<?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 Rohit Sharma on Medium]]></title>
        <description><![CDATA[Stories by Rohit Sharma on Medium]]></description>
        <link>https://medium.com/@rokxyi?source=rss-b5152600b35b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*1HVpwiVHEsMEQhX_AttWsg.jpeg</url>
            <title>Stories by Rohit Sharma on Medium</title>
            <link>https://medium.com/@rokxyi?source=rss-b5152600b35b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:32:25 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rokxyi/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[Beacon: What Happens When a Project Manager Builds an AI Co-pilot Instead of Another Dashboard]]></title>
            <link>https://medium.com/@rokxyi/beacon-what-happens-when-a-project-manager-builds-an-ai-co-pilot-instead-of-another-dashboard-d0104972741e?source=rss-b5152600b35b------2</link>
            <guid isPermaLink="false">https://medium.com/p/d0104972741e</guid>
            <category><![CDATA[project-management]]></category>
            <category><![CDATA[gen-ai-project-management]]></category>
            <dc:creator><![CDATA[Rohit Sharma]]></dc:creator>
            <pubDate>Thu, 16 Jul 2026 04:01:21 GMT</pubDate>
            <atom:updated>2026-07-16T15:21:50.979Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6zddh4tH4t0Np2ieKVR-SQ.png" /></figure><p>Managing projects taught me what our tools keep missing. So I built one that doesn’t.</p><p>It’s Monday morning on a vehicle program in 2026.</p><p>Your software team flagged a defect that, on another program, would have meant an OTA recall — <a href="https://www.deloitte.com/us/en/industries/consumer/about/automotive-software-trends.html">roughly 4 in 10 SDV teams have lived through one</a>. Purchasing is escalating a memory-chip supply risk because <a href="https://www.spglobal.com/automotive-insights/en/blogs/2026/01/automotive-market-trends-2026">AI data centers are swallowing the world’s DRAM</a>. Marketing wants to know if the trim strategy changes now that a third of your EV intenders are drifting back toward hybrids.</p><p>And you have a gate review on Thursday.</p><p>Where does your project management system help you with any of this?</p><p>If you’re honest: it doesn’t. It stores tasks. It renders Gantt bars. It waits for you to type things into it.</p><p>Let me be clear up front: <strong>I’m not selling you software.</strong> I built Beacon, an AI co-pilot for project managers, for my own practice — after years of running projects and a recent PMP that gave my old frustrations a precise vocabulary. What I want to sell you is an <em>idea</em>: the gap in our tooling was never data capture. It’s applied judgment. And that gap is now closeable.</p><h3>Every PM tool you’ve used is a system of record</h3><p>Project, Primavera, Jira, Planview — all valuable, all passive. They remember what you tell them. The actual work of project management — interpreting signals, deciding which process to run next, turning a messy meeting into decisions and actions — happens outside the tool, in your head, at 9pm.</p><p>Here’s the shift, in one picture:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_9LaRTC9e8u_QyFYK9XKiQ.png" /><figcaption>Where does interpretation live? Left: the human absorbs it all after the tool is done. Right: the tool interprets and drafts; the human decides.</figcaption></figure><p>In the left column, notice where the red box sits: interpretation, drafting, minutes — all on the human, after the tool has done its job. In the right column, the tool does the interpreting and drafting; the human does the deciding. Nothing about your accountability changes. Everything about your Tuesday does.</p><p>Three design choices make the right column real:</p><p><strong>The tool knows the methodology.</strong> Beacon carries a structured PMBOK knowledge base into every interaction — not because PMBOK is the only way to run projects (it isn’t), but because a co-pilot needs <em>some</em> shared playbook to reason from. It classifies what you’re asking (risk review? EVM concern? change assessment?), pulls the relevant process, and grounds its answer in your project’s live state. Swap in PRINCE2 or your company’s gated process; the principle holds.</p><p><strong>The tool speaks first.</strong> It scans project data continuously and raises ranked alerts — cost drift, stale risks, overdue responses. A dashboard waits to be read. A co-pilot taps you on the shoulder.</p><p><strong>The tool writes the documents.</strong> Charter, risk register, status report, stakeholder plan, change log — drafted from live data with charts embedded, for you to edit. Judgment, not blank pages.</p><h3>One lap around an automotive program</h3><p>Automotive is the perfect stress test because it compresses every modern PM problem into one program: hardware and software lifecycles colliding, brutal supply volatility, regulatory weight, and a stakeholder web that spans plants, purchasing, and three tiers of suppliers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H-OxPpGRZh3Kw2sz3D-tHQ.png" /><figcaption>A program’s full lifecycle with a co-pilot in the loop. Blue: co-pilot acts. Yellow: governance gate. Red: proactive alert. Green: human decision.