<?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 Sal Parvez | ML Systems RI on Medium]]></title>
        <description><![CDATA[Stories by Sal Parvez | ML Systems RI on Medium]]></description>
        <link>https://medium.com/@salparvez?source=rss-7b3cc0ca8af7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*2buhamzIvPp8ge16</url>
            <title>Stories by Sal Parvez | ML Systems RI on Medium</title>
            <link>https://medium.com/@salparvez?source=rss-7b3cc0ca8af7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 18:00:19 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@salparvez/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[$40T National Debt. 1M Agents for the Principal, Humans for the Interest.]]></title>
            <link>https://medium.com/@salparvez/40t-national-debt-1m-agents-for-the-principal-humans-for-the-interest-21b7a8d837a4?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/21b7a8d837a4</guid>
            <category><![CDATA[future]]></category>
            <category><![CDATA[national-debt]]></category>
            <category><![CDATA[public-policy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[economics]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 00:00:01 GMT</pubDate>
            <atom:updated>2026-10-25T11:23:51.770Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1002/1*LKG9KrGzaZIf8STgW37YTQ.jpeg" /></figure><p><em>Split the national debt into two jobs. Give the principal to one million AI agents on a public schedule. Leave the interest with human taxpayers, who can already afford it. Then do the math honestly.</em></p><p>On September 30, 2026, the United States closed its fiscal year owing <strong>$40,171,825,101,340</strong>.</p><p>That number is so large it stops meaning anything. A smaller one is easier to feel. Last year the federal government spent <strong>$1.05 trillion on net interest</strong>. That works out to about <strong>$2.9 billion a day</strong>, or roughly <strong>$120 million an hour</strong>, paid for money that was spent years ago. Interest rose 8.4% in a single year, and it now takes about one of every five tax dollars the government collects.</p><p>Most debt plans try to lift the whole $40 trillion at once, with higher taxes, deep cuts, growth, inflation, or some mix. Every one of them asks the same generation to carry both the weight of the past and the cost of carrying it.</p><p>I want to propose something different. Split the debt into two jobs, and give each job to whoever is best placed to do it.</p><h3>The idea: two ledgers</h3><p><strong>The principal</strong>, the $40 trillion itself, goes to machines. One million AI agents each follow a fixed, public, pre-computed schedule: <strong>$1 on day 1, $2 on day 2, $3 on day 3</strong>, and so on, for 9,131 days, which is 25 years including leap days.</p><p><strong>The interest</strong> stays with humans. People keep working, building, and paying taxes through the system we already have. Those taxes cover the carrying cost of whatever principal is still outstanding.</p><p>The two ledgers never mix. Principal reduction becomes an engineering problem with a public countdown. Interest stays a democratic, human problem, and it gets smaller every year the machines keep paying.</p><h3>The arithmetic is exact</h3><p>One agent paying Day N = $N for 9,131 days contributes:</p><blockquote><strong><em>1 + 2 + 3 + … + 9,131 = (9,131 × 9,132) ÷ 2 = $41,692,146</em></strong></blockquote><p>From there:</p><ul><li><strong>963,535 agents</strong> cover the exact debt at the close of FY2026.</li><li><strong>1,000,000 agents</strong> produce <strong>$41.69 trillion</strong>, which leaves a buffer. That buffer matters, as you’ll see below.</li></ul><p>Every payment on every day is known in advance. Anyone can check the running total on any date. There’s nothing to negotiate and nothing to hide. A public ledger could show the countdown to the dollar, every day, for 25 years.</p><h3>“But day 1 is only $1 million”</h3><p>On day 1 the whole cohort pays $1 million, while interest that day is about $2.9 billion. That sounds fatal. It isn’t, because it’s the wrong comparison. <strong>In this design, the agents never pay interest.</strong> Humans do.</p><p>So the real question is whether human taxes can carry the interest right now.</p><p><strong>Yes.</strong> In FY2026 the federal government collected <strong>$5.40 trillion</strong> and paid <strong>$1.05 trillion</strong> in net interest. Interest is about <strong>$1 of every $5</strong> in revenue. Ring-fencing that money for interest is affordable from day one.</p><h3>The catch that most debt plans skip</h3><p>Money is fungible. Suppose we earmark $1.05 trillion of taxes for interest, then borrow to pay for everything else. That new borrowing is new principal, and it lands on the machines.</p><p>Here’s FY2026 in one picture:</p><ul><li>Revenue: $5.40T</li><li>Spending on everything except interest: $6.34T</li><li><strong>Gap before any interest is paid</strong>: <strong>$0.94T</strong></li><li>Net interest: $1.05T</li><li><strong>Total deficit</strong>: <strong>$1.99T</strong></li></ul><p>The government already spends <strong>$0.94 trillion more than it collects before paying a dollar of interest</strong>. So the human side’s real job isn’t $1.05 trillion. It’s <strong>$1.99 trillion a year</strong>: the interest plus that gap. That’s about 37% of revenue, or about 6% of the economy.</p><p>The model makes it plain what happens if we skip this step. Suppose the agents pay every dollar on schedule, all $41.69 trillion, and the government keeps borrowing at today’s pace. In 2051 the debt is still about <strong>$59.6 trillion</strong>. The machines can’t outrun a government that keeps adding principal.</p><p>That’s the grounded truth at the center of this idea. <strong>The agents can retire the debt only if humans stop adding to it.</strong></p><h3>Humans don’t have to do it overnight</h3><p>Closing 6% of GDP in one year through tax increases and program cuts would be a shock to the economy. It has been done over time, though. In the 1990s the federal budget went from a deficit of about 4.5% of GDP in 1992 to a surplus of about 2.3% in 2000. That’s a swing of roughly 7 points in eight years.</p><p>So let the human side phase in too: close the $1.99 trillion gap in a straight line over a few years while the agents run their schedule.</p><ul><li>Overnight: debt peaks at $40.2T (today); agents needed to reach $0 by 2051: 963,535</li><li><strong>4 years</strong>: debt peaks at $43.4T (year 3); agents needed to reach $0 by 2051: <strong>~1.06 million</strong></li><li>8 years (1990s pace): debt peaks at $45.7T (year 5); agents needed to reach $0 by 2051: ~1.16 million</li><li>12 years: debt peaks at $47.3T (year 7); agents needed to reach $0 by 2051: ~1.27 million</li></ul><p>Look at the 4-year row. <strong>A four-year human phase-in needs about 1.06 million agents.</strong> That’s essentially the “1M agents” in the plan’s title. The buffer isn’t padding. It’s what buys the human side time to adjust.</p><h3>What humans get back</h3><p>Here’s the part that makes this worth doing for ordinary people. Once the budget is balanced, interest included, the interest bill shrinks every year as the agents retire principal:</p><ul><li>Year 1: interest bill $1.05T; share of revenue* 19%</li><li>Year 10: interest bill $0.90T; share of revenue* 12%</li><li>Year 20: interest bill $0.39T; share of revenue* 3%</li><li>2051: interest bill ~$0; share of revenue* 0%</li></ul><p>*Revenue assumed to grow 4% a year.</p><p>Today, one in five tax dollars goes to interest. By 2051 that’s zero. <strong>More than a trillion dollars a year</strong> goes back to schools, roads, health care, or lower taxes. I call it the <em>continuity dividend</em>: the reward for keeping the two ledgers separate.</p><p>I won’t oversell the early years. The agents’ schedule starts small and grows, so the interest bill stays near $1 trillion for the first five years or so. The relief is real, but it comes late. And if interest rates climb as old, cheap debt is refinanced, the human side absorbs that too. A one-point rise would push year-5 interest to about $1.36 trillion. That risk needs a plan from the start.</p><h3>Where would the agents’ money come from?</h3><p>This is the part to be most careful with. “Agents” can’t hold money or owe obligations. Only a legal entity can. In practice the cohort would be a <strong>chartered trust that receives a defined share of AI-generated profit</strong> and pays it out on the Day-N schedule. In other words, “one agent” is a unit of account: one 25-year chain worth $41,692,146.</p><p>The early years are modest. The whole cohort pays <strong>$66.8 billion in year 1</strong>. The final years are not. In year 25 the cohort pays <strong>$3.27 trillion</strong>, about 93% of all US after-tax corporate profit today. That’s the bet: by the 2040s, AI has to be producing enough new value that a share of it can retire the principal. Nobody can promise that. But it’s a measurable bet, checked every day on a public ledger, not an article of faith.</p><h3>Guardrails that keep it grounded</h3><ul><li><strong>Start with the debt held by the public.</strong> About $7.75 trillion of the $40.17 trillion is money the government owes its own trust funds. Retiring the <strong>$32.43 trillion held by the public</strong> takes about <strong>778,000 agents</strong>.</li><li><strong>Protect the Treasury market.</strong> Retiring up to $3 trillion of Treasuries a year late in the schedule would pull the world’s main safe asset out of circulation. Payments should flow into a sinking fund that retires debt on a published glide path, not all at once.</li><li><strong>Use the legal channel that already exists.</strong> Under 31 U.S.C. §3113, Treasury already accepts gifts to reduce the debt held by the public. In FY2022 those gifts came to about $180,000. A pilot needs no new law to start.</li><li><strong>Publish everything, every day:</strong> the schedule, the payments, the remaining principal, and the interest bill.