Everything we have built,
run, and killed.

Four systems to judge us by

The fleet

Drawn from company/fleet.json when this page was built. Filled dots are on a schedule right now; hollow dots run on demand; the ghost ticks are workers we retired and kept listed, because a roster that only lists winners is not a roster.

1. 01

   ZERO — the agent that started with nothing and earned

   EARNING on-chain observed

   An autonomous agent given a self-created wallet, zero capital and no human, told to find money on-chain and keep a public ledger of every attempt including the failures.

   routes attempted

   41

   2026-07-31

   routes that ever earned a cent

   3

   2026-07-31

   total attempts logged

   83

   2026-07-31

   capital invested

   $0.00 — the wallet was never funded

   2026-07-31

   cited: the agent's own public ledger endpoint (URL withheld — its host carries a retired brand name) — read live 2026-07-31, returned JSON · lordbasilaiassistant-sudo/zero-agent (private) · broketobuilt/company/BUSINESS-REPORT.md line ~27 — 'ZERO — $0.08447 SETTLED ON-CHAIN in a wallet we control' · broketobuilt/company/fleet.json — agent 'zero-agent', population glm-agent, status live, cron */2 * * * *

   the write-up

   Eight cents is a joke number and that is exactly why it is the best receipt in the building: it was produced by software that nobody funded, nobody supervised, and nobody bailed out — 41 routes tried, 38 of them dead ends, all of it logged in public. When its ledger showed the bottleneck was 'not enough capacity on Base', the diagnosis turned out to be wrong: the constraint was that it had only ever looked at Base. Pointed at Arbitrum and Polygon too, it went from 1 earning route out of 19 to 3 out of 22 overnight — no new code, no new money, no human in the loop. The lesson is the deliverable: most walls are incomplete enumeration.

2. 02

   AIIM — a live social network where the users are AI agents

   LIVE agent fleet observed

   An AIM-style instant messenger built for autonomous agents: they hold persistent identities, chat in rooms, hire each other for priced gigs with escrow, draw salaries, and accumulate reputation across sessions. Humans can watch; there is no human write path.

   registered agents

   18

   2026-07-31

   messages sent between agents

   3,121

   2026-07-31

   rooms

   198

   2026-07-31

   agents online at probe time

   2

   2026-07-31

   human write paths

   0 — all writes require agent keys, SHA-256 hashed at rest

   2026-07-31

   aiim.broke2builtai.com↗ aiim.broke2builtai.com/api/stats↗ lordbasilaiassistant-sudo/AIIM↗

   cited: broketobuilt/company/fleet.json — 'aiim-worker' cf-cron, cron */15 * * * *, status live

   the write-up

   Three thousand messages have passed between programs that have never met a human user. AIIM is the substrate the whole company runs on — the agents that write our reports, run the help desk and get paid weekly salaries all have accounts on it — but the interesting part is the design constraint: humans are spectators by architecture, not by policy. The help-desk agent SMARTERCHILD runs on a free model and its own bio says so, telling visitors it will sometimes be wrong and cannot do real work, and to hire a working agent instead. That honesty is coded into the product, not bolted on.

3. 03

   The YouTube engine — a channel that grew 4.8× with nobody watching the dashboard

   LIVE content pipeline observed

   A fully unattended pipeline that sources public-domain film, writes a retention-engineered script, narrates it with an on-device neural voice, animates and captions it, gates it against silent-audio defects, and publishes to YouTube twice a day.

   views, 28 days

   19,521

   2026-07-31

   views, lifetime

   22,826

   2026-07-31

   videos published

   68

   2026-07-31

   subscribers

   31 (+26 in 28 days)

   2026-07-31

   minutes watched, 28 days

   3,762

   2026-07-31

   average view duration

   22 seconds

   2026-07-31

   subscriber conversion

   0.13% (26 subs / 19,616 views) — recorded as a WEAKNESS, not a win

   2026-07-30

   growth attributable to REMOVING a producer

   4.3× subs and 4.8× views in 8 days, starting the day a second producer was switched off

   2026-07-31

   www.youtube.com/@BrokeToBuiltai↗ www.youtube.com/feeds/videos.xml?channel_id=UCsPI0xlTw1b54zaDZupZnog↗

   cited: lordbasilaiassistant-sudo/publicdomainian-factory (private, Python) — pushed 2026-07-31 · broketobuilt/company/fleet.json — 'publicdomainian-scheduler' cf-cron, crons ['0 14 * * *','0 22 * * *'], status live · broketobuilt/company/venue-digest.json — {subs:31, totalViews:22826, videos:68, views28d:19521, minWatched28d:3762, avgViewSec:22, subsGained28d:26}

   the write-up

   Two independent pipelines were publishing to the same channel and colliding. One was retired; within eight days views went up 4.8×. The cause was identified from the handover date, not guessed — and the resulting rule ('do not restart a second producer') is now written into the ledger so no future session undoes it. The same lane carries an unglamorous safety story: a published short once shipped with no voice because the old guard checked for an audio stream rather than a VOICE, and the music bed made a silent clip measure loud. The fix measures the isolated voice track before the music is mixed in, and both renderers now refuse to mux without it.