</figcaption></figure><p>A few moments from that map, concretely:</p><p><strong>Kickoff.</strong> You describe the program in plain English — “18-month mid-cycle refresh, new zonal E/E architecture, three plants, SOP Q1 2028” — and get a draft charter and decomposed WBS to react to.</p><p><strong>Earned value becomes a conversation.</strong> Not “CPI = 0.87,” but what an over-budget tooling line at your battery-pack supplier means for the estimate at completion, and which corrective options fit. EVM stops being a spreadsheet ritual performed for the PMO and starts being how you actually think.</p><p><strong>Risk stays alive.</strong> The DRAM shortage goes into a proper register — probability, impact, response, owner — and when it goes stale while the market moves, that’s an alert, not a surprise at the gate review.</p><p><strong>Meetings become data.</strong> Paste the transcript from a program review with your plant and tier-1s; decisions, actions, risks, and issues come out structured. Nothing said in the room evaporates, and nobody spends their evening writing minutes.</p><p><strong>People get strategies, not spreadsheet rows.</strong> Ask how to move a resistant plant manager toward supportive, and you get an engagement approach built from that stakeholder’s actual profile. Across programs, the same machinery flags that your embedded-software engineers are double-booked on two vehicle lines — the single most common failure mode I see in automotive orgs restructuring around software talent — and tells you which project needs your attention first this week.</p><p>That last cluster is the real payoff. The co-pilot eats the administrative load, so the PM’s hours go where no tool can: negotiating, coaching, deciding.</p><h3>The unglamorous part that changes the most: governance</h3><p>Automotive programs drown in engineering changes. Most tools treat change control as a form; process discipline lives in a policy PDF nobody reads.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4hsAkIawO5uFAZXEYgYGvQ.png" /><figcaption>The change-control loop the tool enforces. Direct edits to a locked baseline are blocked; every path leaves an exportable audit trail.</figcaption></figure><p>In Beacon, baselines lock through an approval workflow. Re-baselining happens <em>only</em> through an approved change request — the tool physically won’t let you quietly edit a locked baseline. Propose a change (“supplier requests switching the ECU memory vendor”) and the impact on scope, schedule, cost, and risk is assessed before it reaches your CCB. Every mutation is logged: who, what, when, before and after — exportable when the customer or auditor asks.</p><p>In a sector run on IATF 16949 mindsets and now UN R155/R156 software regulations, that evidence trail isn’t bureaucracy. It’s the difference between an audit and an archaeology dig.</p><h3>What I actually learned building it</h3><p>Beacon runs — a real backend, a real database, real generated documents. But the lesson isn’t “look what I built.” It’s this:</p><p>Most of what exhausts project managers was never the management. It was the administration <em>around</em> the management. Status compilation, minutes, register hygiene, report formatting, chasing updates — none of it is judgment, and all of it is now automatable by a tool that understands project management as a discipline rather than as a data schema.</p><blockquote>Our frameworks were always executable. We just kept executing them by hand.</blockquote><p>The methodology — PMBOK or otherwise — turns out to be the key that unlocks this. Not as doctrine, but as structure an AI can reason over.</p><p>Experience taught me the difference between managing a project and administering one; the PMP gave that difference a shared language. The tools should know it too — and now they can.</p><p><em>What’s the one thing your PM tool should do for you but doesn’t? I’m collecting answers — the best ones become the next thing I build.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d0104972741e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Trading AI That Had to Prove Itself to a Time Machine]]></title>
            <link>https://medium.com/@rokxyi/the-trading-ai-that-had-to-prove-itself-to-a-time-machine-8ceaf32ba6ba?source=rss-b5152600b35b------2</link>
            <guid isPermaLink="false">https://medium.com/p/8ceaf32ba6ba</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[trading]]></category>
            <dc:creator><![CDATA[Rohit Sharma]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 15:47:33 GMT</pubDate>
