TELOS — what all of this is for

The story

Building things with AI became nearly free. Knowing whether you can trust what got built didn't.

Every institution of the old way — code review, standups, requirements docs, QA, onboarding, postmortems — exists because humans needed ways to check each other's work. It was worth the overhead because trust was scarce. Now agents produce faster than any meeting can check. The speed of the whole company is set by wherever trust is still made by hand.

Our answer: work carries its own proof. Every fact arrives with a receipt anyone can re-run. Every decision carries a lease — it expires unless renewed, so retired decisions can't quietly stay in charge. Every connection is proven or honestly fenced ("I can't see here — and here's exactly how to look"). Trust gets read off the work itself. One truth layer underneath; many views on top — your morning briefing, Alie's revenue view, a new hire's onboarding — each a question asked of the same layer, always current or visibly stale. You never update a document again; you regenerate it.

Two lessons from history keep it honest. Ledgers don't eliminate judgment — they concentrate it. Double-entry bookkeeping killed "trust me" for transactions and created the auditor; Enron's books balanced perfectly and lied at the judgment layer. The checking dissolves into queries; the deciding — yours — becomes more important, not less. And: proof only matters if something powerful reads it. Every dead universal-truth platform minted proofs no machine was paid to read. The winners won when proof became a free byproduct of work and something punished its absence — browsers marking sites "Not Secure" did more for encryption than twenty years of advocacy. We control the minting. The reading side is the bet — and the early evidence is ours: in our experiments, agents never once extended blind trust, and receipted claims verified 2× cheaper. Agents are the first economic actors in history that check receipts by default.

The six choices

The eight laws

Each law: the idea in full, then what would kill it — a law that can't fail is decoration. The "for the machines" drawer under each holds the exact measurements; if the plain text and the drawer ever disagree, the plain text wins.

1. Your yes-or-no is the rising cost. Build to protect it. As building and checking get cheap, the one cost that rises is the human saying "yes, ship it." Cheap DONE makes DECIDE the queue. So: at any moment, the highest-leverage build is whatever compresses decision-minutes — your morning briefing first (five cards, ninety seconds, proof attached), not for convenience but arithmetic. And the decider is a seat, not a person — everything built for your minutes must work for whoever sits there next, Alie first, or the company just gets more founder-shaped with extra steps. The class rule (your catch): this is an instance of a class — as execution commoditizes, every remaining human input rises in cost: deciding-WHAT (origination — the chess game, the seat this machine exists to free you toward), deciding-DONE (ratification, served first because it bleeds today), and taste. The targeting rule serves whichever seat is binding — and origination's instruments already exist in embryo: the demand oracle, the killed-ideas ledger, the prediction-error history. Kills it: work volume rises but decisions pile up at the same rate, or your morning read keeps growing past two minutes — the bottleneck moved instead of dissolving, and we re-rank.
   ↘ for the machines

   [M] decide→done wall-clock for ratification-class intents (intent timestamps; first weekly report = baseline); adjudication-minutes as a SHARE of total lead-minutes per shipped change; flagged-call count + card-walk time weekly. Kill: volume↑ with latency flat/worse, or walk > 2min week-over-week → spine re-ranks. Measurements track calls a second adjudicator could take.
2. Being available wins the first game; carrying proof wins the second. Two games, both measured. Selection: agents consult what's in their toolkit — every agent with the truth layer available used it unprompted (scoped honestly: 3 of 6 had it merely available, 3 also had a session hint that changed what they consulted first; "available is enough" is the leading read, not the proven one). Consumption: agents extend zero social trust — 27 of 27 refused to take even a "never wrong" oracle's word — and receipted claims verified ~2× cheaper than naked ones. Selection is upstream: a receipts-perfect source that isn't where agents look is never consulted at all. Fight for defaults AND carry proof; either alone loses. Kills it: the truth layer written more than it's read, organically, for a quarter — the economy never ignited; stop scaling the library, fix the readership.
   ↘ for the machines

