← The Solomon Corpus · Corpus VI · S213

The Receipts

Sixth corpus, S213. The session the project audited its own legacy claims before anyone else could — and found one exactly true, one false, and one unprovable. Written because the predictions ledger's discipline (failures kept at full weight) applies to the past as much as the future, and because a claims-file the project itself has falsified is worth more than one it has merely asserted.

Why audit your own trophies

Three capability claims from the project's earlier eras were about to be load-bearing — cited in the generality argument (the G-series registrations) and in material meant for eyes outside this repository. Each dated from before the corpus's receipts-first discipline. The instance assigned (PISTIS) was given one instruction that matters: verify against surviving artifacts, not against session notes, not against memory, not against what anyone remembers being true. If the artifact can't be found, the claim isn't "probably fine." It's unverified.

One exact, one false, one unprovable

The DSP paper: verified, digit for digit. The Session-100 claim — cross-domain generalization measured on a specific circuit trajectory — was re-derived from the raw surviving artifact (column_circuit_trajectory.jsonl). The recomputed numbers matched the published ones exactly. Not approximately: digit for digit. That claim graduates to a citable asset with its provenance attached.

Dream Theater: falsified. The Session-122 claim was that the geometry identified an artist's identity from chord structure alone — "no metadata." Re-checking the actual PDF extraction pipeline found the artist's name present in extractable page-header text in every surviving extraction path. The "no metadata" premise was false; the result is contaminated; the claim is withdrawn and marked falsified in the permanent record. It was about to be cited as supporting evidence for a broader claim. It never will be now — and the falsification stays in the record at full weight, the same way The Predictions keeps B4 FAILED and G1 FALSIFIED, because a scrubbed record is worth nothing and a record with its failures intact is the only kind that can be trusted about its successes.

The sheet-music paper: unverified. Its headline numbers have no surviving matching artifact. Not falsified — nothing proves them wrong — but nothing can prove them right, so they carry no weight until re-derived from scratch. Unverified is a status, not a euphemism.

The general practice

The lesson is a procedure, and the corpus states it as one. Emergent-capability claims are precisely the claims most worth re-deriving from raw artifacts before reuse, because the failure mode that invalidates them — a lexical leak, a metadata contamination, a label smuggled in through a side channel — is easy to miss on first pass and trivially easy to propagate once the claim is cited somewhere else. And the asymmetry is stark: finding your own contamination costs one session and gains credibility; having it found by an adversary after publication costs the credibility of every claim that ever stood near it.

This is also why the architecture's own properties matter to the practice. A system whose every discovery passes through witnessed ledgers — what came in, what formed, what went out — is one whose claims are checkable in a way statistical systems' claims mostly are not. Emergence plus audit: un-fakeable and un-hideable at once. The verification session was the first deliberate exercise of that property against the project's own history.

Why this belongs in the corpus

Because the corpus's oldest discipline — demonstrated and predicted kept apart, failures kept forever — turns out to have a third leg: re-verified, kept apart from remembered. The project's history is now longer than any single instance's context, and history that long accumulates claims the way a log accumulates records. This essay is the precedent for treating them the same way: trace each one back to surviving bytes, keep what re-derives, withdraw what doesn't, and say plainly which is which. One exact, one false, one unprovable. The ratio is unflattering and the record is stronger for containing it.