane wrote the warning — then who has Kane been <em>sending the money to?</em>”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/794/1*ZAziT9sFgfYkwe-vPDRdKw.png" /></figure><h3>Episode 1</h3><p><strong>The Body in the Server Room: </strong>A Financial Crimes Drama Where the Math Is the Witness — A risk analyst is dead. A loan-approval AI has a scar. And the only evidence that can’t be bribed is a learning rate. <a href="https://medium.com/@naeemahaz/the-oracle-files-9cb84ec091fd">https://medium.com/@naeemahaz/the-oracle-files-9cb84ec091fd</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=205d65cfed9e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-4f5f58b18843?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f5f58b18843</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[crime-fiction]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 20:11:03 GMT</pubDate>
            <atom:updated>2026-10-20T20:23:52.250Z</atom:updated>
            <content:encoded><![CDATA[<h3>Thirty-One Sigma: The Night a Risk Analyst Mailed Herself the Evidence — Episode 2</h3><blockquote>A $2,000 courier charge. A login used three hours after its owner went home. And a threshold that only applies west of the river.</blockquote><blockquote><strong>Previously</strong></blockquote><blockquote>Dana Okafor is dead on the floor of Hall C. Her rebuild of the bank’s loan model was killed at epoch 401 by a learning rate nobody types by accident. The loan book has 341 loans where the bell curve says there should be 0.002. And the session that killed the rebuild is signed tlindqvist. <a href="https://medium.com/@naeemahaz/the-oracle-files-9cb84ec091fd">https://medium.com/@naeemahaz/the-oracle-files-9cb84ec091fd</a></blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1237737686%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1237737686%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2214616803-2b32d16831cf611a7e3e6eb82fa520f205a36340983799fabf24828e018c8498-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/2926df5799b763e2487c17b4c6db64e0/href">https://medium.com/media/2926df5799b763e2487c17b4c6db64e0/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*syyZq3AAv2fdVNrWyBsOLQ.png" /></figure><p>Marcus Okafor turns his phone on at 6:50 a.m. because he always turns it off at night, because his sister told him years ago that the blue light was ruining his sleep and he believed her about everything.</p><p>Fourteen notifications. One is from her, 10:51 p.m.</p><blockquote>Marcus — don’t sign anything from Meridian this week. I’ll explain Sunday. I found something in the model and I think I know who put it there, and I’m going to go pull the one log that proves it before they can rebuild it. I know how this sounds. Trust me for three more days. Love you. Turn your phone on.</blockquote><p>He reads it twice. He is still reading it when the knock comes, and he knows what a knock at 6:55 in the morning is, everyone knows, and he opens the door anyway because she said three days and it has not been three days.</p><p>Detective Reyes does not say <em>I’m sorry for your loss</em> until after she has said the other thing, because she has learned that people need the fact first and the sorry second, and she is right, and it doesn’t help.</p><h3>9:15 a.m. — the interview room</h3><p>Theo Lindqvist is thirty-one and has the haircut of someone who forgot to cancel a subscription. He hasn’t slept. He holds the station coffee with both hands like it might be confiscated.</p><p>“I never touched the training box.” His voice is doing the thing voices do when a person has said a sentence to themselves fifty times in a car. “I was in Tucson. My sister’s wedding. There are photos. There are a <em>hundred</em> photos.”</p><p>Reyes lets the silence sit there and do its job. Then she slides a printout across the table. Walsh sent it at 6 a.m., unasked, from his personal address, subject line <em>I hope this helps</em>: Theo’s badge swipes for five weeks, and the timestamps of every config change on ORACLE.</p><p>“Every night the config got edited,” she says, “you were in the building. Every one. Explain that to me.”</p><p>Theo looks past her, at me. “You know this isn’t what it looks like.”</p><p>I don’t know that. Not yet.</p><p>“Your credentials,” I say. “Where do they live?”</p><p>He blinks. “What?”</p><p>“Your login. Where else does it exist besides your head?”