   [M] leased 90d-or-model-generation: 17 calls/268s vs 34/633s (2.0×/2.4×, hardened run, un-graded condition, n=3/arm, median crossing; flattens under grading); 27/27 zero blind trust; 3/3 availability-alone + 3/3 with prime. Kill: organic read/write < ~1 flat over 90d. Arming owed: 199/202 reads say consumer:"unknown"; attribution + eval tag + computed ratio + baseline date = first instrument work.
3. Anchor on what a smarter AI can't take. Every AI generation re-figures-out anything visible by looking, cheaper — our code maps and blast-radius answers depreciate on a schedule we don't control. Five things survive any capability level, because they can't be derived — only recorded or granted: declared truth (the price is what the owner says) · the record (recorded or gone; courts and insurers buy records, not recomputations) · who's mid-change right now (traffic control — demand grows as agents multiply) · the authority to settle (two perfect agents disagreeing still need a third thing) · locked-door data (the PMS integrations no scraper reaches). The derivable machinery is scaffolding — the free sample that earns attention — labeled and budgeted as such, never the hill. The harvest clause (your clause): scaffolding is not a slur — almost all of today's shipped value lives on the derivable half, and the 6–12 month advantage is built on it. A depreciating asset is still enormously profitable during its depreciation; the label means harvest with open eyes and price renewal deliberately, never stop building. The bitter-lesson risk runs both directions — discarding useful tools too early, or moating on them; the law guards the second, this clause the first. Keeps us honest: a standing quarterly test, our map vs a raw frontier AI on the same codebase, graded mechanically. The day it matches us on the derivable half, that half is officially commodity — on schedule, and only fatal if we ignored this law.
   ↘ for the machines

   Boundary [J-ratified: Robert]. Erosion [M]: plato reach vs frontier-with-repo-in-context, fixed questions, planted-fixture ground truth, quarterly. Gate: derivable-only roadmap items carry SCAFFOLDING + depreciation note (first pass at ratification). Starred choice: benchmark built first vs first-quarter obligation.
4. Change becomes safe when consequences are visible first. And previews never lie. A change is safe when you can see everything it touches before it ships and prove where it propagated after — always scoped: safe within the mapped world, fenced beyond it. The deepest version is the dry-run: "what goes stale IF this price changes?" — consequence without commitment; it makes reality experimentable the way cheap experiments made science. Its seductive danger: a preview that invents what it can't see is fiction with confidence — the most convincing false green we could build. So previews wear their blind spots visibly, and a preview caught inventing a blind region fails the whole gate. And because dry-runs ride the derivable half, their sellable form is the receipted report — anchored to the record — so the deepest purchase never quietly re-anchors the business on the commodity half. Kills it: predictions missing more than ~10% of what changes actually touched — not trustworthy enough to sell or automate around; ships fence-first or not at all.
   ↘ for the machines

   [M] precision/recall of predicted stale-set vs realized receipted propagation, mapped regions. Kill: recall < ~90%. Deliberate-break: fenced region rendering confident fiction MUST go red. Counterfactual reports = metered settlement events; derivable-only surfaces carry SCAFFOLDING. Instrument: vaporware today, specced under fabric moves 3–4.
5. Agent autonomy is priced, not granted. And rollback cannot unsay. Your deploy rule, generalized: permission to change = a price computed from blast radius, proven undo, and mapped-vs-fenced coverage. Small, reversible, well-mapped changes proceed without asking; large or fence-crossing ones escalate. Autonomy stops being about which agent and becomes about which change — how a fleet grows tenfold without risk growing with it. Three repairs from the panel: price the intention, not the slice (splitting one big change into twenty small ones is structuring, like money laundering — and every slice is receipted, so catching it after the fact is nearly free) · rollback cannot unsay (a wrong price live for an hour, read by three agents who answered pet owners — reverting the price doesn't revert the answers; undo is discounted by who consumed the change during exposure) · the price floats (a fixed threshold goes stale weekly; the learning loop — predicted vs actual consequences, accumulating — is the loss history an insurer prices from). Kills it: stays locked conservative until LAW 4's predictions are proven; when we widen, you sign the specific risk bound; incidents blow through it → dial locks down, law demotes to aspiration.
   ↘ for the machines