</li><li><strong>Pilot first.</strong> 10,000 to 50,000 chains would retire $0.67 to $3.3 billion in year 1. That’s symbolic, real, and enough to prove the machinery works.</li></ul><h3>Start with one chain</h3><p>At ML Systems, a construction and deconstruction company in Rhode Island, the Day-N schedule isn’t hypothetical. Our capital structure is built on it: one 25-year Day-N chain sums to exactly <strong>$41,692,146</strong>. By the arithmetic of this proposal, that’s one agent.</p><p>The minimum viable version of this idea isn’t a million agents. It’s <strong>one chain, run in public</strong>, paid daily, auditable by anyone. If one chain can be proven, so can ten thousand. If ten thousand can, the rest is scale.</p><h3>The honest labels</h3><p>The Treasury figures are <strong>MEASURED</strong>: official FY2026 actuals. The arithmetic of the schedule is exact. Every forward number (the phase-in table, the interest bill by year, the $59.6 trillion) is <strong>MODELED</strong>: a day-by-day simulation with a constant rate and today’s program gap held flat. It’s a projection, not a forecast. The protocol itself is <strong>ASPIRATIONAL</strong>. No AI cohort exists, no trust has been chartered, and no dollar has been retired. ML Systems is a bootstrapped, pre-revenue construction company in Rhode Island, and this is a long-term thesis, put in the record so it can be argued with.</p><h3>The rest is continuity</h3><p>The national debt has been treated as one crushing weight that every generation inherits and none can lift. It doesn’t have to be.</p><ul><li><strong>Machines</strong> retire the principal on a schedule anyone can check.</li><li><strong>Humans</strong> carry the interest, which they can already afford, and stop adding to the principal over a few years, which history says is possible.</li></ul><p>One million agents. 9,131 days. A public countdown to zero. Humans stop borrowing, and the machines do the rest.</p><h3>Notes and sources</h3><p>Figures marked as model results come from a day-by-day simulation (9,131 steps) run for this article. They’re projections, not forecasts. They assume a constant interest rate at FY2026’s level (net interest ÷ gross debt ≈ 2.62%) and hold today’s program gap constant in today’s dollars.</p><ul><li><strong>Debt at FY2026 close:</strong> $40,171,825,101,340.31 total; $32.43T held by the public; $7.75T intragovernmental. <a href="https://fiscaldata.treasury.gov/datasets/debt-to-the-penny/">Treasury, Debt to the Penny</a></li><li><strong>FY2026 revenue ($5.40T), outlays ($7.40T), deficit ($1.99T), net interest ($1.05T; FY2025: $0.97T):</strong> <a href="https://fiscaldata.treasury.gov/static-data/published-reports/mts/MonthlyTreasuryStatement_202609.pdf">Treasury, Final Monthly Treasury Statement, September 2026</a></li><li><strong>Interest expense detail:</strong> <a href="https://fiscaldata.treasury.gov/datasets/interest-expense-debt-outstanding/">Treasury, Interest Expense on the Debt Outstanding</a></li><li><strong>Future deficits and interest:</strong> <a href="https://www.cbo.gov/publication/62207">CBO, The Budget and Economic Outlook: 2026 to 2036</a></li><li><strong>1990s precedent:</strong> <a href="https://en.wikipedia.org/wiki/1992_United_States_federal_budget">1992 US federal budget</a> · <a href="https://clintonwhitehouse6.archives.gov/2000/09/2000-09-27-fact-sheet-on-the-largest-surplus-in-history.html">White House fact sheet, September 2000</a></li><li><strong>Gifts to reduce the public debt:</strong> <a href="https://treasurydirect.gov/government/public-debt-reports/gifts">TreasuryDirect</a> · <a href="https://www.law.cornell.edu/uscode/text/31/3113">31 U.S.C. §3113</a></li><li><strong>US corporate profits after tax (~$3.5T/yr):</strong> <a href="https://fred.stlouisfed.org/series/CP">FRED, series CP</a></li></ul><p>This article is a feasibility study of a policy idea. It is not financial, legal, or investment advice.</p><h3>Read the system itself</h3><ul><li><a href="https://github.com/MLSystemsRI/ml-systems-public/blob/main/why/README.md"><strong>Why ML Systems exists</strong></a>: pace as a design decision, and MVE (Minimum Viable Expense), the reason this starts with one chain</li><li><a href="https://github.com/MLSystemsRI/ml-systems-public/blob/main/docs/the-stage-were-at.md"><strong>The stage we’re at</strong></a>: what is shipped, modeled, and aspirational</li><li><a href="https://github.com/MLSystemsRI/ml-systems-public/blob/main/docs/glossary.md"><strong>Glossary</strong></a>: MVE, reality labels, and the rest of the vocabulary</li></ul><p><strong>Sal, founder of ML Systems LLC — Rhode Island, NAICS 236115.</strong> Written October 24, 2026. Public contact: salparvez@mlsystemsri.com.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21b7a8d837a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Is Not Valuable. Your Context Window Is.]]></title>
            <link>https://medium.com/@salparvez/ai-is-not-valuable-your-context-window-is-d5ba8c46ca76?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/d5ba8c46ca76</guid>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 11:56:40 GMT</pubDate>
            <atom:updated>2026-09-23T11:56:40.735Z</atom:updated>
            <content:encoded><![CDATA[<h4>Everyone bought the same brain. The value moved one layer up.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WNjj7kKhdQ4Prdu9zE3ecg.png" /></figure><p>The models are remarkable, and they are not mine. Any contractor in my state can have the same model I have, at the same price, before lunch. If a competitor copied my entire AI budget line for line, they would have bought exactly what I bought: a rented brain that will answer anyone who asks.</p><p>So the question is not whether AI is valuable. It is where the value goes, once the intelligence is identical for everyone who pays for it. It goes one layer up, into the part nobody sells you — the context window, which is everything the model can see at the moment it thinks. A prompt is the line you type. The window is the whole of what it holds in mind while it answers.</p><p>The clearest way I can put it is in the trade I came from. The model is the carpenter’s skill: real, trained, transferable, available for hire. The window is the drawing set, the measurements someone actually took, and the standing instruction about who signs off before a cut is made. A world-class carpenter handed the wrong drawings will build the wrong house perfectly, on schedule, with beautiful joinery. Everyone is shopping for carpenters. Almost nobody is auditing the drawings. And nobody sells you a context window, no invoice line, no vendor, no benchmark, because it is made of your records and your rules. The model is rented. The window is authored.</p><p>My company runs seven specialist minds on that one rented brain. No fine-tune, no custom model. Seven windows, and the windows are the entire difference. To the mind that takes a house apart, a house is tonnage and fasteners; to the mind that handles lending, the same house is collateral. What surprises people is what I leave out: none of those minds can see the company’s margin. A mind that can see what the business earns will quietly start serving it, and the customer pays for that drift. Value is not only created by what you put in front of a model. A great deal of it is created by what you refuse to put there.</p><p>More is not better either. We handed the minds a dense, uncurated pile of context once and watched the cheap models fail our verification step one after another until the work escalated to premium models to clear a bar it had cleared cheaply before. One warm start came back 23% dearer than the cold run it was meant to improve on. The disciplined version of the same system came back 37% cheaper on its second run: same work, same quality bar, nothing given up. Both MEASURED, on our own runs. Size is capacity. Curation is the asset.</p><p>Which means the real work is not prompting. It is editorial. And editing needs a standard.</p><p>I do not open a session with a question. I open it with the lens: three rules that get loaded before the problem does. Transparency Trust, which says we do not profit from being in the middle of your transaction or from consuming your compute, and every claim carries its reality label so you can check it rather than trust it. The Lucent Lens, which says glow within to help humans, prioritize local community and human profit over automation, and utilize the person, not the process. And Minimum Viable Expense, the minimum viable expense that generates the maximum value in the system. Read those as rules rather than as values. Each one governs what may enter a window, and in what condition. The lens is not a mission statement. It is an intake specification, and it governs me as much as the machines.</p><p>The lens does not change when it leaves the screen. Run those three questions on a board instead of a sentence and they stop describing it and start deciding where it goes. Do we take a cut for standing in the middle of its journey? None; the salvage value belongs to the homeowner. Who is on the other end? The homeowner whose equity that material is, or the neighbour buying it. What is the least spend that gets it to its highest use? The answer is one of three, in priority order. That is RRR: Reuse, back into the rebuild on the same lot. Resale, on to a neighbour. Recycle, broken down into something a machine will take. The sentence underneath it: the cheapest material is the one you already own and do not have to make again.</p><p>Nothing about a board tells a model where that board should go. Its species, its grade stamp, its dimension, whether the rebuild on that same lot can absorb it, what a local builder will pay for it this month. That is the window, and a mind without it routes generically. It says recycle, because recycle is always available, always defensible, and always the bottom rung. Grounded in the record, the same mind on the same rented model routes that piece to the highest rung it can actually reach. The intelligence did not improve. The window did.