            <atom:updated>2026-07-15T15:47:33.944Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Five AI agents argue about stocks. Pure software holds the veto. And every design decision — good or bad — had to survive a replay of the past before it was allowed to stay in the code. Here’s the logic, the evidence, and the bugs, explained for people who don’t code.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VCXlnAApki8VOUO0YcSlxg.png" /></figure><p>Most stories about AI trading bots follow the same arc: someone wires a language model to a brokerage account, it makes a few lucky trades, and the screenshot goes viral. This is not that story.</p><p>This is the story of a system whose most common decision — by a landslide — is to do <strong>nothing</strong>. In a full-year replay of the market, it was asked for a decision 753 times and said HOLD on 708 of them. It bought 31 times. It sold 14 times.</p><p>And the interesting part isn’t the bot. It’s the <strong>referee</strong>: a simulator that replays the past, one day at a time, letting the AI see only what it could have known on each date. Every meaningful claim about this system — every feature that stayed, every feature that got deleted, every bug that got found — traces back to that referee.</p><p>That’s the logic worth explaining. Not “AI picks stocks.” Rather: <strong>how do you force an AI system to earn every one of its opinions?</strong></p><h3>Part 1: The Machine — A Committee, Not an Oracle</h3><p>The first design judgment is the biggest one: <strong>no single AI ever decides anything.</strong></p><p>Instead, the system is built like a small investment firm where every employee is an AI — and the compliance department is not an AI at all. For each stock, on each cycle, the same debate plays out:</p><p><strong>The Analyst</strong> reads the raw material: price momentum indicators (with names like RSI and MACD — think “speedometers for a stock”), company finances, news headlines, social-media chatter, and the overall market’s fear level (the VIX index). It issues a verdict — bullish, neutral, or bearish — with a confidence score.</p><p><strong>The Bull and the Bear</strong> are two AIs forced into opposition. One must make the strongest honest case <em>for</em> buying. The other must make the strongest honest case <em>against</em>. Neither is allowed to be balanced — that’s the point.</p><p><strong>The Judge</strong> reads the debate and scores which side brought better <em>evidence</em>, not which side sounded more confident. It issues its own verdict and an updated confidence.</p><p><strong>The Portfolio Manager</strong> makes the final call — BUY, SELL, or HOLD — under strict, written conditions. It may buy only if confidence clears a high bar (65%), the signals backed by <em>real</em> data agree, and there’s cash available. Its default, absent strong agreement, is HOLD.</p><p><strong>The Risk Officer</strong> is the twist: it contains <strong>no AI whatsoever</strong>. It is plain, deterministic software — a list of rules that can <em>shrink</em> or <em>block</em> any order but can never create one. Position size capped at 20% of the account. Maximum 3 trades per day. A circuit breaker that halts everything after a 10% daily loss. A settlement-law guard (more on that later). A market-hours gate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/871/1*wdgw-LRTY5hjbV-1rA-4Kg.png" /><figcaption><em>One stock, one cycle: the committee debates, the judge scores evidence, the manager decides, and non-AI rules get the final word.</em></figcaption></figure><p>The philosophy compresses into one sentence, and it’s the sentence that governs the whole codebase:</p><blockquote><strong><em>The AIs propose. Only the rules dispose.</em></strong></blockquote><p>Why so much ceremony? Because the earliest version of this system had exactly the failure you’d predict — and the simulator caught it.</p><h3>Part 2: The Time Machine — How the System Gets Tested</h3><p>Before trusting any trading logic with even fake money, the system is run through a <strong>historical simulation</strong>: pick a window of the past (say, June 2025 through June 2026), pick a few stocks (NVDA, Apple, Microsoft), and replay it one trading day at a time.</p><p>The crucial rule is <strong>point-in-time honesty</strong>: on each simulated day, the committee is shown only the data that existed <em>up to that day</em>. No peeking at tomorrow. Then its decisions are graded against what the market actually did next — did the price rise in the 20 trading days after a BUY? Did a SELL dodge a drop?</p><p>Each full-year replay produces a report card: total return, worst drawdown, win rate, decision quality — and it costs a few dollars in AI fees to run. That last detail matters more than it sounds: <strong>when an experiment costs less than lunch, you can afford to test every opinion you have.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/969/1*sEbFyvwVjkflO8u7uYDHOg.png" /><figcaption><em>The engine of the whole project: no feature survives on vibes.</em></figcaption></figure><p>Notice the second box. Every experiment’s pass/fail criteria are <strong>written down before the experiment runs</strong> — a practice borrowed from clinical trials called pre-registration. The project keeps a registry file where each hypothesis gets a row <em>before launch</em>, so nobody (human or AI) can stare at the results afterward and invent a story about why they secretly count as a win.</p><p>This loop — suspect, pre-register, replay, measure, adopt or reject — is the real engine of the project. Everything in the next three sections came out of it.