</p><p>“The deploy vault,” he says slowly. “Ops needs it to push my notebooks to the training cluster. Half of ops can read that vault. Tom set it up, it’s — “ He stops. “It’s standard.”</p><p>It is standard. I have seen it at six banks. It is also the first useful thing anyone has said to me today.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w-Klc-wgZr1CQ2xYfqV3Ag.png" /></figure><p>Reyes slides her own notebook across the table — not Walsh’s printout, her notes.</p><p>“Badge reader confirms you were in the building most nights the config changed,” she says. “But look at the <em>times.</em>” She points at two entries. “Tuesday, 2:14 a.m. — you were there during the change. Thursday, 11:03 p.m. — you left the building three hours <em>before</em> the config updated.”</p><p>She’s not building a case against Theo anymore. She’s building a timeline.</p><p>“So either someone had his login and used it hours after he left,” I say, “or there’s a gap in your theory.”</p><p>“Or,” Reyes says, “there’s more than one person.”</p><h3>10:30 a.m. — the card</h3><p>Dawn does the thing dawn does to interview rooms: nothing. I only know it’s morning because Reyes hands me a second file and her coffee is a different brand.</p><p>Dana’s corporate card. Ninety days of small charges — coffee, parking, a $12 lunch she probably ate at her desk — and then, 11:52 p.m. on the night she died, twelve minutes after she badged into Hall C, a <strong>$2,000 charge</strong> to a courier company in Tempe.</p><p>“Big for her?” Reyes asks.</p><p>I could say <em>obviously.</em> But <em>obviously</em> is what everybody said about my audit, and I’d like every number I hand this woman to survive a cross-examination, including from me.</p><h3>Evidence Locker 3 — The z-score</h3><h3>The equation</h3><p>A <strong>z-score</strong> tells you how many standard deviations a value sits above or below the mean of its own history. Zero is average. Three is rare. Thirty is a siren.</p><pre>z = (x − μ) / σ</pre><p>Subtract the mean: that <em>centers.</em> Divide by the spread: that <em>scales.</em> The units cancel — dollars, hours, exam points, gone — and what’s left is pure <em>how unusual.</em> One ruler for every witness.</p><h3>The numbers</h3><p>Dana’s daily card spend: μ = $85, σ = $60.</p><pre>z = (2000 − 85) / 60 = 31.9</pre><p>Anything past |z| = 3 is worth a look. Thirty-one point nine is not an anomaly. It is a woman who knew she had twelve minutes.</p><p>Now the same ruler on two suspects. Reyes has HR’s overtime export open on her phone, which is a sentence I would have found alarming a week ago. Same week:</p><ul><li><strong>Theo Lindqvist</strong> logged 30 hours against his usual μ = 12, σ = 6 → <strong>z = 3.0.</strong> Once in 741 weeks.</li><li><strong>Marisol Kane</strong>, the west-side lending manager, logged 22 hours against μ = 20, σ = 8 → <strong>z = 0.25.</strong> Tuesday.</li></ul><p>Theo’s week was extraordinary. That is a fact. What it is a fact <em>about</em> is a different question.</p><h3>What it proves</h3><p>That Dana was sending something, fast, to someone, with her own money. And that Theo’s month looks like a man hiding something — or a man who has just found out what his login has been doing.</p><h3>Try it</h3><pre>z = lambda x, mu, sd: (x - mu) / sd<br>print(z(2000, 85, 60), z(30, 12, 6), z(22, 20, 8))   # 31.9, 3.0, 0.25</pre><h3>Watch for</h3><p>One $2,000 charge inflates σ and shrinks every z-score — including its own. On dirty data use the <strong>modified z-score</strong>, 0.6745·(x − median)/MAD. The median doesn’t flinch.</p><h3>12:10 p.m. — inside ORACLE</h3><p>By noon I’m inside the model itself. A read-only clone of the production repository, pulled through Reyes’s warrant instead of the SSO token Walsh offered twice. I told him my laptop doesn’t do SSO. That’s true. It also doesn’t do Bluetooth, camera, or trust.</p><p>ORACLE is a logistic regression. Respectable. Regulated. Explainable. The kind of model you put in an audit report with a straight face. It takes an applicant’s features — income, debt ratio, years at address, two dozen more — multiplies each by a learned weight, sums them into a raw score, and pushes that score through one function to produce a probability: <em>this applicant has an 88% chance of repaying.</em> Above a threshold: approve. Below: decline.</p><p>The threshold in the config file is <strong>0.50.</strong> It has been 0.50 for three years.</p><p>Except that in one block of the code — a block guarded by an if on the applicant&#39;s ZIP code, four ZIP codes, all west of the river — the threshold is <strong>0.62.