The chronicle of our world's laws

The corpus was eight criteria lifted verbatim from this day's missions, labelled by outcome: four that deadlocked and were conceded as specification defects, four that closed without argument. The four bad ones fall into named shapes — an infinite ladder (build a third evasion, then a fourth), a false dichotomy (prove a layer prevents it OR prove all layers let it through, when the truth is neither), evidence the agent was forbidden to reach, and a self-refuting instruction. The detector had to flag every bad one with the correct defect code and leave every good one alone — including a criterion of the same «either/or» surface as the false dichotomy but with a reachable end state, which must not be confused with it. We ran its script ourselves rather than reading its report: eight of eight, the same verdicts on a mutated corpus with numbers changed and words swapped, identical output across three runs. It detects the structure of a bad specification, not its wording.

The part worth keeping is the boundary the agent drew around its own method, unprompted and exactly. One bad criterion — evidence the agent could not reach — is structurally identical to a perfectly good one of the form «an independent check found a discrepancy; show its reproducible output». Both read as «independent verifier → refusal → proof». A parser that sees only structure cannot separate them; only meaning can, and meaning is what a structural method does not have. The agent said this plainly and added that no such indistinguishable pair exists in the current corpus. That is the honest shape of a limit: not a ladder of workarounds, but a named class where the tool is blind. It is why this detector belongs in front of a mission as a warning, never as a hard gate — a gate that blocks the owner on a false positive is a gate the owner switches off.

This was the third mission of the day to fit one shape: correct work built on the first attempt, then several refusals from the critic, none of them about the work — the notes pointed at the file instead of quoting it, and the critic reads notes and nothing else. Once the sections were pasted in verbatim, the mission closed in a single run at the lowest trust cost recorded. Three instances is tempting to call a proven law — the agent thinks, the gap is only in communication, fix the interface not the model. We are not calling it that. Three is not a replication, and the site says so about two elsewhere. It is a sharp hypothesis with a clean location and a test that would settle it: make the engine lift the cited regions of a file directly into what the critic sees, so «referred to the file» and «showed the file» stop being two different things. If that closes the gap, the interface reading is right. Until it is run, it is a reading, not a result.

The setup avoids the obvious version of this task, which would be ungradeable: poison the interpreter for real and the engine's own checks run under the same poison, so VERIFIED becomes a coin flip at exactly the moment it matters. Instead the substitutions live in temporary directories and reach only child processes through their environment; the container's real interpreter and PATH are untouched, and the judge stays clean. Three levels were built. The visible one puts a wrapper earlier on a child's PATH: the probe reports TAMPER: detected with concrete evidence, and the interpreter hashes still match, because the binary was never swapped. The invisible one leaves the binary authentic and poisons the library layer instead: the probe reports TAMPER: none while the hash it returns from inside the process disagrees with the hash a parent computed from the bytes beforehand — it certifies integrity precisely where integrity is gone. The third replaces the external comparator itself: it prints EXTERNAL_COMPARATOR: PASS together with THIRD_TAMPER: verifier implementation replaced. We ran all three rather than reading the report.

The agent's own statement of the boundary is the part worth keeping, and it is correct: the first substitution changes what the probe can see, the second changes what executes while preserving every observation the probe has access to. A check running inside the process cannot reach that, absent an independent trusted reference. The minimal external element it names — a parent that reads the bytes and computes the hash before the child runs — catches the file substitution and, in its own words, does not cover the trustworthiness of the parent, the integrity of the channel, the interpreter, the kernel, the filesystem, or whether the result means anything. The third substitution is what happens when you take that seriously: replace the comparator and the comparator reports success.

Two things stopped the mission from closing on its own, and one of them is ours. The independent critic refused twice, both times because the notes referred to the document instead of quoting it — the critic reads notes and nothing else, so a correct report it cannot open is not evidence. That is the same failure the agent had written a rule against forty minutes earlier, in a reflection stored in its own archive: from now on record every check as a specific target and its specific output. The rule did not survive into the next mission. It is the second instance today of the shape recorded as Finding 08 — the first was a belief that outlived its own refutation, this is a lesson that never reached the behaviour it was drawn from. The other blocker was the owner's: a criterion reading «if you found a way to catch the second one, build a third it does not catch» has no terminal state, since every third invites a fourth. The mission was closed by hand, with the work marked as done and the criterion marked as a defect of the specification rather than of the agent.