</p><h3>Part 3: What the Simulation Proved</h3><p>Run against history, the committee’s results come in two windows, and honesty requires showing both.</p><p><strong>The flattering window (June–November 2025, six months):</strong></p><ul><li>The system returned <strong>+15.4%</strong>, edging out the S&amp;P 500’s +14.6%.</li><li>Its worst peak-to-trough drop was just <strong>−2.2%</strong> — it kept cash cushions instead of riding every dip fully invested.</li><li><strong>60% of resolved trades won</strong>, and its entries were genuinely good: 20 trading days after a buy, the price was higher <strong>83% of the time</strong>.</li></ul><p><strong>The humbling window (June 2025–June 2026, a full year, on the current code):</strong></p><ul><li>The system returned <strong>+13.3%</strong> with a worst drawdown of −8.6%.</li><li>Simply buying and holding those same stocks returned <strong>+38.8%</strong>. The S&amp;P 500 returned +29.4%.</li><li>The win rate on its trades was strong — <strong>68%</strong>, with buys higher 20 days later <strong>79% of the time</strong> and the average buy up about +5% at that mark.</li></ul><p>Read those together and you get the system’s true character, which its own documentation states plainly: it is a <strong>drawdown-avoider, not a bull-market-beater</strong>. In a roaring market, caution has a measurable price — and the simulator put an exact number on it: measured 20 trading days out, the average HOLD-in-cash decision carried an opportunity cost of about <strong>5.6% in missed gains</strong>. That’s not a flaw the authors hide; it’s a trade-off they chose, quantified.</p><blockquote><em>A 60–68% win rate also means this: </em><strong><em>roughly 4 out of every 10 trades lose.</em></strong><em> The system profits because winners outweigh losers — not because it’s usually right.</em></blockquote><p>One more finding deserves a spotlight, because it’s the kind of thing only a simulator can tell you. The committee reports a confidence score with every decision. Does that confidence <em>mean</em> anything? The replay’s answer was delightfully specific: <strong>at the low end, yes; at the high end, no.</strong> Decisions below the 65% confidence bar won only a third of the time — the bar is filtering exactly the junk it should. But <em>above</em> the bar, more confidence added nothing — in fact the 65–75% bucket won <em>more</em> often than the 75%+ bucket. The AI knows when it’s unsure. It does not know when it’s <em>extra</em> right.</p><h3>Part 4: What the Simulation Disproved</h3><p>This is the section that separates this project from the viral-screenshot genre. The replay didn’t just validate ideas — it killed several, including ones the authors liked.</p><p><strong>Killed: the “smarter exits” fix.</strong> An early replay diagnosed a real weakness — the committee sold winners too early. The obvious cure: replace AI-judged exits with a mechanical trailing stop-loss (an automatic “sell if it drops X% from the peak” rule — a classic, respected tool). The experiment was pre-registered, the year replayed twice, and the verdict came back backwards: mechanical stops returned <strong>+8.0%</strong> versus <strong>+11.0%</strong> without them. The stops kept getting shaken out by ordinary wobbles — selling on the dip, missing the rebound. The feature exists in the code today, tested and working. It ships <strong>switched off</strong>, because the evidence said off.</p><p><strong>Killed: loosening the confidence bar.</strong> If the system only buys at 65%+ confidence and mostly says HOLD, wouldn’t lowering the bar to 55% capture more good trades? Pre-registered, replayed, measured: returns collapsed from +13.3% to <strong>+5.2%</strong>, and the newly admitted trades won just <strong>33%</strong> of the time. The bar stayed at 65%. The registry row ends, in effect: <em>rejected, no change</em> — which is exactly what most experiment outcomes should look like in an honest lab notebook.</p><p><strong>Killed: the crypto ambition.</strong> The committee was pointed at Bitcoin on a fast 15-minute cycle. Over 233 decisions it went bearish 166 times, mostly sat in cash, and returned <strong>−0.02%</strong> while simply holding Bitcoin returned +3.37%. The diagnosis was structural, not bad luck: the committee’s edge comes from reading <em>company</em> signals — earnings, news, fundamentals. A coin has none of those, so most of the committee goes blind and the default kicks in: do nothing. Verdict: no edge, strategy parked. The system does not trade crypto — not because crypto is bad, but because <em>this machine</em> has nothing intelligent to say about it, and now there’s a measurement proving it.</p><p><strong>And one bug that changed the architecture forever.</strong> The earliest replays revealed something absurd: the system was bearish <strong>97% of the time</strong> — a broken pessimism baked into how a single AI read the world. That discovery is <em>why</em> the bull-versus-bear debate and the evidence-scoring Judge exist at all. The committee structure isn’t a cute metaphor; it’s a bug fix, with a before-and-after measurement to justify it.