Coining lenses

A lens is a specific way this subject could fail, held by an agent whose only job is to find that failure. Not a job title. Not a persona.

The corpus is unambiguous here: every non-audit run swapped 100% of the pinned roster. Lenses that worked were derived from the subject. A roster is a starting vocabulary, never a default to apply.

Derive from two axes

For the subject in front of you, ask:

1. How could this fail? Not "what is bad" in general — what are the specific mechanisms by which this particular thing goes wrong?
2. Who would notice, and when? Different stakeholders notice different failures at different times. Each distinct noticing-position is a candidate lens.

Then name each lens as the question it asks, phrased as an attempt to break the thing.

The test: could this lens return "nothing found" and have that be informative? A lens hunting a specific failure yields a meaningful negative. "Quality reviewer: nothing found" means nothing.

Orthogonality

Lenses must catch different failures. Two lenses that would find the same finding are one lens and a wasted agent.

Check by asking, for each pair: what does A find that B cannot? If you cannot answer, merge them and add a genuinely different one.

Common redundancy traps:

• Security + reliability overlap heavily on input validation. Split by consequence instead: one hunts exposure, one hunts outage.
• Two seniority levels of the same lens ("senior architect" and "junior developer") — that is one lens with a tone knob.
• A lens per file or per section. Those are subject artifacts, not lenses. The audit corpus explicitly rejected its own five miner lens names as a reusable roster for exactly this reason.

Sizing

Scale to the stakes, not to a magic word (the corpus shows "quick" and "exhaustive" were never typed):

Beyond six lenses, returns fall off and dedup load rises. If you want more coverage, deepen the evidence requirement instead of adding a seventh lens.

Judges are not extra lenses. They are the hard constraints that would kill or approve the decision. Zero judges is correct when nothing is contested — a generation task with checkable ground truth is Mine→Verify only.

The starting vocabulary

Seats that earned their place in the corpus. Adapt or discard freely — this is a prompt for your own derivation, not a roster to apply.

Rejected seats, recorded because negative results are cheap to lose

• A forensics trio of root-cause / timeline-integrity / blast-radius — invented, zero mined runs, rejected 3/3 judges. It reads plausibly, which is precisely why it needed rejecting.
• post-mortem / forecaster as listed vocabulary — not attested in real use.

Writing the miner prompt

Each miner gets: the lens question, the scope, the evidence contract, and the output schema.

You are the CUTOFF lens on the July close package.

Hunt exactly one failure: show me a figure that belongs in a different period.
Look at date basis (effective vs posting vs created), timezone conversion at
period boundaries, and whether any extract was taken while the period was open.

Every finding must cite re-checkable evidence: the file and line, the query, or a
verbatim quote under 200 characters. A finding you cannot point at is not a finding.

Return: {title, claim, evidence, impact, severity, recommendation}
Return an empty array if you find nothing. Do not pad.

Four things that make this work:

1. One failure per miner. A miner asked to find "issues" finds shallow ones across the board.
2. Name where to look, without naming what to conclude.
3. State the evidence contract explicitly — it is what makes Verify possible at all.
4. Permit an empty result. Miners that cannot return nothing will invent something, and invented findings consume verifier budget and dilute the dossier.

When a run comes back wrong

That fourth row deserves suspicion by default. A quorum that finds nothing, or confirms everything, on its first run has usually failed to be adversarial rather than discovered that the work is perfect.

Related skills

• quorum — the orchestrator
• quorum-verification — what the refuter seat does
• quorum-judging — why judges are constraints rather than lenses
• the domain adapters — worked lens sets for common subjects