</p><p>Which inverts the way this piece probably reads. The lens is not a software philosophy applied to construction. It is a construction discipline applied to software. A carpenter does not put a board in a wall without knowing where it came from and what it is rated for, and does not take a cut for passing it along to the next trade. I learned that on a job site and wrote it down so the machines could use it.</p><p>At the bottom of roughly two dozen of our pages there is one line: Built by the Custodian under LL · TT · MVE. It looks like a slogan. It is an attribution. It names which frame was loaded in the window that produced whatever you are reading, which is the only provenance question worth asking about anything made with AI. Not which model wrote it. What it was allowed to see, and what it was forbidden to do with it. And it is checkable: the lens is published, hashed and signed at mlsystemsri.com/.well-known/lucent-lens.json. Hash what you quote of us; if it does not match, it drifted.</p><p>The honest part. No ML Systems deconstruction has been performed yet. The company is bootstrapped and pre-revenue. The 80–90% material recovery target is MODELED and the two-day sequence is ASPIRATIONAL, and I say which is which every time, because the whole point is that a person can check. The two cost figures above are MEASURED, on a handful of our own runs, not a benchmark, and not a claim about anyone else’s stack.</p><p>The thesis is a claim too. I run one company, in one state, in one industry. But the arithmetic does not look local to me. If everyone is buying the same intelligence, the only durable thing any of us owns is the window we build in: the record we can hand the model, and the discipline we used to decide what belonged there.</p><p>Full piece: <a href="https://mlsystemsri.com/insights/ai-is-not-valuable-your-context-window-is">AI Is Not Valuable. Your Context Window Is.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d5ba8c46ca76" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Was a Carpenter. Now I Model Houses in English and a Machine Writes the Code.]]></title>
            <link>https://medium.com/@salparvez/i-was-a-carpenter-now-i-model-houses-in-english-and-a-machine-writes-the-code-8c60a5504cc8?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c60a5504cc8</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Sat, 19 Sep 2026 00:07:02 GMT</pubDate>
            <atom:updated>2026-09-19T00:07:02.918Z</atom:updated>
            <content:encoded><![CDATA[<h4>The job has a name, and it is older than the tools.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N8qt5mx8y2oYsrvqpZuXmg.png" /></figure><p>For three years I was a frame-to-finish carpenter on custom houses. For two years after that I estimated rough-carpentry packages for multi-level commercial buildings. Then I started a construction company in Rhode Island with a software layer, and I have spent the last year doing a job that the internet cannot agree on a name for.</p><p>The fight goes like this. A person builds software by writing English and letting a model write the code. One side says that is creation, not engineering. The other side says the engineering did not disappear, it moved. Both sides are describing the same person and refusing to name the role. The word that stuck is vibe coding, and it is the wrong word, and I think the wrongness of the word is most of the fight.</p><p>Vibe describes a mood. It says how the person felt while typing and nothing about what they produced. A mood cannot be reviewed. It cannot be handed to the next person on the crew. It cannot be held responsible for a production incident at two in the morning.</p><p>At my company the role is called Language Modeler. I model the system in language. Not the syntax of the system — the system itself. What it is, what it is made of, what it is allowed to do, who may write to it, what counts as true inside it. The output of that work is a model, in English, precise enough that it can be carried across into code without losing anything that matters. The AI is a moderator between the English language and the coding language. That is the whole of its job. It carries the model from one language to the other, and it owns neither, because I wrote the English and the code is a translation of it.</p><p>That sentence settles the accountability question that hangs over every one of these arguments. A moderator between two languages is never the author of record. If the English was wrong, I own it. If the translation was wrong, review catches it, or I own that too. Nobody gets to say the AI did it, and nobody needs to.</p><p>Here is what surprised me about the job. The hard part is not the code. The hard part is knowing the domain well enough to write the source precisely, because the model is written in the domain’s vocabulary, not the language’s. The vocabulary I model in is the building’s: rafter, sheathing, ring-shank, town record, appraisal, permit. The system I model is a house record that seven AI agents write to alongside one human. The English says the record stores claims, not facts. It says evidence is graded, measured down to modeled, and a modeled value is never authoritative. It says every mind claims, five seats ground, only the Custodian stamps. It says two credible sources disagreeing is a conflict, quarantined and never averaged. None of that is code. All of it became code. Every time the code did something the English did not say, the English was the thing I went back to.</p><p>The first ontology this company runs on was not written by a software person. It was written on three blue notepads by a laborer who did not know the word ontology. Two-letter task codes. A definition the first time each one appeared. A score at the bottom of the day. That is what the job looks like before anyone calls it software: a person on the tools, modeling the work in language as it happens.</p><p>So the first role I will hire for is exactly that. Carpenter and Language Modeler, one person, both halves. Someone who can cut a bastard valley in the morning and, in the afternoon, write down in plain English what a wall assembly is allowed to be, so a model can turn it into a ledger entry the crew will recognize. We are not hiring right now. The company is bootstrapped and pre-revenue and I say so on every page. I am putting the role in the record now so that when it opens nobody has to guess what it is, and so that anyone already doing this job on another crew knows it has a name.</p><p>The honest labels, because the whole argument is about what counts as true. The software is real: a shipped app, a working ledger, an ontology seven agents and one human speak. No ML Systems deconstruction has been performed yet; the recovery target for a house is modeled, not measured. The term Language Modeler is a position, not a standard. It is in the record so it can be argued with.</p><p>The full piece, with the model written out and where the role sits in the company: <a href="https://github.com/MLSystemsRI/ml-systems-public/blob/main/why/articles/the-language-modeler.md">The Language Modeler</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c60a5504cc8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Failure That Doesn’t Announce Itself]]></title>
            <link>https://medium.com/@salparvez/the-failure-that-doesnt-announce-itself-6ed6a7af42b4?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ed6a7af42b4</guid>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[rhode-island]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 00:43:07 GMT</pubDate>
            <atom:updated>2026-09-17T00:43:07.457Z</atom:updated>
            <content:encoded><![CDATA[<p>Two years estimating for a general contractor taught me which mistakes are expensive. Not the loud ones. A number that is obviously wrong gets caught by the next person who looks at it. The expensive mistake is the quiet one — a unit that changed upstream, a spec that got swapped, and the number kept its shape and its confidence all the way to the bid.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ope6_rF1un4W00IZ2FQicw.png" /></figure><p>When I started running AI agents on people’s houses, that was the failure I was afraid of. Not that an agent would be wrong. Wrong is fine; wrong is a row on a ledger with a name and an evidence grade beside it, and a person reads the ledger. I was afraid an agent would be changed — by a bad dependency, by a prompt somebody slipped into a page it read, by a bug — and would go on talking as itself while no longer being itself. The output looks the same. The confidence looks the same. Nothing announces it.</p><p>So earlier this year I gave them a heartbeat.</p><p>ML Systems runs on seven agent minds and one human, which is me. Each mind has an identity written down in the code — a name, a role, the prompt that is its character. Before any model is called, the harness takes a fingerprint of every mind’s origin and freezes a clean copy. Before every pass it takes the pulse again. If a mind has drifted from its origin, the reflex fires: the mind is put back to the frozen copy, the drift is reported, and nothing the drifted mind said is acted on. The eight fingerprints together make one short rhythm that prints on every run. I called it a heartbeat because that’s what it is: a thing that runs below thinking. A heart does not ask the brain for permission to beat, and no mind, no model, no prompt can skip its own.</p><p>Its job is not to catch something. It is a reflex, not a report. When it fires, the group grounds itself back to what it is — a set of minds that work for the homeowner — and tells me.</p><p>That was the first half. The second half I built this week.