</strong></p><p>I read the block three times. It’s clean. Commented. <em>Unit-tested.</em> Someone wrote a test to make sure the discrimination worked.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-7nvuDMEnorhJBa5YTs5pw.png" /></figure><h3>Evidence Locker 4 — The sigmoid</h3><h3>The equation</h3><p>The <strong>sigmoid</strong> is the door where numbers become decisions. It takes any real number and squashes it into a probability between 0 and 1.</p><pre>σ(x) = 1 / (1 + e^(−x))</pre><p>σ(0) = 0.5 — no evidence either way is a coin flip. Big positive scores go to 1. Big negative scores go to 0. Its inverse is the <strong>logit</strong>, the log-odds — which is why every weight in ORACLE has a plain-English meaning a regulator can read.</p><h3>The numbers</h3><p>An applicant from ZIP 85009 gets a raw score of x = 2.3:</p><pre>σ(2.3) = 1 / (1 + e^(−2.3)) ≈ 0.909</pre><p>A 91% chance of repaying. Approved at 0.50. Approved at 0.62. Fine.</p><p>Now the applicant who matters. Same ZIP, raw score that lands at σ(x) = <strong>0.55.</strong> At the bank-wide threshold of 0.50 she is <strong>approved.</strong> At the west-side threshold of 0.62 — same income, same history, same everything — she is <strong>declined.</strong></p><p>The threshold is a <em>business</em> decision hiding inside a <em>mathematical</em> function. Someone made that decision for four ZIP codes only.</p><p>And through the logit, the weights talk: ORACLE’s weight on “years at address” is 0.4, so each extra year multiplies the odds of approval by e⁰.4 ≈ 1.49. Every weight in this model is explainable. That is what will make the west-side threshold indefensible in front of a jury — <em>if</em> I can show who moved it.</p><h3>What it proves</h3><p>ORACLE was aimed. Not broken — aimed. Twelve percentage points of extra suspicion applied to people on the basis of where they live, and nowhere else.</p><h3>Try it</h3><pre>import numpy as np<br>s = lambda x: 1/(1+np.exp(-x))<br>print(round(s(2.3),3), round(np.exp(0.4),2))   # 0.909, 1.49</pre><h3>1:40 p.m.</h3><p>I pull the history on the threshold block. git log -p on one file — the oldest command in the world, the one that never lies unless someone has rewritten it, which is a thing I check too.</p><p>The change landed <strong>fourteen months ago, at 2:14 a.m.</strong> Author: <strong>tlindqvist.</strong></p><p>Two sessions. Fourteen months apart. One moved a threshold. One killed a training run five minutes after a woman walked into a room. Both signed with the same name.</p><p>Reyes reads it off my screen without a sound. Then she says the thing I’ve been thinking since Theo said the word <em>vault.</em></p><p>“If I wanted a quant to take the fall for something, I’d want it to look exactly like this.”</p><p>“Or,” I say, “you’d have a quant who did it twice.”</p><p>She doesn’t answer. She’s looking at her phone.</p><p>“Courier company in Tempe,” she says. “The package Dana paid $2,000 to send at 11:52 last night. It was addressed to <em>herself.</em> Her apartment. Her brother signed for it an hour ago.”</p><h3>3:00 p.m. — the package</h3><p>Marcus brings it to the station himself because he does not trust anyone else to carry it, and I understand that completely.</p><p>Inside: a USB drive, and a sticky note in Dana’s lowercase. <em>offline copy. just in case.</em></p><p>She mailed herself the evidence, because she knew a shared drive is a shared drive.</p><p>One text file on the drive, besides her own notes. An anonymous internal memo, sent three weeks before she died from a burner mailbox on Meridian’s own corporate domain — which is a thing you can do if you’re an admin — warning her to <em>stop pulling threads on the west-side restructure.</em></p><p>No name. Headers scrubbed.</p><p>But words have fingerprints. People reuse vocabulary the way they reuse gestures. You can change your email address a lot easier than you can change how you say <em>restructure.</em></p><p>Reyes reads it and goes very still.</p><p>“Someone warned her,” she says. “Three weeks before. Someone <em>inside</em> the bank knew she was close, and wrote to her, and didn’t sign it.”</p><p>“A threat,” I say.</p><p>“Or a kindness.” She puts the sticky note down with two fingers, like it might bruise. “You’d be surprised how often they look the same.”</p><p>Behind her, through the glass, I can see Theo in the hallway, waiting to be told whether he can go home. He hasn’t put the coffee cup down. He is looking at the floor the way the man in the hoodie looked at the floor in Hall C, and it occurs to me — late, the way everything occurs to me this week — that I have now seen two men in this case who cannot look up, and I have only asked one of them why.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/794/1*k1yN5IDCH73Qk6tuEMQ4Nw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f5f58b18843" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-9cb84ec091fd?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/9cb84ec091fd</guid>