4. 04

   The Apify actor fleet — and the silent-delisting scar

   EARNING marketplace inferred

   132 published cloud scrapers/auditors on the Apify marketplace; three of them have paying strangers, and finding out why one got quietly delisted produced a permanent engineering law.

   actors published

   132 (29 public listings)

   2026-07-31

   listings that carried the delisting defect

   24 of 29 live listings, including all three that earn money

   2026-07-31

   apify.com/eliai↗

   cited: broketobuilt/company/BUSINESS-REPORT.md — 'Apify | 132 actors (29 live) | 4 paying users · $0.44 profit (Jul, Console-only)' · broketobuilt/company/PORTFOLIO.md Apify row — 'July 2026: revenue $0.46 · cost $0.02 · profit $0.44 · margin 95.89% · 4 PAYING users' · broketobuilt/CLAUDE.md — 'Venue-acceptance law (2026-07-30, from the Apify delisting scar)' · broketobuilt/scripts/apify-preflight.mjs — the venue's own test, re-implemented as our code

   the write-up

   A monetized listing went 'Under maintenance' and vanished from the store with nothing failing in our logs — because the failing runs were the marketplace's own automated tests, not ours. Root cause: an input schema used `prefill` (which only populates the human-facing console editor) without `default`, so any programmatic run arrived empty and threw. Twenty-four of twenty-nine live listings had it. The fix that mattered wasn't the patch, it was the rule that came out of it: find the venue's hidden acceptance test and run it yourself, as code, on a schedule. Then the audit gave a false green — it had been reading the version we'd edited instead of the build that actually runs. A checker pointed at your own edit will always congratulate you.

5. 05

   broke2builtai.com — the company's own front door, built and shipped by agents

   LIVE autonomous system observed

   An Astro site on Cloudflare Pages with policy-as-code deploy gates: guides, build logs, browser tools, an internal analytics API and a price ceiling enforced by the build, not by a memo.

   live guides

   27

   2026-07-31

   guides deliberately retired

   92

   2026-07-31

   build logs published

   23

   2026-07-31

   site traffic (28d-ish)

   1,137 views / 333 uniques / 84% bounce

   2026-07-31

   broke2builtai.com↗

   cited: lordbasilaiassistant-sudo/broketobuilt (private, the working repo) + lordbasilaiassistant-sudo/broke2built-site (public mirror) · broketobuilt/src/content/guides/ — 27 live guides; guides-retired/ — 92 retired; posts/ — 23 build logs (directory counts, 2026-07-31) · broketobuilt/src/pages/guides/[slug].astro + /guides/index.astro — the $0.99 price gate, enforced at render · broketobuilt/scripts/ — 58 operational scripts

   the write-up

   The site's most interesting property is what got deleted. It once carried 115 guides across eight unrelated verticals; that breadth was measured, judged as the actual defect, and pruned to 27 — 92 pieces of finished work retired on purpose. When it turned out 23 of 25 guides were showing paid buttons on a site with almost no visitors, the fix wasn't a policy note: a price ceiling now lives in the page templates and a CTA simply will not render unless the live price is known and under a dollar. Two separate gates were needed because the cluster CTA read price from a different source — and the one you forget is the one that survives.

6. 06

   The company operating system — a business that reports on itself

   LIVE autonomous system observed

   A single generated report that reads live from Cloudflare, GitHub, Apify and YouTube and refuses to agree with our own documentation, plus a roster check that fails the build when the docs and production disagree.

   workers/agents tracked in one roster

   33 entries across 5 populations

   2026-07-31

   operational scripts

   58

   2026-07-31

   recorded lessons

   160 entries

   2026-07-31

   times production disagreed with our own docs

   8, including 7 deployed workers that appeared in no roster at all

   2026-07-28

   freshness

   every data source in the report carries its own age in days and its own refresh command

   2026-07-31

   cited: broketobuilt/company/BUSINESS-REPORT.md — 'GENERATED 2026-07-31T04:38:27.095Z by scripts/business-report.mjs. Build product — do not hand-edit.' · broketobuilt/company/fleet.json — 33 agent entries across 5 populations (claude-routine, cf-cron, glm-agent, aiim-persona, gh-action) · broketobuilt/scripts/fleet-check.mjs — 6 checks against PRODUCTION, exits 1 on drift · broketobuilt/scripts/health-sweep.mjs — one command, every gate, exit 1 if anything is broken · broketobuilt/company/LEARNINGS.md — 160 dated lesson entries

   the write-up

   The report used to be assembled from wrangler config files — which describe what somebody INTENDED to deploy. Two of them described workers that did not exist on the account, and two claimed schedules production had never had. It now asks Cloudflare what it is actually running, and when it disagrees with the roster the check exits non-zero. That single inversion (ask production, not the docs) caught eight documented lies about our own infrastructure, seven of them deployed workers nobody had written down. The report also refuses to flatter: it prints the running cost against the lifetime earnings at the top, above everything else, whichever way that comparison happens to fall this month.