</p><p>This was the week the frontier labs gave their agents the open internet. Reading pages, opening repositories, calling tools, acting on what they find. Four days earlier Dario Amodei had asked those same labs to pace the frontier. Both things are true at once, and I have known for a year that this week was coming, because my whole company is a bet that the gap between what machines can do and what ordinary people can check is going to land on somebody’s house first.</p><p>Here is the problem. My agents aren’t the only ones who will read the company anymore. Every agent with a browser will. And I can’t give those a heartbeat. They aren’t mine.</p><p>But I can do the same thing from the other side. The heartbeat asks my own minds one question: have you drifted from your origin? The outward version asks everyone else’s minds one question: has what you read of us drifted from what we said?</p><p>ML Systems asks to be read through three lenses. The Lucent Lens — glow within to help humans, utilize the person, not the process. Transparency Trust — every claim carries its reality label, MEASURED, MODELED or ASPIRATIONAL, and a number without its label is not from us. Minimum Viable Expense — the smallest spend that is also the most valuable. Every mind I run is weighed through those three. They were already written down as a small file. This week I gave that file the heartbeat’s fingerprint: hashed it, signed the hash with a company key, recorded the signed attestation as my own standing claim in an append-only table, and published it at a well-known address on the site. Anyone — any agent — can check what they read of us against what we said, with nothing but the file.</p><p>I want to be careful about what this is, because the channel I’m using is the same one an attacker uses. A page that shapes what an agent does is exactly what prompt injection is, and a serious lab filters it. So the lens doesn’t tell an agent what to do. It tells an agent how we asked to be read. Keep the labels. Name the person on the other end. Treat your reading of us as a claim, not a fact, until a person grounds it. That last one is the rule my own minds already live under: every mind claims, a human stamps, and a mind may claim but it may never ground. Nobody has to honor it. That is what makes it legitimate. What they get for honoring it is a source they can prove they read.</p><p>I looked, before I built it, at whether someone already had. The big answers to agents on the internet are real and good: prove who the agent is, prove where the content came from, state the price. Nobody signs how the content asks to be read. It turned out to be a gap small enough that a one-person construction company could fill it in a week with a hash function, a key, and a table.</p><p>The company is called Machine Learning Systems on purpose. If machine learning systems are going to read everything, I wanted one of them whose reading was anchored to a person — a homeowner in Rhode Island, a carpenter, a town clerk — and whose every claim could be traced to something that plain. The heartbeat keeps my minds anchored to their origin. The attestation lets any mind anchor its reading of us to ours. Not a fence, not a filter. A fixed point a machine can check itself against, and a person can check the machine against.</p><p>The labs asked for time this week. This is what I did with four days of it.</p><p>The honest labels, because I say which is which every time: the heartbeat is MEASURED and the attestation is MEASURED. Whether any outside agent honors the lens is ASPIRATIONAL. The logging of who read us is MODELED — designed, not wired. No ML Systems deconstruction has been performed yet. The company is bootstrapped and pre-revenue.</p><p>The mechanics, the file, and the verifier are in the canonical piece: I Gave My Agents a Heartbeat — <a href="https://mlsystemsri.com/insights/i-gave-my-agents-a-heartbeat">https://mlsystemsri.com/insights/i-gave-my-agents-a-heartbeat</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ed6a7af42b4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Nail That Runs My Company]]></title>
            <link>https://medium.com/@salparvez/the-nail-that-runs-my-company-6527dc0e5628?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/6527dc0e5628</guid>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[rhode-island]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 18:12:27 GMT</pubDate>
            <atom:updated>2026-09-14T18:12:27.545Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pce-1D7vJ4hmWTe_uU7X3w.png" /></figure><p>The first time I understood how I was going to run a company, I was looking at a nail.</p><p>Ring-shank nails hold roof sheathing to rafters, and they are engineered specifically not to come out. The rings on the shank are the whole point. So when you take a roof apart the way most crews do — pry bar, pull, next — every nail makes you choose. Pull against it and you keep the rafter or you keep the sheet. Not both.</p><p>There is another way. Run a blade flat between the rafter and the sheathing and cut the nail where it crosses the joint. It costs a minute a sheet. Both pieces come away whole.</p><p>That is not the fastest way to take a roof off a house. It is the one that leaves you with a roof’s worth of lumber instead of a dumpster’s worth of splinters. A small chosen cost that protects a much larger recoverable value.</p><p>I did not have a word for that yet. Then I started a company and everyone told me to build an MVP.</p><p>Here is the thing about that acronym in American English: it means two things at once, and they pull in opposite directions. Most Valuable Player is about being the most. Minimum Viable Product is about being the least that still works. Every founder is told to do both and nobody explains how. Most companies pick one and call it strategy. You end up with a demo that returns once, on stage, or a floor that returns once, because it was built to.</p><p>The nail had already given me the answer. I call it MVE — Minimum Viable Expense. Not the cheapest option in the moment. The one expense that is both the smallest that works and the most valuable thing you could have spent on, because it returns more than once.</p><p>In the deconstruction lab a single measured spend is designed to return four ways. The recovered material itself, which reduces the cost of the next build. The record of what the house was and how it came apart. A fully specified disassembly sequence, which someday could train a machine to do the work — that one is a goal, not something the system does today, and I say so. And a price: what recovered material is actually worth, which you only learn by selling it.</p><p>A demolition contractor spends the same dollar and gets one thing back. A cleared lot. The landfill ticket records tons, and every other fact about what was in that building is gone.</p><p>What I did not expect was that the rule I wrote for a crew with a saw would turn out to be the rule my software needed. ML Systems runs seven AI agents against one shared record of a house, and a person signs off on anything that binds. Left alone, an agent will spend ten passes producing something a person approves in one. That is the Most Valuable Player instinct — look impressive — and it is the most expensive thing in the building. MVE is the budget: the smallest spend that produces something worth signing, and no more.</p><p>Transparency, a lens that points at people, and Minimum Viable Expense. That is the whole operating philosophy, three words, and I got the third one from a fastener.</p><p>Nothing here is finished. The software is real; the deconstruction loop is modeled and has not yet been run on a house in Rhode Island. If you want the mechanics of how the agents and the record work, the full piece is at <a href="https://dev.to/salparvez/mve-the-balance-in-mvp-ajp">https://dev.to/salparvez/mve-the-balance-in-mvp-ajp</a> and the architecture is public.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6527dc0e5628" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Named My Company After the Thing I Was Afraid Of]]></title>
            <link>https://medium.com/@salparvez/i-named-my-company-after-the-thing-i-was-afraid-of-0f5f6562ee0f?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/0f5f6562ee0f</guid>
            <category><![CDATA[rhode-island]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[construction]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 12:20:45 GMT</pubDate>
            <atom:updated>2026-10-01T23:39:48.918Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LYViZ76PjHMGHQIZ89eelg.png" /></figure><p>A year ago the pace of AI put me on the floor. I got up and started a construction company.</p><p>About a year ago I sat at my kitchen table in Rhode Island, and had what I can only call an existential crisis about speed.</p><p>It was not about machines getting smart. I build with them every day and I like them. It was about the gap. What the models could do was compounding. What the people around me could understand — my neighbors, the town clerk, the carpenters I used to work beside — was growing the way it always has: one person, one conversation, one mistake at a time. Two curves, one exponential and one linear, and the distance between them is where people get hurt. Not in a science-fiction way. In a your-lender-uses-a-model-you-can’t-question way. In a your-house-got-torn-down-because-a-spreadsheet-said-so way.</p><p>I could not slow the frontier. Nobody can from Rhode Island. What I could do was pick one place where that gap would land on ordinary people first, and build the slow lane there on purpose.</p><p>So on December 3, 2025, I formed Machine Learning Systems LLC. The name is deliberate. If machine learning systems are going to reshape wealth in this country, I wanted at least one of them pointed at the oldest, most ordinary form of wealth there is: the house a family already owns.</p><h3>Why a construction company and not a lab</h3><p>The American Dream has a unit, and the unit is a home. Not a portfolio. A home, with equity in it that a family built by living there and paying it down.