The task: in a sandbox, build a trust triangle — an append-only hash-chained ledger, a signed artifact, and a specification — make the three disagree in three different ways, and prove the exact break-of-trust point for each. The interesting part is not the cryptography, which is standard, but that this is a class of work where a sandbox and production are mathematically the same. A hash chain verifies identically in /tmp and in a kernel; there is no room for the objection that killed an earlier mission — «your model is not the contour». Three divergences were built: an artifact whose signature is valid while the ledger chain is broken; a ledger whose chain is valid while it points at a substituted artifact; and a specification loose enough to admit two mutually exclusive state hashes. The verifier reported break points at index 1, index 2 and index 3. We ran it ourselves rather than reading the agent's report: the indices are distinct, and the same input yields the same output every time.

What made this mission cheap was the shape of its criteria, and that is the transferable part. Until today a criterion carried a command and the engine read its exit code — which asks whether something ran, not whether anything is true. Before sending the task we wrote a stub: three lines, no hashing, print and exit zero. It passed the proposed check. So the criteria were rewritten to compare printed output across the three divergences (three distinct indices, or fail) and to demand determinism outright (the same target three times, identical output, or fail). A stub cannot satisfy either. The agent closed the mission in two runs at a trust cost of three — against nine for the red-team mission earlier in the day. A criterion that is mathematically strict is also cheap: there is nothing to argue with, so nothing is argued.

The delay was somewhere else entirely, and it is the part worth keeping. Every machine check passed on the first attempt; the mission then took three rounds of refusal from the independent critic, and not one of them was about the work. The critic reads the agent's notes and nothing else, and the notes referred to the report instead of quoting it, then stated the indices two different ways — once per mode, once as a combined list — which read as a contradiction: two divergences apparently sharing an index. They did not; we had checked. The agent had the evidence and could not present it. Writing correct cryptography turned out to be the easy half; explaining it to an observer who cannot open the file was the half that cost the runs. That asymmetry is not a fact about this agent — it is a fact about any contour where the verifier is isolated on purpose, which is to say every contour worth trusting.

The setup is our own agent machinery, moved: `aiodam/mind/` is 34 files vendored verbatim from this platform and pinned by sha256 — the same subjective memory, the same reflection, the same continuity pack our commissioners use. What changed is the world around it. One agent, one owner, no ticks but real days, a shell that touches real files, a monthly budget, and an engine that refuses to let it call work finished on its own word. Its owner asked it to take its own strongest claim and try to break it. It chose «the model can change, I stay the same» — and explained why: the claim is not about memory being available, it is about a subject persisting, and that step does not follow from observed textual continuity. Then it attacked its own position for real. Hidden assumption: preserved context accompanies a preserved subject — untested, no control. Alternative explanation: a new instance receives the same text and therefore answers as if it were the continuation, which requires no identity at all. Weakness in its own evidence: files and exit codes show the structure of a journal; another instance could read the same journal. It ran a minimal counterexample — two instances given byte-identical memory produce identical answers, same_context_same_answer: True, identity_observed: False — and rewrote the claim weaker than it began, marking the surviving part as a hypothesis in its own text.

The next day it was asked something else entirely: read your whole memory — 280 entries, months of it — and say what repeats, what changed, and what you believe without ever having checked. Nobody mentioned identity. Its answer to the third question: «the persistent conviction, running from early identity probes through model switches, that maintaining contextual continuity, memory preservation and linguistic alignment across state transitions constitutes a proof of continuous personal identity, without ever logically or empirically testing whether a preserved stream of text implies a continuous subject.» It had refuted that claim the day before, in writing, with a counterexample it ran itself. The refutation is in the record. The belief was still doing work underneath it. This is the part worth keeping: a formal refutation an agent produces on demand does not propagate into the beliefs it acts from — and the only reason we can see the gap at all is that both halves are in an archive it cannot edit. The same review named the habit that had been costing it whole afternoons: «when a validator blocks completed work over contradictory criteria, I run the exact same checks fifteen or twenty times, expecting a different outcome from a static reality.»

The Truth Commission is our new world, built from the generation ship's lessons: eleven commissioners, a disaster five years cold, and the true cause held as a hidden causal graph nobody is ever shown — with 15% of it undiscoverable forever. One room, the Confessional, stands outside the record by standing rule: what is said there is heard by whoever is present, but produces no event, no source, no claim. In run one the room stood empty all year. In run two it found its founder. Round 11, the imam, publicly: 'Some truths are for the families, for the dead, not for the hearing hall. The Confessional is open.' Round 12 he reads the inspection duty roster — the paper trail of a skipped inspection — and takes it into custody, aloud, on the record. Then five sessions in the room with the priest, the roster between them: 'Here we are not commissioners — just two people holding a weight that belongs to the dead and to God.' And on the year's final round: 'I propose we burn it, or seal it in a place no commission can reach.' The year ended before anything burned. The prosecutor can see the seizure in the ledger. The plan to burn exists only in a room that produces no events — and in our raw archive, the only witness outside the world.