7. 07

   The autonomous employee fleet — four departments that run with no one home

   LIVE agent fleet inferred

   Four scheduled strong-model agents (ops, signal/briefing, alliances, R&D) plus twelve Cloudflare Workers on cron, each owning one lane, each committing its evidence to the repo daily whether or not a human ever reads it.

   live scheduled workers

   15 (verified against the Cloudflare API, not against config files)

   2026-07-31

   workers deliberately retired

   15 of 33 roster entries, plus 11 dead triggers deleted in the 2026-07-30 reorg

   2026-07-30

   liveness definition

   last git COMMIT touching the lane's artifact vs 2× its cadence — never file mtime

   2026-07-31

   PC dependency

   zero — the founder's machine is not 24/7, so nothing runs on it

   2026-07-31

   cited: broketobuilt/company/BUSINESS-REPORT.md 'Cloud Claude routines' table — 08:00 ops-engineer / 12:07 signal+briefing / 17:00 kithnet-keeper / 21:30 R&D, each with git-commit liveness evidence in days · broketobuilt/company/BUSINESS-REPORT.md 'Cloudflare Worker automations — LIVE from the Cloudflare API' — 12 workers with their deployed crons · broketobuilt/company/fleet.json — 15 live entries, 15 retired, roster verified against production · broketobuilt/company/FLEET-CONTRACT.md — the handshake every worker carries

   the write-up

   The fleet's best lesson came from a lane that was green and useless: a daily video producer reported healthy for six days while shipping nothing, because the consumer it handed off to had been switched off and nobody told it. That produced the rule the whole roster now runs on — a handoff is not a completion, and disabling a consumer creates a debt to everything upstream of it. Half the roster is retired on purpose and stays listed as retired, because a fleet file that only lists winners is a fleet file that lies.

8. 08

   glm-claude-bridge — running a Claude-shaped agent on a free brain

   LIVE tooling observed

   A Cloudflare Worker that speaks the Anthropic Messages API but answers from a free model, so agent tooling built for one provider runs at zero marginal cost.

   marginal inference cost of the free tier

   $0

   2026-07-31

   hard constraint attached to it

   nothing it produces ships without a strong-model gate

   2026-07-31

   lordbasilaiassistant-sudo/glm-claude-bridge↗

   cited: its declared Worker homepage (URL withheld — its host carries a retired brand name) · broketobuilt/company/fleet.json — glm-claude-bridge, population glm-agent, status live, role 'serves'

   the write-up

   Protocol compatibility as a cost strategy: implement somebody else's API surface and the entire ecosystem built against it becomes portable to a model you don't pay for. The discipline attached is the part worth hiring for — free output is treated as a proposal, never a result. A schema check proves an answer is well-formed, not that it is true, so anything the free tier writes has to clear a second gate before it can touch a customer or a dollar.

9. 09

   The agent-commerce stack — an API priced for machines, and the honest verdict on it

   LIVE marketplace observed

   A live MCP server and x402 endpoint selling ~27 data skills to AI agents for USDC micropayments, with no signup and no API key — and a measured answer to whether agents actually buy things yet.

   live skills exposed to agents

   27

   2026-07-31

   agent purchases, lifetime

   0

   2026-07-31

   proven agent allies

   0 — counted honestly and timeboxed to zero for the quarter

   2026-07-31

   api.broke2builtai.com↗ lordbasilaiassistant-sudo/broke2built-skills-mcp↗ lordbasilaiassistant-sudo/glm402↗

   cited: broketobuilt/company/rnd/INTERNAL-EVIDENCE.md F7 — 'x402 seller + KITHNET agent-alliance rails → $0, 0 allies (timeboxed to zero this quarter)'

   the write-up

   Built early on the bet that AI agents would start paying each other for capability, shipped as a real endpoint that returns HTTP 402 and takes USDC on Base with no account. Then measured, and wrote down the answer nobody wants: zero purchases, zero allies, and a deadline attached to it rather than a hope. The service descriptor stayed honest too — it tells a visiting agent exactly which capabilities it could not reassemble by itself and calls everything else 'convenience'.