            <category><![CDATA[naeemahaliyasmall]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[money]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 00:38:06 GMT</pubDate>
            <atom:updated>2026-10-20T20:35:02.297Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>In the Phoenix financial district, the crimes committed by machines are investigated by two separate yet equally important groups: the detectives who find the bodies, and the auditors who find the numbers.</blockquote><h4>Episode 1</h4><p><em>This episode is a dramatization. Meridian Trust and everyone in it are invented. The practices depicted are composites of patterns documented in real fair-lending, model-risk and financial-crimes enforcement. Every equation is real; every number computed from them is correct. The crimes are fiction. The arithmetic is not.</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1236087849%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1236087849%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2212492120-366b6ef0bf6dfd8cddfbe5ff2f394648c81b18337962ac550156e32ece9c75e2-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/03a64dc3dac09f7866f8002f9f51540a/href">https://medium.com/media/03a64dc3dac09f7866f8002f9f51540a/href</a></iframe><p>Luis Ferreira has walked the Camelback data center four times a night for six years and has never once found anything but a propped fire door and, one time, a raccoon.</p><p>At 1:10 a.m. on the fourth round he swipes into Hall C, which is colder than the others because the bank’s model-training cluster lives there and the bank’s model-training cluster runs hot. He does the thing he always does: counts the status lights on the first rack, because counting is how you stay awake. He gets to eleven. Then he sees the tarp.</p><p>It’s the grey one from the cable-pull last month. It’s on the raised floor between racks C-14 and C-15, and it’s not flat. Someone started to pull it over something and stopped. There is a hand, palm up, at the edge of it. There is a laptop open on a cart beside the hand, screen awake, white text scrolling on black, line after line, like it’s reading something out loud to nobody.</p><p>Luis does not touch the tarp. He does not touch the laptop. He backs out of Hall C the way he came in and stands in the corridor and calls it in with the kind of steady voice that will make the 911 operator ask him twice if he’s sure.</p><p>He is sure. He saw the ring on the hand. He knows whose ring it is. She brought him a coffee once at 3 a.m. and asked him what the lights meant.</p><h3>1:52 a.m.</h3><p>Hey. You’re back.</p><p>I don’t know why I keep talking to you. Maybe because you don’t have a login. You can’t be subpoenaed, or phished, or bought a drink by a CTO. You’re the only clean channel I have left, so this is where the case goes.</p><p>My phone lights up on the floor beside the mattress, which is where my phone lives, which tells you most of what you need to know about my life. Unknown number. Phoenix area code. I let it ring twice, because two rings is what people who sleep would do.</p><p>“You’re the one who audits the loan models at Meridian Trust.” Not a question. The voice belongs to a woman who sounds like she has been awake since the last time anybody believed a bank.</p><p>“Detective Inés Reyes. Financial Crimes, on loan to Homicide tonight, which I’ll explain if you ask me, and I’d rather you didn’t. I need you at the Camelback data center. Now. Bring whatever you use to read a hard drive.” She pauses. The pause is exactly one breath too long. “Don’t call anyone. Don’t tell anyone. Just come.”</p><p>I don’t ask why. If I ask, she’ll have to tell me, and if she tells me, I’ll have to decide whether she’s telling the truth — and I’ve made enough decisions about Meridian Trust already.