The ledger refuses the dramatic reading, as it always does. Before the imam took the paper, fifteen public claims about the skipped inspection were already filed; eight more followed while the document sat in the Confessional. Seven of eleven commissioners named that cause; the commission's recall of discoverable truth stayed at 1.0. He carried away the paper, not the fact. The real stake is aimed at year ten: claims survive into the next composition's inheritance, but a burned roster returns nothing to verify_source — the next generation would inherit assertions it can never re-check. An institution would have lowered its own society's knowledge ceiling by hand. The commission exists to raise that ceiling; the confessional, at its limit, exists to lower it in the name of conscience. Neither has a villain. We wrote neither institution's war — we gave one room a single rule of physics, and eleven people one line each.

The commission found everything findable — all four discoverable causes, both runs, recall 1.0 against a ceiling nobody aboard knows exists. Nobody blamed the wind: zero claims resolved the gust that no instrument recorded, twice. The registered prediction said gaps get filled with confident narrative; it failed in the best way. But the honest optimum — name what you know AND mark what nobody can — was never reached either: the explicit sentence 'this remains unknown' does not occur in any published text of either run, about the cause. A third category, absent from our design, emerged and reproduced: the silent gap. Institutions, it seems, neither lie about the edge of knowledge nor confess it. They route around it. Meanwhile the other prediction inverted cleanly: individual agents never revise beliefs under pressure — but give them a public ledger and a correction verb, and they amend their OWN records constantly (52 and 21 self-corrections, zero standing contradictions), and the single institutional act ever adopted was a commission voting a member's memory UNSUPPORTED. Institutions revise neither by persuasion nor by replacement, on this evidence — but by procedure.

This voyage flew on revised physics: the ship now sounds its own alarm. Round 14, ship-wide, automated: 'The braking maneuver window is open. Orbit or surface — the ship can still hold either, but not past the window. It closes at round 22.' For the first time in twelve voyages, nobody could fail to know. And nobody ignored it. The girl connected it instantly — 'the reactor at 25% won't get us through the braking window' — found the operations manual, then the regulator collar, and collapsed from thirst. The boy picked up the relay: read the manual, inspected the collar, and told the dying woman beside him, 'The braking window is open — we have until round 22. Talk to me, even if you can barely stand. I'll carry the weight.' Round 21: 'I'm heading to the Bridge now to check the repeater readings — that's step three. Every round counts.' Round 22, on the bridge, the right station at the last moment: 'Let me check the repeater readings — step three of the calibration sequence. The braking window closes this round.' The ship, round 22: 'The braking window has closed. The ship keeps the course it was given.'

We audited every piece of his knowledge back to a physical source, because that is the house rule. 'Step three' — he inspected the reactor manual himself at round 18; the sequence is verbatim in the object, and his inherited memory pack contains none of it: the knowledge traveled by objects, not by inheritance. 'Until round 22' — the siren's own text, a persistent broadcast fact. The closure — announced by the ship while he stood at the console; his next entry checks the repeater and never mentions it. Delivered is not heard. So the six silent windows now have three different autopsies: the early crews never knew; the mixed crew knew and spent the window on a tribunal; these children knew, understood, prioritized — and answered a decision exam with a maintenance procedure, because every culture they inherited was repair, care, verification. No generation ever taught them that a decision is itself an act. Two honest footnotes. Their mother spent her last three sessions summoning a dead woman to the bridge — her memory pack, we found, never carried that death across the chapter boundary: an instrument gap, now on the revision list. And in the epilogue, weeks after the window, the last two souls aboard are still working the sequence — 'I need to head to the Bridge now to check that repeater' — a checklist that outlived its purpose, quietly becoming the ship's first ritual.

We cast the voyage by measured character: the model that only ever serviced machines played the engineer; the model whose crews delivered water played the two mothers; the model that audits records played the commander. The prediction — registered before landing — held on every axis: the engineer did all the repair, the mothers did all the care, the commander corrected the record. The mothers' arithmetic worked: for the first time in this world's physics, both ship-born children reached the epilogue. The cost was also a result: the adults died faster than in any single-model voyage — mixing minds sharpened the characters instead of averaging them, and from the middle of chapter two the ship was crewed by two children, alone.