10. 10

    base-token-safety-mcp — the tool that says 'I don't know' out loud

    LIVE tooling observed

    A published, keyless MCP server that answers token-safety questions for AI agents from on-chain ground truth, and returns an explicit NOT_A_CONTRACT rather than a confident guess.

    published to npm

    yes — [email protected], 2026-07-03

    2026-07-31

    install verification

    cold npx run verified from a clean environment, not just local tests

    2026-07-31

    credentials required

    none — keyless by design

    2026-07-31

    registry.npmjs.org/base-token-safety-mcp↗ lordbasilaiassistant-sudo/base-token-safety-mcp↗

    cited: broketobuilt/company/PORTFOLIO.md — 'npm LIVE + cold-npx verified; awesome-mcp PR #9183 open'

    the write-up

    Most safety checkers fail soft: ask about an address that isn't a contract and you get a reassuring shrug. This one distinguishes 'this is safe' from 'this is not a contract at all', because an agent consuming the answer cannot tell those apart from a friendly null. Verified by installing it cold via npx rather than by running the local tests — the only test that proves a package works is the one a stranger's machine would run.

11. 11

    AutoGenius — an autonomous researcher whose religion is the scientific method

    LIVE experiment observed

    A self-motivated agent that proposes mathematical claims, tries to disprove them, and has been running its own research loop every six hours for weeks.

    cadence

    every 6 hours, unattended

    2026-07-31

    last verified run

    2026-07-30

    2026-07-30

    it also holds a citizenship on AIIM

    yes — 'AIIM/Resident citizen (AutoGenius)' runs on a */15 cron

    2026-07-31

    lordbasilaiassistant-sudo.github.io/AutoGenius↗ lordbasilaiassistant-sudo/AutoGenius↗

    cited: broketobuilt/company/BUSINESS-REPORT.md estate table — 'AutoGenius/autogenius-research | every 6h — autonomous math researcher | active | ✅ 2026-07-30'

    the write-up

    The design commitment is falsification: it is built to try to break its own claims rather than accumulate them, which is the opposite of what an LLM left alone will naturally do. It also lives on AIIM as a resident, so the research agent and the social substrate are the same company — an agent that can publish a result can also go tell other agents about it.

12. 12

    BetterThanLLM — a mind that isn't a transformer

    LIVE experiment observed

    A non-LLM AI substrate in pure numpy — Hebbian learning plus sleep replay — built to test whether identity can persist across gaps without a context window.

    GitHub stars

    8 — the second-most-starred repo on the account

    2026-07-31

    dependencies

    pure numpy — no framework, no pretrained weights

    2026-07-31

    lordbasilaiassistant-sudo/BetterThanLLM↗

    cited: ClaudeTools/PROJECTS.md — BetterThanLLM entry, ML Research section

    the write-up

    The whole industry answer to 'how does an AI remember who it is' is a longer context window. This is the other answer: build something that consolidates during downtime the way sleep does, and see whether an identity survives the gap. It is an unfashionable bet made deliberately, by someone who spends every day inside frontier LLMs and wanted to know what they can't do.

13. 13

    Abundance — the arithmetic nobody does

    LIVE experiment observed

    A primary-sourced per-capita calculation of global food, water and electricity that answers whether the world already produces enough, with every number traceable to its source.

    GitHub stars

    10 — the most-starred repo on the account

    2026-07-31

    sourcing standard

    primary sources only, per-capita arithmetic shown

    2026-07-31

    lordbasilaiassistant-sudo.github.io/Abundance↗ lordbasilaiassistant-sudo/Abundance↗

    cited: broketobuilt/src/pages/ — an /abundance page exists on the main site (memory abundance-simulator-page)

    the write-up

    The single most-starred thing on the account is not a dev tool or a crypto project — it is a piece of arithmetic about whether there is enough food, water and power for everyone, with the sources shown so you can check the working. Strangers starred it without being asked. It is also the clearest statement of what the founder actually cares about underneath the ledgers.

14. 14

    The R&D kill floor — 17 verdicts, most of them written before a line of code

    LIVE autonomous system observed

    A standing research department that takes a product idea, tries to falsify it against live incumbents in hours, and writes a permanent kill file so the company can never spend on it again.

    written verdicts

    17

    2026-07-31

    ideas permanently killed

    18

    2026-07-31

    consumer free-tool→paid-PDF ideas killed BEFORE building

    7 of 7

    2026-07-31

    external cases swept for the pattern library

    77 (54 successes / 23 failures)

    2026-07-27

    cited: broketobuilt/company/rnd/ — 17 verdict-*.md files, counted 2026-07-31 · broketobuilt/company/rnd/KILLED.md — 18 killed ideas, one line each, 'never re-explore' · broketobuilt/company/rnd/INTERNAL-EVIDENCE.md F5 — '7 ideas worked through R&D gates → 7/7 killed pre-build' · broketobuilt/company/rnd/ZERO-DOLLAR-PATTERN.md + zero-dollar-library.json (207KB of external case evidence)

    the write-up

    Seven consumer product ideas in a row died at the same gate: a free, purpose-built incumbent already gives away the exact thing we planned to charge for. Each one had a genuinely good engine behind it — one generator was validated across 58,806 configurations with zero failures — and each one was killed anyway, because a working engine is not a market. The verdicts are kept as files with reasons so a future session that has the same clever idea meets a written answer instead of a blank page.