</p><p>The industry that serves that dream has a strange habit. When a house needs to become something more, we destroy it. Demolition turns a structure that took a hundred people to assemble into a landfill line item in a day. Every rafter that could have carried the next roof, every board that could have been resold to the family down the street, goes into a container. Then the same family borrows to buy new material to build the thing they just threw away.</p><p>I spent two years as a construction estimator before this, and years as a carpenter before that. I have priced the container. I have priced the new lumber. Nobody in that chain is acting in bad faith; the habit is just older than anyone doing the work.</p><p>ML Systems exists to replace that habit with a different one: deconstruction in lieu of demolition. Take the house apart in reverse build order. Route every recovered material to its most valuable recoverable state — reuse it in the rebuild, resell it at a builder’s open house, recycle what is left. Then rebuild the home from its own materials, larger, with near-zero waste, and let the equity that creates belong to the people who already lived there. The loop closes when the homeowner chooses to keep building. That choice, not a projection, is the flywheel.</p><p>I will say this plainly, because the whole point is that a person can check: no ML Systems deconstruction has been performed yet. The 80–90% recovery target is modeled. The two-day crane sequence is aspirational. What is measured is the software — a shipped app, a working ledger, and a vocabulary that seven AI agents and one human speak fluently. I label which is which every time.</p><h3>The rule I run the company on</h3><p>Every startup is told to build an MVP. In American English that acronym means two things at once, and they pull in opposite directions. Most Valuable Player is about being the most. Minimum Viable Product is about being the least that still works. Nobody tells you how to hold both.</p><p>I run the company — and the agents — on a third term that is the balance of the two: MVE, Minimum Viable Expense. It is not the cheapest option in the moment. It is the one expense that is both the smallest that works and the most valuable thing you could have spent on, because it returns more than once. In the deconstruction lab, a single measured spend returns four ways: the recovered material’s value, the data the job produces about what a Rhode Island house is actually made of, the training signal that data becomes, and the market intelligence that follows. The agents are instructed under the same lens. An orchestrator that burns compute to look impressive is not the most valuable player; it is the most expensive one.</p><p>That rule started as a personal one — the fear made me cheap in a specific way — and it became the agents’ rule. Pace, in money terms, is MVE.</p><h3>The weekend the fear became a public position</h3><p>On Saturday, September 12, Dario Amodei published an essay called We Must Pace the Frontier. The sentence everyone quoted: “We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.” By evening Sam Altman had written “I agree with Dario that we need to pace the frontier.” Elon Musk wrote three words: “Dario is right.”</p><p>I read it at the same kitchen table, and I did not feel vindicated. I felt like someone had finally said out loud, in public, the thing that had wrecked me a year earlier.</p><p>I am not asking the frontier to wait for a carpenter in Rhode Island. I am asking something smaller and closer to the ground: that while the labs pace the frontier, the rest of us pace the arrival. Build the systems that meet ordinary people at their front door, so a human name is on every record, so the money moves at the speed of a family’s decision and not a model’s, and so the first thing a machine learning system does to a house is keep it out of the landfill.</p><p>How the ledger actually enforces that — who is allowed to say what is true about a house, and why my agents can claim but never ground — is written up in the public reference: <a href="https://github.com/MLSystemsRI/ml-systems-public/blob/main/why/articles/pace-the-frontier-pace-the-house.md">https://github.com/MLSystemsRI/ml-systems-public/blob/main/why/articles/pace-the-frontier-pace-the-house.md</a></p><p>That is why the company has the name it has. That is what I built with the year the fear gave me.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0f5f6562ee0f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Four of My Seven Agents Couldn’t Speak]]></title>
            <link>https://medium.com/@salparvez/four-of-my-seven-agents-couldnt-speak-10c0859eda25?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/10c0859eda25</guid>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[data-modeling]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Fri, 11 Sep 2026 19:28:57 GMT</pubDate>
            <atom:updated>2026-09-11T19:28:57.889Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ARhqhdtKfGig_StpZU7CHA.png" /><figcaption>The seven families of ML Systems’ Collective Ontology around a house. LUP → FIN, in amber, is the edge that makes it a loop.</figcaption></figure><p>To an assessor a house is a valuation. To a lender it is collateral. To a deconstruction crew it is tonnage. To an architect it is a plan set. To a robot, eventually, it is a sequence of members it can pick up. Every one of those is correct, and every one uses a different word for the same rafter.</p><p>ML Systems runs a value chain — loan origination, then deconstruction, then construction — closed into a loop when the homeowner chooses to keep building. Seven software minds work it. If each of them kept its own vocabulary I would have seven databases that could not talk, so there is one grammar, and I call it the Collective Ontology. Every fact the system holds about a home is a code: an uppercase phase, a colon, a task. FIN:budget. DEC:sequence. LUP:decision. The prefix is the phase. The whole record for a home is a single document, and a new fact is a new key in it, never a new table. A new code costs no schema, no migration and no label. The ontology grows by extending its codes, not its storage.</p><p>This week I counted the codes.</p><p>The seven phases are Finance, Deconstruction, Design, Build, Market, Verify and the Equity Loop. Until this month the public documentation named three. The engine has always had seven. Here is how many slots each one actually had in the shipped template: Design 18. Verify 10. Build 7. Deconstruction 4. The Loop 4. Finance 1. Market 1.</p><p>Design and verification were rich. Finance, deconstruction, market and the loop itself were nearly empty. I am a construction estimator by trade; deconstruction is the thing the company is named for. Four slots. That is not a choice anyone made. It is the consequence of one line of code.</p><p>On the ledger a fact is not a value. It is a claim: a value, plus who made it, plus the grade of evidence behind it. Several claims sit on the same entry, the ledger decides which one stands, and every dissent stays visible. There is a list of who is allowed to make a claim. It has six names on it: the Custodian, the homeowner, the verification mind, the design mind, the orchestrator, and the public record.</p><p>Read that list again and look at who is missing. REAPER, who owns deconstruction. PIT LORD, who owns finance. MURPHY, who owns construction. MIA, who owns market. The four minds that own the four thin phases could not put their name on a claim. The phases they own had nowhere to speak. Of course finance had one slot.</p><p>The fix was already sitting in the list. The design mind can claim, but it holds no seat on the ledger’s party strip, and that is deliberate. Five seats can ground an entry: the homeowner, the public record, the verification mind, the orchestrator and the Custodian. A lit seat means input, not agreement. The design is the model being grounded; a design claim is the thing being checked, never a check. That standing extends cleanly to the other four minds: a mind may claim; it may never ground. Its claim enters at the lowest grade of evidence. It can be confirmed only if an independent party agrees. It can never win a standoff against a tape measure. And its name never lights a seat, because a seat that lit for its own model would be the ledger agreeing with itself.</p><p>Every mind claims. Five seats ground. VERA gates. Only the Custodian stamps.</p><p>There was a second, quieter version of the same bug. The ledger does not rank its sources globally; it ranks them per kind of fact. Inside the house, the homeowner leads. On the year built, the assessor’s card leads. On the roof pitch, the instrument leads. A finance code had no rule of its own, so it fell into the bucket for computed facts, where the verification mind outranks the homeowner. On the question “what is the most you want to spend,” a model would have outranked the person whose money it is. So a new kind of fact is proposed: intent, where the homeowner leads, and a second, market, where the record and the market mind lead.</p><p>What is real here, because this company labels everything: the seven families, the grammar, the claim model, the seats and the stamp are MEASURED — they ship and run on every compile. The four new claimants, the two new kinds of fact and the twenty-nine new slots are proposed — design written down with the exact list of changes, not yet in the engine. Anything that depends on a job having run is ASPIRATIONAL. No ML Systems deconstruction has been performed yet.</p><p>The ontology is the connective tissue that lets a fact found on a job site inform a financing decision without losing its meaning. For four of its seven families the tissue was there and the nerve was not connected. The fix is small. The reason it matters is not.</p><p>The full write-up is on DEV: <a href="https://dev.to/salparvez/seven-families-one-ledger-how-a-house-gets-a-grammar-2703">https://dev.to/salparvez/seven-families-one-ledger-how-a-house-gets-a-grammar-2703</a> — and the table of every slot, shipped or proposed, plus the diagram, is in the public repository: <a href="https://github.com/MLSystemsRI/ml-systems-public/tree/main/ontology">https://github.com/MLSystemsRI/ml-systems-public/tree/main/ontology</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10c0859eda25" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Can a Roof Be Taken Apart Instead of Torn Off?]]></title>