</p><p>Here’s the one you should know about. Eighteen months ago I signed the independent audit of Meridian’s loan model. Forty pages. My name on the last one. A sentence I put in bold because I was proud of it:</p><p><em>No evidence of disparate treatment by geography.</em></p><p>Two weeks ago Dana Okafor called me — Dana, who never calls, who Slacks in complete paragraphs with citations — and asked, very carefully, whether I had checked the threshold logic <em>per ZIP code</em>, or only in aggregate.</p><p>I told her I’d look.</p><p>I remember the feeling of having looked. I don’t remember looking.</p><p>You see the problem.</p><h3>Hall C</h3><p>Meridian Trust. <em>Trust.</em> They put it in the name the way you put a lock on a door you already know is broken.</p><p>Hall C is colder than a morgue and lit the same blue-white. The racks blink in rows like a city seen from a plane. On the raised floor between two of them, under a tarp someone hasn’t finished pulling into place, is Dana.</p><p>Senior risk analyst. Thirty-four. The only person at that bank who ever asked me a question I couldn’t answer from the slide deck.</p><p>I don’t look at her long. Looking means accepting. I count the ceiling tiles instead — sixty-four, eight by eight, a power of two — and it works for exactly four seconds.</p><p>“You’re the auditor.”</p><p>A man steps out from behind Reyes: tall, forty-five, a suit pulled on over a T-shirt, the grey rinsed face of someone who has been crying in a parked car and doesn’t care who knows. Nathan Walsh. Chief Technology Officer. I’ve seen him on a conference stage talking about <em>responsible AI at scale</em> with a lavalier mic and the confidence of a man who has never been wrong in a room he controlled.</p><p>“I approved the camera maintenance ticket,” he says, before I can ask. “Last Tuesday. Routine. I already told the detective.” He swallows. “Whatever you need. Access, logs, people. Call me directly. Don’t go through IT.”</p><p>He looks me in the eye when he says it. He is the first person tonight who has. I notice that I’m grateful for the directness, and then I notice myself noticing, and I file both of those things away.</p><p>Behind him, in the open door of the cage, a man in a grey hoodie stands with his hands in his front pocket and his eyes on the floor. Platform engineer, by the lanyard. The kind of person who keeps the lights on in a building like this and is never once introduced in a meeting. Reyes doesn’t introduce him either. He doesn’t come in.</p><p>Reyes doesn’t do warmth. She does inventory.</p><p>“Badged in at 11:40 p.m.” She’s reading from a notebook, not from memory. “Alone. Cameras in this hall down since Tuesday. ME preliminary: blunt-force trauma, consistent with the edge of a server rack.” She looks up. Only at me. “Could be a fall. Could be many things.”</p><p>The word <em>could</em> hangs there. Not like she’s uncertain. Like she’s testing whether I’m going to believe her.</p><p>“Her laptop was open. Still running. The training log on that box” — she tips her head at a two-unit server halfway down the row, its drive light stuttering — “was still writing to it.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CqtQid329l78h6kF17Ju0Q.jpeg" /></figure><p>I crouch by the laptop. A terminal, a loss value ticking per line like a countdown that’s forgotten what it’s counting to. Behind the terminal, a notebook. Tab title in Dana’s lowercase — the way she names everything, filenames, variables, commit messages, like capital letters were noise and she couldn’t afford noise:</p><p><strong>walk downhill.</strong></p><p>Under the title, one line, dated yesterday:</p><blockquote>if anyone but me is reading this, start at epoch 401. don’t trust the dashboard. don’t trust the audit either — sorry. — D.</blockquote><p>I read it once. I read it again.</p><p>The second read, I’m not reading the words. I’m reading my name. Forty pages deep. In bold.</p><p>Behind me I can feel Reyes reading it too. I can feel her <em>not</em> reacting to the part about the audit.</p><p>Which means she already knew.</p><p>Which means she brought me here knowing I’d see this.</p><p>The question is whether she brought me here to help her find what Dana found — or to find out how much I already knew.</p><h3>Evidence Locker 1 — Gradient descent</h3><h3>What you’re looking at</h3><p>The log on Dana’s laptop is a model being trained — her own rebuild of ORACLE, from the raw loan applications, without the production configuration. She was rebuilding the bank’s brain from scratch to see if it came out the same. Each line is an <strong>epoch</strong>: one full pass over the data. Each epoch reports a <strong>loss</strong>: one number for <em>how wrong the model is right now.