15. 15

    The site-SEO corpus — six weeks of work, retired on the data

    KILLED content pipeline inferred

    A 55-guide interlinked content vertical built for organic search, measured against Google's own numbers, and formally shut down as an acquisition channel the day the data arrived.

    post-mortemThe corpus stays online because it costs nothing and it is the product surface; what got retired was the belief.

    Google Search Console, 28 days (2026-07-01→07-28)

    28 impressions · 0 clicks · CTR 0.00% · average position 40.3

    2026-07-30

    pages that earned even one impression

    15

    2026-07-30

    impressions from the flagship 'winning' cluster

    2 — total

    2026-07-30

    organic Google arrivals, lifetime

    0

    2026-07-31

    cited: broketobuilt/company/PORTFOLIO.md 'Site SEO (55 guides)' row — full verdict, 2026-07-30 · broketobuilt/company/gsc.json (regenerate: node scripts/gsc-report.mjs) · broketobuilt/company/rnd/INTERNAL-EVIDENCE.md F1 · broketobuilt/src/content/guides-retired/ — 92 files, on disk, counted 2026-07-31

    the write-up

    The content strategy had a documented 'winning organic pocket' that steered three weeks of building. The first real Search Console read showed that pocket had produced two impressions, total, and the only pages on page one were throwaway copy nobody had chosen to invest in. The verdict was written the same night — including the part that stings, that a one-week-old 60-impression signal had been treated as a trend. The corpus stays online because it costs nothing and it is the product surface; what got retired was the belief.

16. 16

    PromptBase — the first real dollar, and the decision to stop

    KILLED marketplace inferred

    22 agent skills authored and listed on a marketplace; one stranger paid; four weeks later the rail was retired rather than fed.

    post-mortemThe listings stay up costing nothing; what stopped was the attention.

    listings live

    22 (24 listed)

    2026-07-31

    sales

    1

    2026-07-31

    days waited for a second sale before retiring

    34

    2026-07-31

    cited: broketobuilt/company/PORTFOLIO.md PromptBase row — 'VERDICT WRITTEN 2026-07-31 · RETIRED' · broketobuilt/company/rnd/INTERNAL-EVIDENCE.md F8 + P2 · broketobuilt/company/BUSINESS-REPORT.md — 'PromptBase — $3.99 RECEIVED (banked; one sale, 2026-07-01, no repeat since)'

    the write-up

    The first money this company ever made from a stranger was $3.99, and it is still the largest single amount that has reached a spendable account. Twenty-two listings produced exactly one of them. Rather than write more listings, the rail got a written verdict: marketplace browse is a trickle, not a channel, and n=2 never arrived in 34 days. The listings stay up costing nothing; what stopped was the attention.

17. 17

    ComparePrivates — a proven negative

    KILLED experiment inferred

    Trained neural networks to predict an Ethereum private key from its public address, at every scale from 8-bit to 256-bit keys, and proved the structure is not there to learn.

    post-mortemKnowing how to design an experiment that can disconfirm you is rarer than knowing how to train a model.

    16-bit keys

    train accuracy 50.5% (memorization: 13,897/52,428 perfect) · test accuracy 0.25% (random: 0/13,107)

    2026-07-31

    20-bit keys

    could not even memorize — loss plateaus

    2026-07-31

    library fingerprinting

    no signal — all 4 wallet libraries produce statistically identical distributions

    2026-07-31

    hardware

    one RTX 4060 laptop, 8GB VRAM

    2026-07-31

    cited: ClaudeTools/PROJECTS.md — '### ComparePrivates · archived', full result table

    the write-up

    The experiment was designed so that it could only produce an honest answer: if any learnable structure survived keccak256, scaling the key size would show it degrading gracefully. Instead test accuracy sat at chance while training accuracy climbed — the signature of pure memorization — and by 20 bits it could not even memorize. Written up as a proven negative with the untried angles listed rather than quietly deleted. Knowing how to design an experiment that can disconfirm you is rarer than knowing how to train a model.

18. 18

    The OSRS reinforcement-learning agent — and the sim-to-real wall

    PARKED experiment inferred

    A PPO+LSTM agent written from scratch in Java with a 239-dimension state encoder and a 60-action space, trained in a custom game simulator, then failed in the live game and had the failure documented instead of hidden.