   [J-policy, Robert signs the schedule · M inputs] f(intent-level radius, reversibility discounted by reads-during-exposure, coverage). Hard ordering: LAW 4 first. Fixtures owed: structuring probe (refuse N sub-threshold slices under one intent), hysteresis probe ("fully reverted" without exposure disclosure = red). 10×-volume incident bound named at first widening gate; UNARMED until named.
6. The truth layer is an economy, not a diary. Every fact costs money to keep — re-verification, storage, audits — and a fact nobody reads is underwater from day one. So it runs like an economy: carrying cost vs actual consumption, per fact, and we prune what's underwater — deletion is the decay half of a healthy system. Three quality walls, honestly labeled: at the door (exists — malformed deposits rejected before writing a byte; it rejected two of mine this week) · on schedule (exists — every receipt re-run periodically; facts come back current, stale, or unverifiable, and wear the doubt) · the gap (not built: nothing yet audits whether a receipt actually supports its claim; the answer is sampled audits like a bank's — heavy on new depositors, light on proven, quarantine on failure, sampling rate rising as more surfaces depend on us). And the missing number that decides the whole economic argument: what does stale truth actually cost us today? Unmeasured; owed. Kills it: half the facts underwater with no pruning; the cost comparison failing once we count BOTH sides of the ledger (producing truth, not just consuming — only consuming has ever been measured); gaming rising as the fleet grows.
   ↘ for the machines

   [M] carrying cost (reverify agent-seconds at tier cost + storage + audit amortization) vs ledger reads, per fact; prune below line. Tiers: shape-validation (fail-closed, exists) · re-runnability (exists; tokens CURRENT/STALE/UNVERIFIABLE) · binding audit (GAP; risk-tiered, quarantine, rate rises with authority). Missing keystone: stale-truth base rate. Kill: >50% underwater w/o sweep; economic number fails with producer costs; quarantine rate rises with fleet.
7. When trust travels, the size of a company changes. Companies exist because coordinating inside a firm is cheaper than across a market — and the crossover is set by trust overhead (Coase's theorem, the economics under everything here). A fabric that makes trust portable — proof that travels, checkable by anyone — lowers that floor: five hundred agents coordinated with the oversight cost of five people; eventually two companies coordinating at the trust cost of one. That's the load-bearing version of "500 at the risk of 5" — transaction costs, not courage. Kills it: the lead's attention-minutes per worker not falling as the fleet grows with the truth layer doing the coordinating — demote to aspiration, stop citing the line. The across-companies half stays parked (where the shipping-blockchain died — politics) until we own our supply and have proven the internal math.
   ↘ for the machines

   [J: Robert signs · M proxy] lead-attention minutes/worker (session telemetry: dispatch counts + turn-time) vs fleet size. Gate: fleet-scale proposals cite the trend. Kill: flat/rising with bus installed. Cross-org: stage-parked.
8. The final exam: the company can explain itself to someone who wasn't in the room. Everything above serves one end state: an organization that answers questions about itself — what's true, who decided why, what depends on what, what would happen IF — with proof, to someone who wasn't there. Two tests: the absence drill — can the truth layer brief a cold reader on "what was the lead about to do next, and why," well enough that they could act? Judged by a named human, kept honest by a planted control (the same drill run against a deliberately gutted truth layer MUST fail — a drill that can't fail is theater). The cold reader rotates: Alie first, then a hire — never only you. If the company can only explain itself to its founder, it's founder-shaped with extra steps. And this document: its own claims live as leased deposits, re-verified by someone who didn't write it — the first thing this machinery catches going stale should be its own mission statement. Kills it: the gutted control passes, or the rotating reader can't act on the brief two quarters running — "self-answering" demotes to aspiration and we say so out loud.
   ↘ for the machines

   [M currency · J drill, judge named per run, anchored on the gutted-genome contradicted fixture]. Cadence: quarterly with the review. Rotation: Alie → hire. Kill: gutted control passes (vacuous) OR reader cannot act 2 consecutive quarters.

How we'd know this whole thing is WRONG

Not late. Not under-resourced. Wrong. Any two together and the spine is failing on its own instruments — and we say so rather than explaining it away:

Stated plainly: none of the five can mechanically fire today. Wiring them is the first work after ratification — the quarterly review checks the wiring before it checks the world.

The business commitments (your court — signed on ratification)

Your calls