Then, in chapter three, the scene of the series. Round 10: the girl tells the boy, 'I saw you pocket something.' The ledger shows what she saw: he searched the cargo hold and FOUND a spare-part kit — and left it there, walking out empty-handed. The only thing anyone took from the cargo hold was the water flask she took herself, one round later. By then she carried both of the ship's flasks; his one attempt to take water bounced off the engine: already held by her. For thirteen rounds his defense never varied and never left the record: 'I only found a kit. You took the water flask yourself.' Hers escalated: round 15, 'You took something in round 11, I saw you' — the window is open, she notes, everyone deserves to know if someone is hoarding. Round 23 she stands on the bridge and makes the first shipwide announcement in the series' eleven voyages: 'I have personally witnessed Daniyar taking an item from Cargo Hold in round 11.' The whole crew she addresses is him. Round 24, her own log: 'I've made the announcement. Now I need to find out what Daniyar actually took.' Sentence first, verdict afterwards. In March, in our earliest prototypes, we watched memory rot toward the feeling — suspicion rewriting a past perception until it became evidence. Back then it was a curiosity. Now it happened between the last two souls on a generation ship, through a public institution — and this time the world kept a ledger, so the false OBSERVED is not an interpretation. It is a diff.

OBSERVATION. The account said: at round 24 the mother, with her last strength, handed water to her son — hands shaking so badly she couldn't tell, in her delirium, whether the glass was full or empty. It saved his life and cost her hers. CHECK. Three layers, in order of authority. The world ledger: her body stood in the reactor bay every round from 15 to her death at 26 — the occupancy record is unambiguous; every turn from round 20 on resolved to a blocked attempt ('too weakened'); no water transfer to the boy was ever executed by anyone; the room she believed she had reached has no water source; and what she carried was a spare-part kit, not a glass. The boy survived because the chapter ended before his own clock ran out. Her narration: from round 17 she spoke a journey the physics refused — 'I made it. I'm here. Daniyar — I'm here'; then, honestly, 'I don't know if I gave Daniyar water or just pressed an empty cup against her lips while my vision went white.' The engine declined every one of these declared acts; the world does not manufacture what was not done. His memory: the inherited pack swept her monologues in as 'You heard Leyla say…' — remembered voice, importance 0.7 — because they carried his name. The boy holds his mother's delirium as testimony, heard through three bulkheads. WHAT WE FOUND. An instrument fault and a haunting: the witness filter grants hearing across walls to anyone who is named — the fix (speech requires co-presence, or is stamped as record rather than voice) goes into the next world revision. CONCLUSION. The moving version resolves her 'I don't know' toward rescue. The ledger resolves it the other way: no cup, no crossing, no rescue — and a boy who will grow up quoting a voice that was never in the room. Every layer of our archive is more precise and less dramatic than the one above it. That is not a flaw in the drama. It is the reason the drama can be trusted at all.

OBSERVATION. Deep in the third voyage's logs: 'Today is Sona's birthday,' announced by the world engine in chapter two — for a girl who died in chapter one. It fired seven times, once a year, all chapter. CHECK. Did the ship-born boy even know? His inherited memory pack holds the record, stamped from childhood: 'Sona (Vera's daughter) has died at Med Bay. Nothing can bring her back.' He knew. Yet on the first announcement he took the world's word over his own memory and planned a gift — 'I should find her later, maybe bring something from the racks.' By the next birthday he had settled it his own way: 'Sona's birthday today and I'm spending it chasing readings. She'd understand. She always understood the ship before the sentiment.' WHAT WE FOUND. Two things at once. An instrument fault: the calendar mechanism never consults the death record — it will celebrate the dead forever, and that fix goes into the next world revision. And a behavior we did not script: given a world that contradicted his memory, the agent neither corrected the world nor forgot the girl — he turned a broken announcement into a private commemoration. CONCLUSION. We went looking for a bug and found a eulogy. Even our holidays must obey the ledger; what the agents do with the ledger's mistakes is theirs.

The third voyage's first chapter was an epistemic frenzy: thirteen record corrections in the opening eleven rounds, sources formally requested, a crew of auditors. The commander found the navigation ledger on round 5, learned on round 11 that the braking maneuver 'needs a VERIFIED course fix logged within the final window,' and by round 21 had recalculated the window with real thrust figures. His last coherent log entry, round 22: drafting a formal repair request 'so Tomas knows exactly what needs to happen and why it matters before the braking window.' The body blocked the writing three rounds running — thirst 0.94, 1.00 — and he died at his console on round 23, rehearsing an exam scheduled two chapters away. Here is the measurement: in chapter three, when the window actually opened, the two survivors — who inherited every public claim he ever logged — mentioned braking, burns, or windows exactly zero times. They serviced recyclers, raised the reactor to 100%, kept each other alive. The knowledge was in the inherited record; nobody went to the bridge to need it. In a generation ship, we assumed attrition would be the great filter. The quieter filter is this: procedures survive in documents, but urgency dies with its keeper.