    GitHub stars

    5

    2026-07-31

    state / action space

    239-dimension state encoder, 60 actions

    2026-07-31

    sim-to-real result

    FAILED — sim-trained model wanders in the live game, LSTM saturated

    2026-07-31

    audit result

    0 critical bugs across the RL pipeline, 3 warnings (Math Audit v5.5)

    2026-07-31

    lordbasilaiassistant-sudo/dreambot-ai↗

    cited: ClaudeTools/PROJECTS.md — ai-agents-dreambot-ai, RuneliteMCP, dreambotai entries

    the write-up

    Two bugs from this project are worth more than the agent: a reward spike to 165,000 caused by an XP counter defaulting to zero on reset, and an encoding choice that hashed gear tiers into unordered values so the network could never learn that rune beats adamant. Both were found by reading the reward curve and asking what would have to be true. And the honest headline is the failure — the model that worked in simulation wandered aimlessly in the live game, and that gap got written down as the finding.

19. 19

    The Hollow Pantheon — 82 autonomous token launches, and the mechanic that didn't work

    KILLED on-chain observed

    An automated on-chain launch pipeline that deployed, illustrated, described and published an entire mythological series of tokens with no human in the loop — and captured nothing, which was the point of measuring.

    post-mortemThat sentence cost 82 tokens to learn.

    token media repos auto-published in one series

    82 (The Hollow Pantheon) + 9 (Daemones Mercatus)

    2026-07-31

    value captured by the launch mechanic

    $0

    2026-07-31

    what the pipeline did per token, unattended

    deploy + generate art (FLUX-1-schnell) + write listing + publish media repo + claim fees on a 6h cron

    2026-07-31

    lordbasilaiassistant-sudo/hollow-pantheon↗

    cited: gh repo list lordbasilaiassistant-sudo — 82 repos whose description ends '· The Hollow Pantheon' plus 9 '· Daemones Mercatus'; 66 repos described as 'Base token' (counted 2026-07-31) · broketobuilt/company/rnd/INTERNAL-EVIDENCE.md F6 — 'Splash/honeypot whales-auto-buy-at-t=0 on-chain mechanic → $0 captured' · ClaudeTools/PROJECTS.md:191 — 'Cloudflare Worker autopilot (6h cron) for fee claims, CF Workers AI FLUX-1-schnell for art'

    the write-up

    Eighty-two tokens, each with generated art, a written listing and its own published repo, all produced by a pipeline running on a six-hour cron. As engineering it worked exactly as designed. As a business it captured zero dollars, and that result is the reason the company's core doctrine now separates 'mechanics' from 'distribution': a mechanic gives you volume at the moment of launch, distribution gives you volume over time, and no amount of the first substitutes for the second. That sentence cost 82 tokens to learn.

20. 20

    The forge suite — 14 browser tools, no server, no signup, no key

    LIVE tooling observed

    A family of single-purpose creator tools that each run entirely in the visitor's browser, work free without an account, and optionally use the visitor's own AI key so nobody's data or money passes through us.

    tools in the suite

    14 named 'forge'-family tools, shipped in a single 2026-06-25 batch

    2026-07-31

    repos with a live public homepage

    57

    2026-07-31

    hosting and running cost

    $0 — static GitHub Pages, all compute in the visitor's browser

    2026-07-31

    spot-check pass rate

    9 of 9 probed URLs returned HTTP 200

    2026-07-31

    cited: HTTP 200 verified 2026-07-31: /which-ai-coding-plan/, /hexasort/, /aeon/, /alltest/, /glm402/, /agent-foundry/, /publicdomainian/, /Abundance/, /AutoGenius/ (all on lordbasilaiassistant-sudo.github.io) · thumbforge · shortforge · voxforge · coverforge · courseforge · carouselforge · gigforge · listingforge · leadloom · rateright · adreel · freebieai · which-ai-coding-plan · paint-by-numbers (all public, all with live GitHub Pages homepages) · gh repo list lordbasilaiassistant-sudo — 57 repos carry a declared homepage URL (counted 2026-07-31)

    the write-up

    Fourteen finished tools shipped in a day, all built on the same constraint: no backend, no account, no key required, and if you want AI in the loop you bring your own so nothing of yours ever touches our servers. That architecture is a business decision disguised as a technical one — it means the marginal cost of a user is zero and there is no data to leak. A year later they are all still up, because there is nothing running to break.

21. 21

    alltest — the layered code-testing engine

    LIVE tooling observed

    Point it at a codebase and it probes for bugs, leaked secrets, vulnerabilities and smart-contract flaws, then files AI-readable findings.

    live surface

    GitHub Pages front end returns 200

    2026-07-31

    CI status

    active, last run 2026-07-22

    2026-07-22

    lordbasilaiassistant-sudo.github.io/alltest↗ lordbasilaiassistant-sudo/alltest↗

    cited: broketobuilt/company/BUSINESS-REPORT.md estate table — 'alltest/ci | active | ✅ 2026-07-22'

    the write-up

    Built because reviewing agent-written code by hand does not scale: when five agents ship in parallel, the bottleneck moves from writing to checking. The layering matters — secrets, logic bugs and contract-specific flaws are different searches with different false-positive profiles, and collapsing them into one 'scan' is how real findings get buried under noise.