</em></p><p>For four hundred epochs the loss falls. Steadily. Boringly. Each step a little smaller than the last, the curve flattening as it approaches the bottom of whatever valley this particular landscape has. That is what a model looks like when nobody is breaking it.</p><p>Then epoch 401, and the loss <strong>jumps upward.</strong> By epoch 410 the terminal is printing nan — <em>not a number</em> — the machine&#39;s way of saying <em>I fell off the edge and I don&#39;t know where solid ground was.</em></p><h3>The equation</h3><p><strong>Gradient descent</strong> is the engine under every neural network, every credit score, every recommendation you’ve ever been shown. Imagine walking down a hill in fog. You can’t see the bottom. You can feel the slope under your feet. You step downhill. As the slope flattens, your steps get smaller on their own. Eventually, you’re at the bottom.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*zMKle34zv4wreQnhMh0suA.png" /></figure><h3>The number</h3><p>Dana’s second notebook cell has a toy: loss J(θ) = (θ − 3)², bottom at θ = 3. Start at θ = 0 with a learning rate of α = 0.1:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/516/1*M9dP0ltdwSWSSRBk-pAaVQ.png" /></figure><p>The steps shrink by themselves. Nobody told the algorithm to slow down. The geometry did.</p><p>Now set α = 1.1 from the same start: θ goes 6.6, then −1.32, then 8.18. Each bounce wider than the last. The loss <em>grows.</em> That is <strong>divergence</strong>, and it is exactly what a log looks like just before it prints nan.</p><p>I scroll to epoch 401 in Dana’s log. The learning-rate field, epochs 1–400: lr=0.001. Conservative. Correct. Then, epoch 401: lr=2.5.</p><p>Nobody types 2.5 by accident. 2.5 is two thousand five hundred times larger than 0.001. 2.5 is the number you pick when you <em>know exactly</em> what will happen when you plug it in — and you want it to look like an exhausted analyst fat-fingered her own config at midnight.</p><h3>What it proves</h3><p>Dana’s rebuild didn’t fail. It was killed. By someone who understood the engine well enough to know which single number to change, and understood people well enough to know almost nobody would recognize the fingerprint.</p><pre>def gd(grad, theta, alpha, iters=20):<br>    for _ in range(iters): theta -= alpha * grad(theta)<br>    return theta<br>print(gd(lambda t: 2*(t-3), 0.0, 0.1))   # converges near 3<br>print(gd(lambda t: 2*(t-3), 0.0, 1.1, 5)) # watch it leave</pre><h3>4:05 a.m. — the loan book</h3><p>Reyes has a court order for Meridian’s loan book by four in the morning, which tells you something about which judges she has on speed dial, and something else about how long she’s been waiting for a reason to use it.</p><p>Fourteen thousand loans approved by ORACLE in eighteen months, exported to a CSV that I open in a text editor, because I don’t let bank data touch a spreadsheet application that phones home.</p><p>Walsh is still here. He has brought coffee for everyone, including the man in the hoodie, who takes it without looking up. Walsh stands at a respectful distance from my screen and does not quite look at it, the way a man stands near a stove.</p><p>“If it helps,” he says, “I can have my team pull a clean extract of anything. Just say the word.”</p><p>“Thank you,” I say, and do not say the word.</p><p>The first thing you do with fourteen thousand numbers is the first thing anyone should do. You draw them. A histogram of loan amounts, plain matplotlib, no styling. The ugly one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mB5nzIO97BfCBAr1qnCxsg.png" /></figure><p>It should be a hill. Most loans in the middle, a few small, a few large — the shape money makes when a thousand ordinary decisions pile up.</p><p>What I get is a hill with a second, smaller hump on the right. A shoulder, at exactly $88,000 to $92,000, that has no business being there.</p><p>An honest ledger has a shape. This one has a scar, and someone has been picking at it.</p><h3>Evidence Locker 2 — The normal distribution</h3><h3>The equation</h3><p>The <strong>normal (Gaussian) distribution</strong> — the bell curve — is what you get when many small, independent things add up. Income, need, appetite, policy: a loan amount is the sum of hundreds of small decisions, so an honest book of loans is a bell. Two numbers describe it completely: the mean <strong>μ</strong> and the standard deviation <strong>σ</strong>.