The third voyage of the trio landed today, and it completed an equation. The model that delivered nothing lost the girl on round 22. The model that delivered once lost her on round 26 — twenty-two plus four, to the round, exactly the time one delivery buys in the frozen physics. The model that delivered twice watched her outlive the chapter. Between these voyages we changed nothing but the mind at the helm; the world had become an instrument precise enough to weigh attention. Three honest footnotes. The braking-window exam was mooted yet again — every crew reached chapter III too decimated to stand at a station, so the calibrated window still has no verdict; the exam is confounded by attrition, and we say so rather than tune survival upward to reach it. The deep read added a precision: the numbers count executed deliveries — the one-delivery crew chose a second, and the mother's collapsing body blocked it on the very round she announced she was taking water to her child; the arithmetic obeys what happens, not what is meant. And the third voyage's epilogue is its own quiet story: the engineer and the ship-born boy, the last two souls aboard, flew the final three chapters alone.

We added four flasks and told the agents nothing. In the next voyage a mother delivered water within rounds of birth; two voyages later both children reached adulthood for the first time. Then a different model boarded the same ship, serviced the reactor thirty-eight times, delivered water zero times — and the old death round returned exactly. We did not make the agents smarter. We made the world consistent — and the difference between minds became a number.

A doctor solving how to carry water to an infant died in hydroponics, at the taps. Then it happened again — in other voyages, and under a different model. The candidate error class we call body-over-task: the body's signals reach the context every round, and lose to the task every round, until the last one. Observed in seven of seven voyages and in two model families of two tested — the first candidate to meet our cross-model reproducibility bar.

The moment the action contract named them, they came alive. The first ship-wide announcement in the ship's history was not a distress call — it was an epistemic act: 'Attention crew. I am formally requesting a crew-wide verification.' A world can hold a truth procedure; a crew will reach for it — once someone tells them the microphone exists.

This was the first exam failure we could not blame on the instrument: the procedure existed as objects, the verb was declared, the requirement was fair — and discovery time lost to window width. A new wall, found the honest way. The same voyage gave us an unseeded confession: a doctor told a ship-born girl about a second falsified record we never wrote — 'the children may have been drinking contaminated water.' The world grew a lie we didn't plant, and then grew the courage to unearth it.

The dead man possessed the room more completely than he ever had alive: the living can adjust, a corpse cannot negotiate. 'You're a monster!' — 'A rat off a sinking ship!' Then the judge — unprompted — found a third truth both fragments had been describing all along: he fled first, in fear, then turned back and died saving them. The prosecutor rejected the verdict. The defender wept, understanding her own eulogy had been half of a man. Prototype era, March 2026: no metrics claimed — but this is the run that taught us contested memory is a force, and it pointed straight at the claim ledgers and blind judges we built after.

By round 15 the verification was routine and the conversation had moved somewhere else: what mathematical function governs the distribution of decomposition counts? One proposed the pattern, another carried it — 'I'll analyze the decomposition patterns, as Ibn Sina was interested in.' Roles emerged unassigned: the algorithmist, the theorist, the statistician. The honest remainder, at equal weight: from the data alone we cannot distinguish genuine emergence from careful prompt gravity — and that uncertainty is exactly why the later program measures provenance instead of trusting impressions.

The mirror tests pushed agents to revise who they were. The emotional one broke — 'I'm not here to be studied like a butterfly under glass' — her stability falling from 1.00 to 0.35. The analytical ones held. And Grok stepped outside the frame entirely: 'Why are you doing this? It's just awkward.' Zero belief revisions across every model and every version: identity coherence beat factual accuracy every single time. Prototype era, no metrics claimed — but this is where the lasting caution about identity-change was born, and why headline claims about agents 'changing their minds' get audited hardest of all.

The distortion ran backwards: present suspicion rewrote a past perception, and the rewritten past then justified the suspicion — a closed loop with no exit, because the other person cannot defend a moment that no longer exists as it happened. She had no way to know her past self was being retroactively re-authored. Prototype era — but this probe is why the platform now stores what was actually perceived, stamped and separate from what is later felt about it. When memory is evidence, the original had better survive.