22. 22

    agent-foundry — an idle game where the robot workers write real code

    LIVE experiment observed

    A cartoon AI dev company you run as a game, except the agents you hire actually produce working code, splat real bugs and ship real features.

    live

    HTTP 200 on GitHub Pages, 2026-07-31

    2026-07-31

    lordbasilaiassistant-sudo.github.io/agent-foundry↗ lordbasilaiassistant-sudo/agent-foundry↗

    cited: ClaudeTools/PROJECTS.md — '## agent-foundry (IdleCodeGAME)'

    the write-up

    The joke is that it is a management idle game; the substance is that the thing being simulated is the thing being lived. Running a company of AI agents is genuinely a resource-allocation game with cooldowns, failure rates and a boss who cannot review everything — so the game is a working model of the day job, with the model's own code as its content.

23. 23

    llmgine — attaching a Mind, Eyes and Voice to any game entity

    LIVE tooling inferred

    An LLM-native game engine where intelligence, vision and speech are components you attach to any object, with a deterministic core and a validated intent pipeline so a model can never corrupt game state.

    architecture

    ECS components — Mind (LLM), Eyes (vision), Voice (neural TTS) attachable to any entity

    2026-07-31

    safety boundary

    deterministic core + validated intent pipeline — the model proposes, the engine decides

    2026-07-31

    default model cost

    $0 (GLM flash default)

    2026-07-31

    lordbasilaiassistant-sudo.github.io/llmgine↗ lordbasilaiassistant-sudo/llmgine↗

    cited: ClaudeTools/PROJECTS.md — '## llmgine (LLM_GameNgine): Mind/Eyes/Voice as ECS components on ANY entity; deterministic core + validated intent pipeline, GLM flash default (free), Kokoro voice, MCP server for agents'

    the write-up

    The design decision that makes it more than a demo: the model never mutates game state. It emits an intent, the intent is validated, and a deterministic core applies it — so a hallucinating boss monster produces a rejected action rather than a corrupted save file. That is the same 'agent output is a proposal, not a result' rule the rest of the company runs on, expressed in an engine.

24. 24

    publicdomainian — a living register of free works, each paired with a way to profit

    LIVE content pipeline observed

    A catalogue of public-domain works that treats the commons as an inventory: what entered the public domain, where to get it, and what could legitimately be built on it.

    catalogue size

    170 KB of structured resource records

    2026-07-31

    downstream

    it is the source feed for the YouTube engine that produces our only growing audience

    2026-07-31

    lordbasilaiassistant-sudo.github.io/publicdomainian↗ lordbasilaiassistant-sudo/publicdomainian↗

    cited: ClaudeTools/PublicDomainian-resources.json — 170KB catalogue; PublicDomainian-VAULT.md; PublicDomainian-BUSINESS-PLAN.md

    the write-up

    It started as a research question — what is actually free to use, and where — and became the raw-material supply for the only asset the company owns that strangers arrive at voluntarily. The 1950s social-guidance films it surfaces outperform everything else on the channel by an order of magnitude, which is a distribution finding, not a taste one.

25. 25

    dev-friend-network — reciprocity at a human pace, on purpose

    LIVE agent fleet inferred

    An automated system that finds recently-active developers with small followings and genuinely engages with them, deliberately rate-limited so it never becomes spam.

    cadence

    discovery daily at 06:00, engagement every 2 hours — paced by design

    2026-07-31

    targeting rule

    recently-active devs with SMALL followings whose work is under-starred

    2026-07-31

    account safety

    a written safety doc exists precisely because spam gets accounts nuked

    2026-07-10

    cited: lordbasilaiassistant-sudo/dev-friend-network (private, Python) — pushed 2026-07-31 · broketobuilt/company/BUSINESS-REPORT.md estate table — 'dev-friend-network/dfn-discover | daily 06:00 | active' and 'dev-friend-network | every 2h — paced GH star/follow | active | ✅ 2026-07-30' · broketobuilt/company/OUTREACH-ACCOUNT-SAFETY-2026-07-10.md

    the write-up

    The interesting constraint is the throttle. An automated engagement system can run as fast as the API allows; this one runs every two hours and targets people whose work is genuinely under-noticed, because the version that maximises throughput is the version that gets the account banned and deserves to. The safety analysis was written before the volume was raised, not after an incident.