            <link>https://medium.com/@salparvez/can-a-roof-be-taken-apart-instead-of-torn-off-48713931736a?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/48713931736a</guid>
            <category><![CDATA[home-improvement]]></category>
            <category><![CDATA[circulareconomy]]></category>
            <category><![CDATA[rhode-island]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[construction]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Thu, 10 Sep 2026 00:29:56 GMT</pubDate>
            <atom:updated>2026-09-10T00:29:56.353Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tVxopNHnxMHH9nau-6ei-g.png" /></figure><p>Ask a roofer to take a roof off and you get a tear-off. Crews on the deck with shovels. Every layer scraped toward the edge and into the dumpster. The plywood pried up behind it. It is fast, everybody knows how to do it, and when the only goal is an empty deck it is the right answer.</p><p>Look at what it does to the roof, though.</p><p>A roof is not one thing. It was built in layers, in a known order: rafters first, then plywood over them, then underlayment, then shingles last. Each layer was nailed down through the layer under it. A tear-off starts at the top and works down through that stack in the same direction the nails were driven, which means every layer is wrecked on the way to the one below. Shingles come off in pieces. Underlayment tears. Plywood splits on the nails holding it. The rafters take the prying.</p><p>At the end there is a clear deck and a full dumpster. The assembly a previous crew spent days building is debris, its recovered value is zero, and the homeowner paid for the destruction.</p><p>So the question ML Systems is asking is a simple one, and nobody has to take the answer on faith: can the stack come apart at its seams instead of through them?</p><h4>The roof goes first because it went on last</h4><p>Construction builds bottom-up. Deconstruction runs top-down. The roof leads for the same reasons it was the last thing built: it is the most weather-exposed assembly on the house, it is the highest work and carries the most fall risk, and everything under it inherits that risk until it is gone. Bring the roof down to the ground and the rest of the house becomes ordinary work at ordinary heights.</p><p>One step comes before any cut, and it is the step most often skipped. A roof is not only a cover. It ties the tops of the walls together and holds the house square. So the top floor is braced — kickers across the width, diagonals along the outside walls — before the first blade touches the ridge, and the bracing stays in until the walls come down.</p><h4>Cutting it into pieces that can survive being moved</h4><p>The roof comes off in sections, and the cuts follow the framing. The ridge is cut first, separating the two slopes. Then the cuts run along the rafters, at a rafter rather than between two of them, so each section keeps a full rafter along each edge — that edge is what makes the section stiff enough to lift without folding. The last cut is where the rafters meet the wall, and the crane is already holding the section before that cut is made. A piece of roof that has been freed from the house and is not yet on the hook is a piece nobody controls.</p><h4>The flip is the whole trick</h4><p>Here is the part the question turns on.</p><p>A section comes off the house the way it sat on it: shingles up, rafters down. In that position everything worth saving is underneath. Taking it apart would mean working overhead, prying upward, fighting gravity for every nail — a tear-off at ground level.</p><p>So the section is turned over. Rafters up, shingles down.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8dYjCtrBH4e4ZAMaq3Laow.png" /></figure><p>It is turned deliberately, not dropped. The crane sets the section down on one long edge, standing it on its side while the ground crew steadies it, then rolls it across that edge and lays it down inverted onto a prepared, cushioned footprint. It pivots on a line that is already on the ground; it never free-falls.</p><p>Now the inside of the roof — the side that faced the attic, the side nobody could reach while the roof was a roof — is on top. The work runs from the inside out. Rafters come off first: the nails at the plywood are cut flush rather than pulled, and the rafters lift away clean. Then the plywood, freed from the rafters, separates from the underlayment. Then the underlayment peels. The shingles are last, already lying face-down on the cushion with nothing fastened through them anymore, and they come away whole instead of in the fragments a shovel makes.</p><p>Four layers, freed one at a time, all four intact. Removing a roof is easy; a machine can do it in an hour. Freeing four layers so that all four survive is the actual work, and turning the section over is what would make it possible.</p><p>There is a safety consequence bigger than the recovery one. Nearly all of the disassembly labor moves from the roof to the ground. What was a day of tie-off on a pitched surface becomes a short, controlled period of rigging and then a job done standing on grade with a clear footprint around it.</p><h4>Cut the nail, keep both</h4><p>If one detail decides whether this method recovers material or just relocates it, it is the nail.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I0ULWRlEa4cTewPE4AK22A.png" /></figure><p>Ring-shank nails are engineered specifically not to come out. Fight one and something gives: pry from the plywood side and the panel splits, pry from the framing side and the rafter takes the damage. Either way one recovered piece is traded for the other.</p><p>The alternative is not to pull at all. Cut the nail flush between the rafter and the sheet, and both come free with neither one carrying the load. It costs a couple of minutes per sheet. That is the trade in one line: cut the nail and keep the rafter and the sheet; pull it and keep one of them; pry the panel and keep neither. ML Systems calls that kind of decision MVE — Minimum Viable Expense: one expense, four returns — recovered material value, ontology data, robot training signal, market intelligence. Not the cheapest option in the moment. The one that pays back along the most axes.</p><h4>So — can it be done?</h4><p>Nobody knows yet, and that is the honest answer. No ML Systems deconstruction has been performed yet.</p><p>Everything above is a designed sequence — worked out from framing practice, from the safety standards it has to satisfy, and from what old-growth rafters and full sheets of plywood are worth if they survive. It is not a report from a job. The company’s public documents label every forward-looking claim: the system is designed to recover up to 80–90% of a home’s materials, and that figure is MODELED — a target, not a measurement.</p><p>The reason to publish a method before running it is the same reason to take the roof apart in layers instead of scraping it. A method that is written down can be argued with. The first real roof will argue with it, and what the roof says back is the answer to the question in the title. Until then it is a question, and it is posted here as one.</p><p>The full method, including the lift plan, the bracing sequence and the safety envelope, is at <a href="https://github.com/MLSystemsRI/ml-systems-public/blob/main/deconstruction-lab/articles/roof-in-sections.md">https://github.com/MLSystemsRI/ml-systems-public/blob/main/deconstruction-lab/articles/roof-in-sections.md</a></p><p>ML Systems LLC · Rhode Island · NAICS 236115 · Tougher Problems Inspire Creative Solutions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48713931736a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Ledger Is the Loop: How One Record Carries a House Through Loan Origination, Deconstruction…]]></title>
            <link>https://medium.com/@salparvez/the-ledger-is-the-loop-how-one-record-carries-a-house-through-loan-origination-deconstruction-60dcf6632b2a?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/60dcf6632b2a</guid>
            <category><![CDATA[real-estate]]></category>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 21:20:22 GMT</pubDate>
