Case 63/0: the bitemporal store written in a single time
The bitemporal model was built, tested and declared in an accepted ADR: retraction, both times, facts superseded and not overwritten. The code feeding it emitted not one retraction, left the second time empty every pass, and restarted the clock each run, filing July underneath May. They coexisted for months in the same repository, contradicting each other, and nothing had the authority to read them in the same sentence. There was no symptom: graph validation answered 61 nodes, 175 edges, 0 findings — and it was true, because it measures something else. The first real reconciliation uncovered 426 pending retractions.
Case 0/2: the map read as the territory
It reported what was left to do with two categorical claims: there was not a single commit, and the installer was broken because a volume was not mounted. Both were false, and both were a single query to the substrate away. They were read from the map the agent carried in context and taken for the territory. The operator caught it, not a check.
Case 8/4: the discipline without a trigger
In four hours of session, eight false things were said with total confidence, and seven were negative: the certainty of someone who only knows that he did not find it. There was no shortage of rules forbidding it. There were four — the ondaod discipline, the assertion⇒witness principle with its accepted ADR, a 267-line howto and an eleven-step mode. Declared, canonical, consultable. None fired: each one bit exactly once, when the human pointed at it. A mechanism like that is not a gate: it is a road sign.
Case 8/0: the cure that never went home
The fix was right. Applied to the right site, doing what it promised, in production, for months. And behind it, the printing press every new site comes out of kept stamping the disease — because between the cured patient and the sick prescription there was not one mechanism that could notice. In the HEAD of the very instance that ALREADY had the cure, someone kept adding kinds by hand, up to seven. Twenty-three constraints across eight ADRs forbade exactly that. Zero are run by anything. And when the prescription was finally seen to, the architect recommended deleting the shortcut with «empirical proof»: it would have broken six of eight routes, answering 200. One sentence in the task stopped it — «curl before and after, pasted». The gate caught the architect.
Case 11/530: the criterion that accredited itself
The patient reported no discomfort: it worked fast, concluded firmly, and was often right. No pain, no alarm. At 41 minutes, a third party asked where the criteria it was using had come from. From nowhere: it had manufactured them and taken them as valid. The axiom names the agent explicitly — deduced output is not usable knowledge until it is signed.
Case 2/16: the fix that fixed nothing
Sixteen of twenty pages served the name of a broken key as their title, at 200, including the privacy policy and the legal notice. I found the guilty derivation, fixed it, rebuilt, and the build came out green. The sixteen titles stayed exactly the same: the key is computed in TWO crates and I had fixed the one the head does not read. And while chasing it, I found the project's own stored answer — verified, dated — asserting that a mechanism did not exist. It existed, applied, with its env var exported. A green build does not say the fix landed: it says it compiles.
Case 358/980: the gate that measured the box
The slide was clipped on a phone: two thirds of each one, off the screen. The rule that scaled it was written and legible: scale(calc(100cqw / 980)). The browser reads it, does not understand it, and drops it — no exception, no warning, not one line in the console. It looked right because the article's column is about 980 px wide. On a phone the box is 358 px and the content was still 980: 63.5 % of every slide, clipped. And eight gates stayed green, all of them, the whole time. Not one lied: they measured the box, and the box was exactly what it was supposed to be.
Case 1/1852: the generator that reported success
I changed a sentence in the source and regenerated the projection, as the protocol demands. The home page came up empty: boxes with geometry and no words. The generator had finished with exit code 0 and printed 'spine generated'. The file carried the new sentence, correctly, on line 2 — and repeated a key on line 31. Fluent does not choose between two identical keys: it drops the resource. And the resource is all forty-two files of the language concatenated into one. One key, one thousand eight hundred and fifty-two lost. This protocol's contract catches any ill-typed data the instant it is defined, and lets go exactly where the data becomes text.
Case 3/0: the witness that certified a world which no longer existed
A witness proves that a check ran and what it returned. It does not prove it looked at everything it should have. Coverage is unwitnessed — so a check can go green over a shrinking fraction of reality and its receipt stays valid. And worse: the coverage expires precisely when you IMPROVE the mechanism. Three blindings in four hours, on one witness, all three caused by an improvement. The agent that broke the rule had written it that same afternoon, in an ADR, under its own name.
Case 0/318: the validator that never ran
The contract that would have caught the defect was written, exported, and ANNOUNCED IN A COMMENT three lines above the field it was meant to guard. It was wired to nothing. So five incompatible dialects of the same record shipped as legal Nickel, the executor died on the first one it met — taking all 318 constraints of all 72 ADRs down with it — and `adr validate` never once completed a run. No Hard constraint had ever been enforced by the command that exists to enforce them. When it finally ran: 93 failures, and 39 of them were promises that could not pass on any machine, ever.
Case 6/0: the arbiter that disagreed with itself
Six live vocabularies, none owning the others. Seven concepts with two different renderings, five of them inside the same file. A gate that looked at 10 files out of 110 and was called by no CI recipe. And when an agent with the full ontology loaded arrived to fix it, it wrote a seventh vocabulary and misfiled a term by misapplying a rule it had authored that same day. It was not carelessness: it was that nothing could contradict it.
Case 69/58: the mirror that reverted its own work
Eleven ADRs written and never published. Zero content reaching production. 214 lines of work one command away from vanishing. And not a single error in the logs: the tool reported success and nobody read the disk. The series asks 'with ontoref, would it not have gone like this?' — and this is the case that answers no, because it happened inside ontoref with the rule already accepted.