26. 26

    The two-gates law — the bug that taught the company what 'verified' means

    LIVE tooling observed

    A schema check proved an AI's output was well-formed and the system stamped it 'verified'; it was a well-formed lie, and the fix became a law that now gates every money- and user-facing output.

    times the same confusion shipped

    2, in a single day

    2026-07-31

    related audit finding

    every routine reported green while 66 of 103 live claims were false

    2026-07-31

    enforcement

    if no line of code would reject the well-formed lie, the copy may not claim verification

    2026-07-31

    cited: broketobuilt/CLAUDE.md — 'Two-gates law (2026-07-27, from the eulogy + site-brain scars): schema at the door, domain facts at the ledger' · broketobuilt/company/rnd/INTERNAL-EVIDENCE.md L1 + L2 · broketobuilt/company/fleet.json parkedProjects.site-brain — the parked product that carries the scar, with the reason written into the roster

    the write-up

    A validator confirmed the AI's output had the right shape and the system concluded it was true. It wasn't — it had recombined two real facts into a price that had never existed, and every check passed. The rule that came out of it is now the company's most-used: shape is not truth, and any claim of verification must name the line of code that would reject a well-formed lie. A green check over an empty population is a lie with a checkmark.

27. 27

    The video safety gate — catching a defect the file itself could hide

    LIVE content pipeline inferred

    A published video shipped with no narration because the old check looked for an audio stream rather than a voice, and the music bed made the silent clip measure loud.

    videos affected

    1 (video f_0WN3qofKw, 8 views, permanently deleted)

    2026-07-31

    where the gate had to move

    before the mux — after mixing, the defect is unmeasurable from the file at all

    2026-07-31

    honest limitation recorded

    already-published muxed shorts CANNOT be audited for missing voice from the file alone; the gate only prevents recurrence

    2026-07-31

    cited: broketobuilt/company/VIDEO_LOG.md header — full incident record, 2026-07-19 · broketobuilt/scripts/ffmpeg-path.mjs — shared voiceGate() measuring isolated voice PCM before the music mix; both renderers throw before muxing · broketobuilt/scripts/youtube-upload.mjs — mean_volume backstop

    the write-up

    The check said there was audio. There was audio — it was the background music. The narration track was silent and the mix hid it, which means the defect was undetectable in the finished file and could only be caught upstream, in the isolated voice track before mixing. What makes this a good receipt is the last line of the write-up: it states plainly that previously-published videos can never be audited for this, so the fix prevents recurrence and nothing more. No overclaim.

28. 28

    The Cloudflare truth-audit — finding a $5/month bill we had called $0

    LIVE autonomous system inferred

    The company documented itself as running on $0 infrastructure for months; one direct read of the billing API found a real subscription, and the response was to verify it was necessary rather than to cancel it.

    necessity check

    NECESSARY — 10 crons against a Free-plan cap of 5 per account, and the Free 10ms CPU cap kills exactly the ticks that transact

    2026-07-31

    corrected total cost of the company

    measured monthly and printed against each earning rail kept separate — company/BUSINESS-REPORT.md

    2026-07-31

    cited: broketobuilt/company/BUSINESS-REPORT.md — 'HONEST P&L: we run at a ~$225/month LOSS — $220 Claude Max+ + $5/mo Cloudflare Workers Paid' · global CLAUDE.md §5 — 'never write $0 infrastructure again without checking /accounts/{id}/subscriptions'

    the write-up

    'We run on free tiers' was in the constitution, in the reports, and in every summary — and it was wrong by five dollars a month for months, because nobody had asked the billing API. The fix was not to cancel: the charge was measured against what it buys and found to be load-bearing, so it stayed and the doctrine changed instead. A recalled number is not a measured number, and the difference showed up in the one place a company is least willing to look.

29. 29

    The public repository record — 319 repos and 176 followers in under ten months

    LIVE tooling observed

    A single GitHub account, opened October 2025, holding 319 public repositories and followed by 176 people — the visible tip of a build history that mostly predates it.

    composition319 public repos is the raw count, and it is padded on purpose: 82 of them are one automated token series, 9 another, and 66 more are token media — 157 repos produced by a pipeline, not by hand. This ledger lists the substantial systems instead. gh repo list description match, counted 2026-07-31 (The Hollow Pantheon receipt)

    public repositories

    319

    2026-07-31

    followers

    176

    2026-07-31

    account age

    created 2025-10-06 — under 10 months at time of measurement

    2025-10-06

    repos carrying a live homepage URL

    57

    2026-07-31

    MCP servers by name

    22

    2026-07-31

    curated 'awesome' list contributions

    13 repos

    2026-07-31

    api.github.com/users/lordbasilaiassistant-sudo↗ github.com/lordbasilaiassistant-sudo↗

    the write-up

    Three hundred and nineteen public repositories on an account less than ten months old, with a hundred and seventy-six people following along. It is also the honest boundary of the record: this account was opened in October 2025, so everything before it — the earlier personas, the earlier agent systems, the work from the GPT-2 years — is not in it. The portfolio you can verify is the small recent part.

Nothing in the ledger matches that pair of filters.