How a certificate becomes a record.

Was: C-1 stated that "as at 16 Aug 2026, 170 of its certificates had been fetched from the laboratory's own host and each was byte-for-byte identical to the copy we hold."

Now: The figure is 169. Freedom Diagnostics rows carrying a lab-fetched sha256 in verification_response: 169; rows with verification_status = lab_confirmed: 169 — two derivations agreeing, measured 21 Aug 2026 with the row set unchanged since before C-1's stated date, so the entry was off by one on arrival rather than overtaken by intake.

A dated correction is exempt from re-derivation — the frozen-literal doctrine protects it — but the exemption assumes the literal was true at its own stated date. This one was transcribed, not derived, and the only write it can refer to touched 169 rows. A correction is a new claim starting at zero evidence: it ships with its own derivation or it ships a new defect into the one surface whose subject is defects.

Scope: the 169 byte-matched Freedom Diagnostics rows of its 368 in test_results @ epoch 20260821031815; the corrected count reproduced at epoch 20260821035548

Was: Six hand-written sentences describing relationships in the data: that a certificate from one named laboratory could not yet reach the top tier, that every certificate confirmable against a laboratory database came from one of two named laboratories, that confirmation and contamination-panel depth did not overlap at all, that independent confirmation had not been attempted on any vendor-published certificate — and, in this ledger itself, that every one of a named laboratory's fetched certificates had matched, stated in the present tense with a count that had since moved.

Now: Each of those surfaces states counts drawn from the same records it renders beside, and draws no conclusion the counts do not carry. Where a figure is zero the sentence still derives, so it moves when the data does. The ledger's own entry now carries the date it was measured, which is what makes it a record of a reading rather than a standing claim.

THE USEFUL LESSON IS NOT THAT FIVE SENTENCES WENT STALE. It is that C-1 had already retracted one of them, and the replacement it published SOFTENED the categorical instead of removing it: "can never reach our top tier" became "cannot reach our top tier yet". The shape survived the correction — a universal about a named laboratory, with a live count welded beside it — so the same sentence went stale a second time, by the same mechanism, in the text written to fix it. A correction that edits the wording and keeps the assertion has changed how confidently a false thing is said. Retract the SHAPE: state counts, and let the reader draw the conclusion. Everything else here follows from that. The five were true when written and were falsified by ordinary intake — a third laboratory became confirmable, and certificates carrying a contamination panel were confirmed — because nothing re-measures a paragraph. And the surfaces were re-read after the previous change only where a sentence had been edited, which is why static copy nobody had touched was the last place still asserting it.

Scope: Six sentences across three published surfaces and this ledger. Four were identified from the production database; the fifth was found by sweeping every universal or exclusive claim about confirmation, laboratories or panels across all four lab-testing surfaces before changing any of them; the sixth was found by the test written in the same change, which exempts a retraction's quoted claim but not the sentence offered in its place. No figure on any page changed as a result of this correction and no record was re-assessed: the counts were already correct, and the prose beside them was not.

Was: A confirmation metric that counted one of the two ways a certificate can be matched against the issuing laboratory. Four vendors were shown as having no lab-side evidence at all; on the largest, that was 98 certificates reading zero. The same records were labelled "Vendor-published" on the certificate browser and on their own record pages, and the purity figures on peptide pages left them out.

Now: The metric counts both routes — documents taken from the laboratory's own database, and vendor-published documents whose bytes we matched against the laboratory's own copy — everywhere it is computed, and it is named for what it measures rather than for the word "confirmed". Each record states which of the two routes applies to it.

One word covered two populations. The overview counted 407 records as confirmed while the leaderboard scored 238, one click apart, because the tier read the evidence and the score read a single status value. The peptide-page purity figures had the same gap. Correcting one surface and not the other would have left the same defect with a smaller blast radius, so both were corrected together. No vendor's underlying evidence changed and nothing was re-checked: every one of these records already carried its confirmation, and the surfaces were reading past it.