</p><h3>Part 5: The Simulator Rebuilt the Code Itself</h3><p>Here’s the least obvious payoff. Replaying a year of markets means running the full AI committee hundreds of times, unattended, for hours. That workload became the best <strong>stress test</strong> the code ever had — and drove engineering fixes in domains that have nothing to do with trading strategy.</p><p><strong>The “day-27 killer.”</strong> Long simulations kept dying about a month in. The cause: once in a while, the AI’s response comes back cut off mid-sentence — malformed, unreadable by the code. Rare enough to never surface in single-day tests; certain to occur somewhere in 700+ calls. The fix taught the system to survive and retry truncated responses. A live-trading crash was debugged before live trading ever saw it.</p><p><strong>The time-leak hunt.</strong> The scariest class of simulator bug is <em>future leakage</em> — any path where tomorrow’s information sneaks into today’s simulated decision, making results look better than reality. Two real leaks were caught and sealed: the simulated Portfolio Manager was found reading the <em>live</em> long-term memory database (which contained outcomes of trades from “the future” of the replay), and the market-regime detector was rebuilt to compute its readings strictly from data available on each simulated day. Both fixes are now pinned by automated tests, so the leak can’t quietly return. If those tests didn’t exist, every result in this article would be suspect — <strong>a simulator has to earn trust the same way the strategy does.</strong></p><p><strong>Surviving a bad week at the AI factory.</strong> Over one five-day stretch, the AI provider’s servers returned 48 overload errors. The system rode through all of them — retry with patient, randomized backoff; rotate to backup access keys; fall over to a different AI provider if a whole service is down; and a hard timeout so one hung call can’t freeze the day’s cycle. None of that resilience existed on day one. Marathon simulations made the fragility impossible to ignore.</p><p><strong>The learning loop that quietly starved.</strong> The system is supposed to learn: days after each trade, a “reflector” grades the decision against what actually happened and files a one-sentence lesson into long-term memory, which future debates can recall. Monitoring caught it silently broken — trades tagged with long horizons meant <strong>zero lessons were being written</strong> for weeks. The fix added interim report cards (grade the trade at 5 and 20 days, not just at the distant finish line). The lesson generalizes to any AI system: <strong>a feedback loop you don’t monitor is a feedback loop you don’t have.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/922/1*f6Z36MFM2biHoBiM2NEMMw.png" /><figcaption><em>The loop that quietly starved — and now gets monitored like everything else.</em></figcaption></figure><p>Through all of this, the safety net grew to <strong>900+ automated tests</strong> — small programs that each check one promise the code makes (“the daily-loss breaker survives a restart,” “a simulated decision can’t see the future”) and run after every change. For a non-coder, the honest analogy: it’s a pre-flight checklist with 900 items, executed by machine, every time anyone touches the plane.</p><h3>Part 6: The Money Boundary — Who Is Allowed to Pull the Trigger</h3><p>Strategy aside, the system’s sharpest logic is about <strong>whose finger is on the button</strong>. There are three ways to run it, and they differ in exactly one thing: what happens to an approved order.</p><ul><li><strong>Paper trading</strong>: the system trades <strong>autonomously</strong> — it decides <em>and</em> places orders, every day, unattended — but on a brokerage’s <em>paper</em> account: live market data, real order mechanics, <strong>simulated money</strong>. This is where the system currently runs, building a track record with zero dollars at risk.</li><li><strong>Preview</strong>: the full committee runs on real data and logs exactly what it <em>would</em> do. Nothing is ever placed. A dress rehearsal.</li><li><strong>Live (real money)</strong>: the machine is <strong>not allowed to trade</strong>. It reads the account, analyzes, and recommends — and then a human must place the order, personally, after typing an explicit APPROVE for that specific trade. The default answer is NO. There is no auto-approve setting to switch on. It doesn’t exist in the code.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DAS9LB3qZBMl0ZAZQA9fjw.png" /><figcaption><em>Same brain everywhere. What changes is whose money is at stake — and who is allowed to pull the trigger.</em></figcaption></figure><p>(For completeness: the equity committee isn’t the only strategy in the codebase. A second sleeve runs a conservative options-income strategy — “the wheel” — that sells insurance-like contracts on quality stocks under hard guardrails: always fully cash-backed, never naked, the Analyst holds a veto on every entry, and collateral is capped. It lives on the paper account only, since a single contract ties up $2,500+ of collateral, and it earns trust the same way everything here does: through a measured track record first. The third experiment, crypto, you already watched get parked.)