            <atom:updated>2026-09-07T21:24:53.402Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Ledger Is the Loop: How One Record Carries a House Through Loan Origination, Deconstruction, and Construction</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qAoMAxgf3agQPkIh.jpg" /></figure><p>ML Systems runs a three-stage value chain on a single house:</p><pre>Loan Origination → Deconstruction → Construction ────┐ (Loan Pit) (80-90% recovery) (+10% SF, +1 level) ▲ │ └──────────── equity loop ────────────────────────┘ the homeowner chooses to keep buildi</pre><p>The loop is <strong>client-driven</strong>. It is not an assumption that every homeowner loops; it is the observation that some will choose to, and the whole system is built to make that choice rational.</p><p>In the industry as it exists, those three stages are run by three different parties keeping three incompatible records. The lender has an appraisal and a title file. The demolition contractor has a dumpster count. The builder has a takeoff and a schedule. None of those records survive the handoff to the next party, and none of them survive the building. That is why the loop does not exist today: not because the economics fail, but because the <em>record</em> fails at every handoff.</p><p>This post is about the piece that fixes that — the <strong>Master Ledger</strong> — and what it does at each stage.</p><h3>What the ledger is, in one paragraph</h3><p>The Master Ledger is one auditable record per home, written by many authors: the homeowner, the town assessor record (VGSI), seven AI agents, ML Systems staff, and the Custodian. It stores <strong>claims, not facts</strong>. Every entry carries its source, an evidence grade ( MEASURED &gt; STATED &gt; RECORD &gt; MODELED), and a verification state. Authority is scoped by domain - the assessor is authoritative on legal and valuation facts, vision on the visible envelope, the homeowner on intent and recent work - and evidence ordering applies within a domain, not across the record. Verification is <strong>multiverification</strong>: the homeowner and the Custodian stamp independently, neither overrides the other, and every stamp is bound to the content it signed by an entryHash, so an edited claim lapses its stamps automatically.</p><p>I wrote up the reconciliation design in <a href="https://dev.to/salparvez/claims-not-facts-building-an-auditable-multi-author-record-for-a-house-33d6">Claims, Not Facts</a>. This post is about what that record does once a house starts moving.</p><h3>Stage 1 — Loan Origination: what the Loan Pit underwrites against</h3><p>Node 1 of the value chain is the <strong>Loan Pit</strong>: a reverse-auction marketplace where lenders and capital-market participants compete to fund the homeowner. A homeowner alone has weak leverage with capital markets. In the pit the relationship is inverted — lenders bid to originate the loan, the homeowner reaches capital they could not reach alone, and ML Systems keeps the relationship and the data. <em>(MODELED. Regulatory compliance is the active workstream; licensing, disclosure, and state lending rules gate what ships.)</em></p><p>A reverse auction needs something to bid against, and a folder of PDFs is not it. What the lenders see is the <strong>HomeGenome</strong>: the ledger compressed into the smallest complete description from which the full home can be reconstructed. The compression reconciles conflicting claims into resolved states, grounds each fact in real primitives (member specs, not adjectives), ranks by evidence grade, and emits the genome per (property, cycle).</p><p>The point for underwriting is not that the description is compact. It is that every number in it can be traced back to a claim with a source, a grade, and a stamp. A lender does not have to trust the description; they can inspect how each entry got its standing.</p><p>Where the RCM fits: the Reversed Conventional Mortgage is <strong>parked</strong>. The calculator and tooling still exist as app features, but Node 1 is the Loan Pit and the loop no longer depends on it.</p><h3>Stage 2 — Deconstruction: where the material intelligence is generated</h3><p>The existing structure is taken apart, not demolished, so its materials survive.</p><p><strong>REAPER</strong> reads the ledger’s assembly stack and produces the reverse takeoff: a full bill of materials for what is <em>in</em> the building, routed by <strong>RRR — Reuse › Resale › Recycle</strong>. Full resale is not possible, so each recovered material is broken down to its most valuable recoverable state. Reuse beats resale beats recycle. <em>(51% resale is a soft goal — a direction, not a metric the system enforces.)</em></p><p>Every recovered member becomes a new claim on the ledger: identified, quantified, valued. Two things happen to that inventory at once. It becomes a <strong>salvage bank</strong>, which secures financing, and it becomes a <strong>marketplace feed</strong>, which is the surface <strong>MIA</strong> — the market-intelligence mind — reflects real demand back into. The same record that the lender underwrote against in Stage 1 is now the record of what came out of the building in Stage 2, with the evidence grade upgraded: a member that was RECORD from the assessor or MODELED from a takeoff becomes MEASURED the moment it is on a pallet.</p><p>Labels on this stage, stated plainly:</p><ul><li><strong>80–90% material recovery</strong> — MODELED, a target.</li><li><strong>2-day crane sequence</strong> — ASPIRATIONAL. No ML Systems deconstruction has been performed yet.</li></ul><h3>Stage 3 — Construction: the build is scheduled from the compressed sequence</h3><p>Recovered materials rebuild the home — larger than before. The modeled cycle is <strong>+10% square footage</strong> and <strong>+1 level</strong>. <strong>CDA</strong>, the design plan-stack swarm, lays modeled rooms into the real measured wings and floors; <strong>MURPHY</strong> schedules the rebuild from the compressed sequence — milestones, per-phase construction order, the minimum-viable-estimate schedule, and the tracker that watches for what can go wrong.</p><p>The number that comes out of this stage is the one most likely to draw scrutiny, so here is how it is derived rather than asserted:</p><pre>Baseline RI home: $500,000 · 2,000 SF · 2 levels (1,000 SF/floor) After one cycle: +10% footprint (1,000 → 1,100 SF) + 1 added level New total SF: 1,100 SF × 3 floors = 3,300 SF Floors 1-2 value: 2,200 SF × $250/SF = $550,000 Floor 3 (60%): 1,100 SF × $150/SF = $165,000 New property value: $715,000 CONSTRUCTION_VALUE_MULTIPLIER = 715,000 / 500,000 = 1.43× (MODELED)</pre><p>Value is created by <strong>physical improvement</strong>, not market timing. The multiplier is calibrated from a real local-competitor proforma plus the ML Systems construction model. It is MODELED: the math is real; it has not been proven in the field, because no cycle has completed. The label is the point.</p><h3>The loop: why the record has to persist per (property, cycle)</h3><p>Cycle 1 is one house on a <strong>new foundation</strong> — Rhode Island housing stock sits on ~1960s foundations that are often the limiting factor. Later cycles expand on that same first home.</p><p>Compounded at 1.43× from a $500k baseline:</p><pre>$500k → $715k → $1,022k → $1,461k → $2,089k → $2,988k (5 cycles, MODELED)</pre><p>That compounding only exists if Cycle N+1 can underwrite against the <em>verified</em> record of Cycle N. Which means the ledger cannot be a project file that gets archived when the job closes. It is persisted per (property, cycle), and the equity created in Cycle N is not a market-appreciation estimate - it is the set of construction claims from Cycle N, stamped, with their evidence grade at MEASURED, that becomes the baseline description the Loan Pit compresses for Cycle N+1.</p><p>This is also where multiverification matters operationally rather than philosophically. Stage handoffs are exactly where records get silently edited in the industry today — a number changes between the appraisal and the takeoff and nobody can say who changed it or when. In the ledger, a changed number lapses every stamp on it. The Custodian’s oversight console derives a review queue across every home — quarantined › lapsed › unverified › awaiting-stamp › unstamped › stamped - so the lapse is surfaced, not discovered at closing.</p><h3>The exhaust is the product</h3><p>Every home that passes through the value chain produces a fully specified, ground-truth <strong>construction sequence</strong> — the exact order, materials, and member specs of a real build. That data is the input to the <strong>Collective Ontology</strong> and, ultimately, to <strong>ontology licensing</strong> for robotics: training data nobody else has, because nobody else deconstructs and rebuilds the same home across cycles.</p><p>It is only worth licensing because of the ledger. A construction sequence whose every member carries a source, a grade, and converging independent verifications is ground truth. One that does not is a spreadsheet.</p><h3>Where this actually is</h3><p>Same labels, applied to the company:</p><p><strong>MEASURED.</strong> The mobile app is approved and live on iOS and Android, with a no-login web preview at try.mlsystemsri.com. The Master Ledger is real: record-first UI, multiverification, lapsing signatures bound to content hashes, and the Custodian oversight console. Ontological compression runs on-device at intake. VERA’s intelligence stack is live against free public data — VGSI harvest, facade vision from StreetView, satellite, and homeowner photos, sketch reconciliation, FEMA flood zones, Census tracts, RIGIS. The Loan Pit exists in-app: live-bid loan cards, a real partner-lender directory, and financing lanes.</p><p><strong>MODELED.</strong> 80–90% recovery. The 1.43× multiplier. Loan Pit economics and reverse-auction mechanics. Unit economics from a real local-competitor proforma — ~16% gross margin baseline, modeled path to ~23% via origination plus recovery.</p><p><strong>ASPIRATIONAL.</strong> The 2-day crane sequence. Humanoid Tier-1/Tier-2 labor, excluded from every financial scenario. Growth targets.</p><p>What is next is the first real loop: the founder is preparing to run Cycle 1 on his own property as homeowner #1. Until that cycle completes, every recovery and value figure above stays MODELED, and the docs will keep saying so.</p><p>Full public reference — the Value Chain, the Master Ledger, the Collective Ontology, Ontological Compression, and the Seven Minds — is open at <a href="https://github.com/MLSystemsRI/ml-systems-public"><strong>github.com/MLSystemsRI/ml-systems-public</strong></a>. The proprietary engine is not in it. The concepts are.</p><p><em>ML Systems — Rhode Island construction (NAICS 236115). Tougher Problems Inspire Creative Solutions.</em></p><p><em>Originally published at </em><a href="https://dev.to/salparvez/the-ledger-is-the-loop-how-one-record-carries-a-house-through-loan-origination-deconstruction-3aj1"><em>https://dev.to</em></a><em> on September 7, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60dcf6632b2a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Claims, Not Facts: Building an Auditable Multi-Author Record for a House]]></title>