Scope: 167 of the 807 records this site scores; 169 of the 1,012 held @ 2026-08-16 (epoch 20260816120451). On the leaderboard's own population — 23 vendors, the certificates each publishes — 4 vendors understated, 0 overstated, 19 unchanged. On the peptide pages, which count only primary results: 204 rising to 266, moving the median purity on 20 compounds by at most 0.79 points, and adding a purity summary to four compounds that had too few results to show one. These are three different denominators and are not comparable with each other.

Was: Nine certificates shown under "Lookup pending — the lab publishes a lookup; we have not run it."

Now: Confirmed. The lookup had in fact been run on 30 July: each of the nine was fetched from the issuing laboratory's own host and its SHA-256 compared with ours. All nine matched.

The result was written to the record and the status field was never moved, so every surface deriving a tier from that field went on reporting the check as outstanding. The evidence was in the database the whole time; the answer was read off a status column instead of off the evidence. That is the defect this entire ledger exists for, and it is ours.

Scope: 9 records, all at one vendor, of the 1,012 held @ 2026-08-16. Confirmed 2026-07-30/31, corrected 2026-08-16 — seventeen days. The date the evidence was recorded is measured (first_checked_at, in the data); exactly which sentence rendered on which day is not, because the tier vocabulary changed during the period.

Was: A column headed "Confirm" showing, for each vendor, the share of its certificates at labs whose lookup WE had automated.

Now: A column headed "Lab lookup" showing the share whose issuing lab publishes a lookup a reader can work. The leaderboard reordered. NO VENDOR'S SCORE FELL — the metric moved every score in the same direction, so every apparent drop in rank is displacement by another vendor rising, not a decline by the vendor that moved down.

The score read a column recording our own pipeline coverage and rendered it as a claim about the laboratory. The ranking was not wrong about the vendors; it was measuring us and labelling it them.

Scope: all 23 ranked vendors; 430 of 661 ranked records changed value @ 2026-08-15

Was: That consecutive lab report numbers showed two brands had submitted a shared batch together.

Now: Stated as what it is — an inference from matching compound, purity and measured mass. Report-number adjacency does not hold for most of these pairs and is no longer offered as evidence.

A pattern that held for the first examples looked at was generalised into a stated mechanism, and the mechanism was then published as the reason.

Scope: the shared-batch pairs rendered on the leaderboard @ 2026-08-15

Was: A corpus of 728 certificates, with tier counts of 238 / 201 / 289 — a partition of OUR QUEUE STATE presented as a partition of what the laboratories can be checked against.

Now: The partition derives from the laboratory's own lookup_status, and it is published in full: 238 taken from a lab's database, 167 confirmed by byte-match, 291 pending and 267 unverified at labs that publish a lookup, 17 at a lab that publishes none, and 30 at labs nobody has assessed.

The tier was labelled with a property of the LABORATORY and computed from a property of OUR PIPELINE. Those are different partitions, so the counts were a correct measurement of the wrong question — and the largest single instance of the defect this ledger exists for.

Scope: the 1,010 records carrying a resolved compound, of the 1,012 held @ 2026-08-16 (epoch 20260816120451). The seven figures sum to 1,010 and are counted independently, none by subtraction.

Was: Horizon Analytical recorded as publishing no lookup, with a note calling its certificates structurally uncheckable.

Now: Horizon Analytical publishes a public lookup, no login, keyed on a report number printed on the certificate. Our own Horizon records still carry no such reference, which is a separate fact and is now recorded as one.

The register field describing the LABORATORY was set from a fact about OUR RECORDS. Both statements were true; only one of them was about the lab, and the wrong one had been stored.

Scope: 14 of the 807 records this site scores; 14 of the 1,012 held @ 2026-08-16

Was: That a certificate from Freedom Diagnostics could never reach our top tier because there was no portal to check it against.

Now: Freedom Diagnostics does publish a public lookup, and we have run it: as at 16 Aug 2026, 170 of its certificates had been fetched from the laboratory's own host and each was byte-for-byte identical to the copy we hold.

Two surfaces read two different columns for the same claim. The overview read lookup_status; the register read lookup_url_template, which is a display field and was empty for this lab. The sentence stated a fact about the laboratory that was drawn from neither.

Scope: 365 of the 1,012-record corpus; 236 of the 807 records this site scores @ 2026-08-16