Generated from cat-harness/skills/sdlc/sdlc-core/untainted-verification.md — do not edit here.
✎ Edit this page’s source

Untainted verification — the producer never writes the verdict

Skill id: untainted-verification · Package: sdlc-core

A secret scan, a schema validation, a back-translation and a systematic literature search differ in what they read and in what counts as a finding. They do not differ in this:

The party that produced the artefact does not get to write the verdict on it, and the checking party is given a controlled context rather than the run of the repository.

Everything below is that sentence made operational. It is generic on purpose — translation-manager.md §”The agentic round trip” is where it was first written down, in translation vocabulary, and taking the translation out changed nothing structural.

The parties

party	is given	produces
producer	the task	the artefact
checker	the artefact’s declared `checker_sees`, and nothing that would let it shortcut	an independent rendering or finding
adjudicator	`adjudicator_sees` and the checker’s output — never the artefact itself	`pass` / `warn` / `fail`, each drift named

Three rules, and the failure each one stops

1. Neither party sees what would let it shortcut. A back-translator shown the English writes the English back and the check passes vacuously. An adjudicator shown the French can talk itself into any reading of the back-translation. Neither failure raises an error; both produce a green verdict that measured nothing, which is worse than a red one.

2. The checker uses no tools, and says so. The artefact is in the repository. A checker with filesystem access can find what it was not given, and the verdict then measures its search rather than the artefact. Ask for a TOOLS_USED line and record the answer — UntaintedParty.tools_used. Absent means nobody asked, which is not the same as “no”.

3. One party doing both halves is not this check. It compares a thing with its own paraphrase of itself.

The context is declared, not composed

UntaintedDispatch on the criterion (schemas/block-qa.ts) declares checker_sees, checker_withheld, adjudicator_sees and drift. A caller names a subject and a criterion; it does not hand-assemble a brief.

That is not ceremony — the party composing a brief by hand is the producer, and rule 1 then breaks silently, one call site at a time. A declaration is also checkable, and untaintedPartitionDefects checks it:

defect	what it means
undeclared	no `untainted` block. Reported, never skipped: nobody said is not nothing to check
overlap	a role visible to both parties — one of them is grading its own input
unpartitioned	a companion in `depends_on` that is in neither visible set; usually a role added later. Nobody decided
phantom	a role named in the declaration that the criterion does not depend on — the declaration describes something else

Totality is the point of checker_withheld. Without it a newly-added companion role defaults into the checker’s view, or out of it, and either way the separation moves without anybody choosing. With it, the partition simply stops being total, which has a name.

drift must say both halves. Told only what to look for, an adjudicator returns a style review and everything fails. The translation round trip states it well: synonyms, articles and re-ordering are not drift; a claim added, dropped, weakened, strengthened or reversed is. The check verifies the string is non-empty and nothing more — “states what is not drift as well as what is” is a reading, and a gate pretending to measure it would be the exact defect this skill exists to stop.

Three states, and the third is what lets it gate

verified — the adjudicator ruled. Two entries, adjudicator first.
could not dispatch — no dispatch capability was available. One entry, result: "n/a", metrics.dispatch: "unavailable", and a required reason. The producer may write this one, and only this one.
not attempted — no entry at all.

State 2 is the owner’s ruling of 2026-09-21, and without it the discipline cannot gate anywhere. A producer forbidden from writing a verdict has nothing to write in an environment with no subagents, so the gate reads an absence it cannot tell from negligence. Recording it keeps “could not dispatch” distinguishable from “nobody tried” — the same three-state rule ci-health and readme-sections already apply, reached here from the opposite direction.

It is never a pass. isVerified() is the predicate a gate asks, and it is false for state 2 and for state 3 alike. A consumer that reads n/a as “nothing to see” turns the honest gap straight back into the silent one.

Recording

Both parties are written as witnesses. The adjudicator leads — the first entry per criterion is the operative one everywhere in this repo. The checker sits behind it with result: "n/a" and its output in notes, because a reader asking on what basis? needs the intermediate and a reader asking who did this? needs both names.

Record model with model_source, or not at all. A subagent’s serving model is not observable from the session that dispatched it: it inherits the parent unless the harness overrides, and the hand-back does not say which model served the turn. A bare model string is an inference printed as a fact. Absent both, a reader gets “not recorded”, which is true and acceptable.

A sweep replaces only entries whose reviewer is itself. mergeUntainted carries everything else through. This rule exists because translation-block-qa once deleted a round trip a pair of agents had produced, on an unrelated re-run, silently, with nothing in the output to say so. A human ruling is never superseded here.

A verdict is hashed to what it was about. Edit the subject and the verdict goes stale. This matters more for an agent ruling than a script one, because nobody can cheaply re-run it.

Where the verdict is kept: the attestation store, not the derived file

Both entries are judgements — reviewer.kind is agent (or human) — so their durable home is the attestations graph, test/attestations/<family>/<mirrored subject path>.attestations.json (qa-attestations/v1), on main. The derived test/results/ file beside them is a working copy whose record lives on the qa-reports branch, and a regeneration cannot reproduce an agent ruling (arc 3fva, owner ruling D2 (a)).

recordUntainted writes the store first, through the same resolvePrior / finalizeCriteria API as every other attesting writer (bean 8iqt), and refreshes the derived report only if one already exists — it never invents one. So the verdict lands whether or not the derived file is on the branch, and the store file it returns is the one to commit. What puts a subject in scope is the subject existing (<subject>.md or .ts), not its derived report, which no longer lives beside it. A corrupt or unreadable store reads UNKNOWN and is refused, with nothing written, rather than overwritten.

Why the instrument is the thing to watch

A measurement shipped with a description explaining its own failures away as expected “with limited vocabulary back-translator”. The back-translation map it scored against held six entries for thirty-six strings, so the failure count was a count of absences.

A measurement whose author has to explain it away is about the instrument, not the subject. All of it was removed — the numbers, the field, the badge and both scripts. The incident in full, with its numbers, is in translation-manager §”The instrument, not the translation”, where it happened.

This is one rule with two homes

It is not a platform invention. Guideline methodology separates the systematic review team from the guideline development group for the same reason, and 8rwa carries the identical rule into evidence-based literature review for living guidelines. When adding a domain’s rules here, run them through domain-fencing’s three questions first: a rule that fails all three is a folio’s rule, correctly located, and fencing it and saying so is the honest move.

translation-manager.md §”The agentic round trip” — the first instance, and the translation-specific drift definition that does not generalise.
permissions.json qa-reporting — who may emit a verdict at all. Bean a58y gives it its first consumer.
pipeline/untainted-verification.ts — the recorder. It deliberately does not call a model: the payload is written by whatever ran the parties, so recording works the same for two subagents, two API calls, or two people.

Processes that run this skill

process	step(s) that name it
Adjudication	Dispatch with adjudicator_sees — never the artefact
Test-plan execution	Re-execute a sample, blind to the tester’s verdicts
Ingestion subprocess — the L1 completeness gate	Round-trip translation QA