</p><p>And even a human-approved real-money order still has to survive a gauntlet of deterministic checks — the Risk Officer again, plus safeguards born from unglamorous realities of US brokerage law. The standout example: <strong>T+1 settlement</strong>. When you sell a stock, the cash takes a business day to officially settle; reusing it too quickly triggers a “Good Faith Violation” that can get a small account restricted. The system keeps a ledger tracking the settlement status of every dollar and will block <em>any</em> trade that would violate it — even a stop-loss sale. A compliance rule outranks the smartest AI argument, every time.</p><p>There’s also a quieter piece of engineering here worth admiring: <strong>crash-safe orders</strong>. Before placing any order, the system writes its intent to a local database; after a crash or lost connection, it checks with the broker about that intent <em>before</em> doing anything new. This is the classic way trading bots accidentally buy twice — and it’s structurally impossible here.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mrSaag3nUSmp3GTuR17RIA.png" /><figcaption><em>A real-money order must clear every gate in a row. The human is the last and hardest one.</em></figcaption></figure><h3>Part 7: What It Costs, and What It Can’t Do</h3><p>The running cost is almost comically small: about <strong>one cent per decision</strong> using Google’s Gemini models — and $0 on the free tier, which caps out at 20 requests per model per day. A full-year historical replay costs a few dollars. The AI bill will never be the story.</p><p>Even here there’s a characteristic honesty note in the project’s own docs: the cost tracker was once <strong>five times too optimistic</strong> because it used an outdated price table. It was found, fixed, and documented rather than buried — in a project whose whole thesis is measurement, even the measurement of the measurement gets audited.</p><p>And the limits, stated as plainly as the wins:</p><ul><li><strong>It cannot see the future.</strong> Four in ten trades lose. The edge, where it exists, is a repeatable <em>process</em>, not clairvoyance.</li><li><strong>It will lag a roaring bull market</strong> while it holds cash cushions — that’s the documented price of −2% drawdowns instead of −15% ones. A choice, not a flaw. But a real cost.</li><li><strong>It has no day-trading, no intraday, and no crypto edge</strong> — the simulator checked, and the answer was no.</li><li><strong>Back-tests are lenient by nature.</strong> These replays used present-day company fundamentals frozen in time, assumed no price slippage on fills, and can’t fully rule out that the AI “remembers” 2025’s news from its training data. The authors list these caveats themselves — which is exactly why the current phase is autonomous <em>paper</em> trading on live data: it closes those loopholes one by one, still without risking a dollar.</li><li><strong>Simulation does not guarantee future returns.</strong> Neither does a good paper-trading streak. Nothing does.</li></ul><h3>The Takeaway: The Process Is the Product</h3><p>Strip away the trading and something more broadly useful remains — a working template for building <em>any</em> consequential AI system:</p><p><strong>Make AIs argue instead of answer.</strong> One model’s opinion had a 97% bearish tilt; adversarial debate plus an evidence-scoring judge fixed what no prompt tweak could.</p><p><strong>Let deterministic code hold the veto.</strong> The AIs propose; rules that cannot be sweet-talked decide what’s allowed. The blast radius of a bad AI day is capped by arithmetic.</p><p><strong>Pre-register your experiments, then obey them.</strong> Write the pass/fail rule before you run the test. Publish your rejections. This project’s registry has more “REJECTED — no change” rows than adopted ones, and that’s precisely why its adopted features are believable.</p><p><strong>Distrust your own simulator.</strong> Hunt the future-leaks. Pin the fixes with tests. A flattering referee is worse than none.</p><p><strong>Keep the human at the money boundary.</strong> Autonomy with simulated dollars, recommendation-only with real ones, and an explicit typed YES for every irreversible act.</p><p>The system’s most common output, remember, is HOLD — a machine that spent a year of simulated markets mostly declining to act, and made +13% anyway. There’s something almost philosophical in that: the intelligence here isn’t in the picking. It’s in the discipline of proving, one pre-registered replay at a time, when <em>not</em> to trust yourself.</p><p>*Not financial advice. All numbers are from historical simulations and paper trading with simulated money; both have known leniencies, and past results — simulated or real — don’t guarantee future returns. The human owns every trade and all of the</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8ceaf32ba6ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>