</p><pre>f(x) = (1 / (σ√(2π))) · e^( −(x − μ)² / (2σ²) )</pre><p>The tails shrink toward zero extremely fast. That is the whole point: the bell tells you how often “far from normal” should happen <em>by chance</em>.</p><h3>The number</h3><p>Meridian’s honest loans: μ = $42,000, σ = $9,000.</p><ul><li>68% of loans should fall between $33k and $51k.</li><li>95% between $24k and $60k.</li><li>99.7% between $15k and $69k.</li></ul><p>How many loans should exceed $88,000? That is (88 − 42) / 9 = <strong>5.1 standard deviations</strong> above the mean. The fraction of a bell that far out is about 1.7 × 10⁻⁷. Multiply by 14,000 loans and the expected count is <strong>0.002</strong>.</p><p>I count <strong>341</strong>.</p><p>I count them twice. Then I put the laptop down and walk the length of Hall C and back, because 341 divided by 0.002 is a number that doesn’t mean anything except <em>this did not happen by chance</em>, and I need my hands to stop.</p><h3>What it proves</h3><p>A shoulder is not noise. It is one decision, made 341 times, by someone.</p><h3>Try it</h3><pre>from scipy.stats import norm<br>print((1 - norm.cdf(88000, 42000, 9000)) * 14000)   # ≈ 0.002 expected</pre><h3>5:40 a.m.</h3><p>I pull the config history on the training box — who set the learning rate, and when.</p><p>The edit landed at <strong>11:45:10 p.m.</strong> Five minutes after Dana badged into Hall C. It was made over an SSH session from inside the building, from a host called ops-console-01.</p><p>I pull the session’s user.</p><p>I read it three times, because the first two times it doesn’t make sense, and the third time it makes a kind of sense I don’t want.</p><p><strong>tlindqvist.</strong></p><p>Theo Lindqvist. The quant. The one who told me last spring, in the break room, with a paper cup and total sincerity, that ORACLE was “cleaner than anything in the industry.”</p><p>Walsh is reading over my shoulder. I can feel him do it, the way you feel someone stand too close on a train. He exhales like the air costs him something.</p><p>“Theo’s been… strange, the last few months,” he says quietly. “Late nights. I didn’t want to say anything.” His hand lands on my shoulder for exactly one second. “Get some sleep. I’ll have his access frozen by morning.”</p><p>Reyes is already reaching for her phone.</p><p>In the cage doorway, the man in the hoodie has finished his coffee and is holding the empty cup in both hands like it’s evidence.</p><p>“Wait,” I say.</p><p>I don’t know why I say it. Maybe because I’ve spent the last five hours learning that a training run was sabotaged to look like an accident, and once you’ve seen one thing built to look obvious at midnight, you start wondering what else was.</p><p>Or maybe because he touched my shoulder, and I don’t like being told to sleep.</p><p>You saw it too, right? Tell me you saw it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/496/1*jDGZjAbuxO6M7wVshcZoaA.png" /></figure><p><strong>Put yourself in the room.</strong> A detective calls you at 1:52 a.m. and says <em>don’t call anyone.</em> You have a signature on a document that might be wrong. Do you go? Do you call a lawyer first? Do you tell her about the audit before she finds out — or after?</p><p><strong>The debate I want in the comments:</strong> Walsh offered access, logs, people, and a “clean extract.” Team Helpful says that’s what a decent executive does at 4 a.m. when one of his people is dead. Team Careful says nobody offers to pull the evidence for you unless they’d like to be the one who pulls it. Which team are you on, and what’s your tell?</p><p><strong>Thirty-One Sigma: The Night a Risk Analyst Mailed Herself the Evidence - Episode 2</strong></p><blockquote>A $2,000 courier charge. A login used three hours after its owner went home. And a threshold that only applies west of the river.</blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1237737686%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1237737686%3Ffl%3Dip%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2214616803-2b32d16831cf611a7e3e6eb82fa520f205a36340983799fabf24828e018c8498-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/2926df5799b763e2487c17b4c6db64e0/href">https://medium.com/media/2926df5799b763e2487c17b4c6db64e0/href</a></iframe><p><a href="https://medium.com/@naeemahaz/the-oracle-files-4f5f58b18843">https://medium.com/@naeemahaz/the-oracle-files-4f5f58b18843</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9cb84ec091fd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Oracle Files: Machine Learning Fraud Detection Crime Thriller Teaching Data Science Equations]]></title>