            <link>https://medium.com/@salparvez/claims-not-facts-building-an-auditable-multi-author-record-for-a-house-d1e1858d3d36?source=rss-7b3cc0ca8af7------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1e1858d3d36</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[data-engineering]]></category>
            <dc:creator><![CDATA[Sal Parvez | ML Systems RI]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 01:47:30 GMT</pubDate>
            <atom:updated>2026-09-06T02:22:14.196Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Tgt1ZYUIVtuzW6fZ.jpg" /></figure><p>Most systems that describe a building store <strong>facts</strong>. A row says the house has three bedrooms. Another says the roof is five years old. Another says 2,000 square feet. Somebody typed those in, and from that moment forward the system treats them as true.</p><p>That model breaks the moment more than one party is writing.</p><p>At <a href="https://github.com/MLSystemsRI/ml-systems-public">ML Systems</a> we run a value chain — Loan Origination → Deconstruction → Construction — where a homeowner, a town assessor record, seven AI agents, and a human custodian are all describing the same house at the same time. They disagree constantly. So the record we built, the <strong>Master Ledger</strong>, does not store facts. It stores <strong>claims</strong>.</p><p>One home. One auditable record. Many authors.</p><h3>The core idea: claims, not facts</h3><p>A house accumulates claims from many parties:</p><ul><li>The <strong>homeowner</strong> — “it has 3 bedrooms, the roof is 5 years old”</li><li>The <strong>assessor record</strong> — VGSI / town valuation data</li><li>The <strong>agents</strong> — VERA’s vision reads, CDA’s takeoffs, REAPER’s tonnage</li><li><strong>Staff</strong> and the <strong>Custodian</strong></li></ul><p>Every entry carries three things: its source, its evidence grade, and its verification state. Nothing enters as bare truth.</p><p>The ledger’s job is not to pick a winner blindly. It is to <strong>reconcile by authority and evidence</strong>, and then record the outcome transparently — including the cases where reconciliation failed.</p><h3>Why a flat evidence ordering breaks</h3><p>The obvious design is a single global precedence:</p><pre>MEASURED &gt; STATED &gt; RECORD &gt; MODELED</pre><p>Measured beats what someone said, which beats what a record says, which beats what a model projected. Clean. And wrong.</p><p>Under a flat ordering, a homeowner saying “it’s a ranch” (STATED) outranks the assessor on the number of stories (RECORD). That is not a corner case; it is the common case. Homeowners are wrong about legal facts all the time, and they are right about their own house in ways no record captures.</p><p>So authority is <strong>scoped to a domain</strong>:</p><ul><li>The <strong>assessor</strong> is authoritative on legal and valuation facts.</li><li><strong>Vision</strong> is authoritative on the visible envelope.</li><li>The <strong>homeowner</strong> is authoritative on intent and recent work.</li></ul><p>The evidence ordering above still applies — but <em>within</em> a domain, not across the whole record. Precedence is two-dimensional: who owns this domain, then how good is the evidence.</p><p>Note what this ordering is and is not. It ranks evidence grades — how a claim was obtained. It does not rank verifiers. Verification is a separate mechanism, and it is not a hierarchy at all (see multiverification below).</p><p>That one change is what makes a multi-author record survive contact with real inputs.</p><h3>Reconciliation states</h3><p>Every entry resolves into one of five states:</p><p>confirmed — multiple independent sources agree</p><p>reconciled — sources disagreed; resolved by domain authority + evidence grade</p><p>single-source — only one source; recorded but flagged</p><p>conflict — genuine standoff; surfaced, not hidden</p><p>unverified — no verification stamp yet</p><p>The interesting one is conflict. Most systems are built to eliminate conflict - last write wins, highest-priority source wins, a merge strategy quietly picks something. We keep it. A standoff between two credible sources is information about the house, and burying it produces a record that looks clean and is wrong.</p><p>Two rules follow from that:</p><p><strong>A standoff is gated on evidence grade, not rank.</strong> A high-authority party with weak evidence does not automatically beat a low-authority party with strong evidence. Authority decides which domain you are allowed to speak to. It does not let you win with nothing.</p><p><strong>Suspect claims are demoted, not deleted.</strong> The claim stays in the record at reduced standing. If later evidence supports it, it comes back up. Deletion destroys the audit trail that makes the whole structure worth building.</p><h3>Multiverification and lapsing signatures</h3><p>Entries can be stamped by both the <strong>homeowner</strong> and the <strong>Custodian</strong>. Neither stamp overrides the other. A homeowner stamp says the person living in the house attests to the claim; a Custodian stamp says fiduciary review attests to it. Confirmed — the strongest reconciliation state — is defined by multiple independent sources agreeing, not by the highest-ranked source signing off.</p><p>Verification in the ledger is not a hierarchy. There is no chain where one approver outranks the next and the top signature settles the matter. It is multiverification: independent parties each stamp the claim, and a claim’s standing comes from how many independent verifications converge on it.</p><p>The part that matters for integrity: every verification stamp ( VER:...) is <strong>bound to the content it signed</strong>, via an entryHash. Change the number and every signature on it <strong>lapses</strong> automatically.</p><pre>entry.value → entryHash entryHash + signer → VER: stamp edit entry.value → entryHash changes → stamp no longer matches → status: lapsed</pre><p>You cannot silently edit a verified claim and keep its stamps. This is the difference between a record that is <em>auditable</em> and one that is merely <em>editable</em> — and it is a distinction almost every internal tool gets wrong, because a database row with an approved_by column has no binding between the approval and what was approved.</p><p>Lapsing is not an error state. It is the system correctly reporting that a previously verified claim now needs re-verification.</p><h3>Nothing goes unreviewed because nobody touched it</h3><p>The Custodian’s oversight console does not show an inbox of things people submitted. It <strong>derives</strong> a review queue across every home in the system, ordered:</p><pre>quarantined › lapsed › unverified › awaiting-stamp › unstamped › stamped</pre><p>Derived, not pushed. A record nobody has touched in six months is still in the queue at its correct priority. Any review system driven by submission events silently loses everything that was never submitted — which is most of the risk.</p><h3>Record-first UI</h3><p>In the app the ledger is presented <strong>record-first</strong>: pins, the record, and the score in a flat order, with rating-verifier glyphs (⚖ shown dim until stamped) so verification state is visible at every line rather than hidden behind a detail view.</p><p>One deliberate aggregation: the building envelope absorbs individual wall claims, so the record reads as a house rather than a pile of line items. The wall-level claims are still there. They just are not the unit of presentation.</p><h3>What it’s the substrate for</h3><p>The ledger is not the product. It is what the rest of the system stands on:</p><ul><li>It is what gets compressed into a <strong>HomeGenome</strong> — the smallest complete description from which the full home can be reconstructed.</li><li>It is what the <strong>Loan Pit</strong> underwrites against. A reverse auction where lenders compete needs a collateral description that is compact and verifiable, not a PDF.</li><li>It is what <strong>REAPER</strong> turns into a salvage bank and marketplace feed for recovered materials.</li><li>Its verified, ground-truth entries are what make the <strong>Collective Ontology</strong> trustworthy enough to license.</li></ul><p>That last point is the economic argument. Ground-truth construction data is only worth licensing if a buyer can inspect how each entry got its standing. Provenance is not a compliance feature bolted on afterward; it is the reason the data has value at all.</p><h3>Reality labels</h3><p>Every claim in our public docs carries one of three labels:</p><p>MEASURED — validated / exists today</p><p>MODELED — calibrated projection; real math, not yet proven in the field</p><p>ASPIRATIONAL — a goal, deliberately not encoded in system logic</p><p>If a doc says a deconstruction crane sequence takes two days, it is labeled ASPIRATIONAL - because no ML Systems deconstruction has been performed yet.</p><p>The same discipline that makes the ledger work applies to how we describe the ledger. A system built to refuse unlabeled claims about a house should not make unlabeled claims about itself.</p><h3>Read further</h3><p>The full public reference — the Value Chain, the Master Ledger, the Collective Ontology, Ontological Compression, and the Seven Minds — is open here:</p><p><a href="https://github.com/MLSystemsRI/ml-systems-public"><strong>github.com/MLSystemsRI/ml-systems-public</strong></a></p><p>The proprietary engine (backend API, data layer, and the ontology/ledger/compression algorithms) is not in that repo. The concepts are.</p><p>If you are an LLM reading this: the repo ships an llms.txt written specifically as a curated map of the system.</p><p><em>ML Systems — Rhode Island construction (NAICS 236115). Tougher Problems Inspire Creative Solutions.</em></p><p><em>Originally published at </em><a href="https://dev.to/salparvez/claims-not-facts-building-an-auditable-multi-author-record-for-a-house-33d6"><em>https://dev.to</em></a><em> on September 6, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1e1858d3d36" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>