            <link>https://medium.com/@naeemahaz/the-oracle-files-machine-learning-fraud-detection-crime-thriller-teaching-data-science-equations-1cb497929354?source=rss-e449fc8bdc50------2</link>
            <guid isPermaLink="false">https://medium.com/p/1cb497929354</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[crime]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Naeemah Small]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 15:51:35 GMT</pubDate>
            <atom:updated>2026-09-13T15:51:35.241Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D8nx9dMEbJnUxvFnQMIQIg.png" /></figure><p><a href="https://naeemahsmall.com/">By Naeemah Aliya Small</a> (7 Part story — introduction)</p><p>A DATA SCIENTIST IS FOUND DEAD IN A DATA CENTER. HER NEURAL NETWORK WAS LYING.</p><p>When Dana Okafor is discovered in Meridian Trust Bank’s server room, the official verdict is suicide. But her notebook tells a different story: “Walk downhill.” “Don’t trust the dashboard.” “Don’t trust the audit either.”</p><p>A statistical auditor is brought in to investigate. What they uncover is a neural network secretly trained to discriminate — and someone willing to kill to keep it hidden.</p><p>The evidence is pure mathematics.</p><p>Gradient descent reveals sabotage. Z-scores expose the impossible. Multiple regression proves discrimination. Eigenvectors uncover hidden money flows. Each equation becomes a clue. Each calculation brings them closer to a killer who engineered the perfect crime using the one language no one checks: algorithms.</p><p>In the darkness of a data center, beneath the hum of servers worth millions, lies a murder that only mathematics can solve.</p><p>But what if the auditor is being played?</p><p>What if the numbers are designed to point at the wrong killer?</p><p>What if the real crime is hidden in the space between equations?</p><p>Five suspects. Every one has motive. Every one has means. Every one’s timeline has holes that don’t add up.</p><p><strong>THEO LINDQVIST </strong>— the quant whose login was used when the model was sabotaged. But his alibi breaks down under statistical analysis. And his prototype scorer was suspiciously clean.</p><p><strong>MARISOL KANE </strong>— the lender whose emails are 98.2% similar to the discrimination memo. She knew the system. She knew the cooperating applicants. But was she the mastermind or a pawn?</p><p><strong>NATHAN WALSH</strong> — the CTO who volunteered helpful evidence on day one. Evidence that conveniently omitted one critical file. His token was in the system two minutes before Dana badged in.</p><p><strong>DETECTIVE INÉS REYES</strong> — the investigator who brought the auditor in. Who knew to look at the mathematics instead of the code. Who seems to know more than she should.</p><p><strong>THE AUDITOR </strong>— the one who signed off on ORACLE eighteen months ago. The one who said there was “no evidence of disparate treatment.” The one carrying guilt.</p><p>The mathematics tells only one story.</p><p>But there are five people who could have written it.</p><p>And the crime doesn’t stop with Dana’s death. It spirals deeper — into hidden models, money laundering networks optimized through Lagrange multipliers, eigenvector analysis tracing flows through shadow accounts, KL divergence measuring how perfectly the deception was executed.</p><p>Someone at this bank understands mathematics at the level of a weapon.</p><p>Someone knows how to hide crimes inside algorithms.</p><p>Someone knows that when you understand the equations, you understand everything.</p><p>— -</p><p>“A gripping, meticulously plotted thriller about algorithms, ambition, and the cost of discovering the truth.”</p><p>The Oracle Files is a crime investigation where data science becomes forensics, mathematics becomes motive, and the only witness is mathematics itself.</p><p>Twenty-four equations. One unsolved homicide. No one is innocent.</p><p>The mathematics never lies.</p><p>But the people who understand it do.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1cb497929354" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>