redteam

/redteam — the Tau Ceti red team

The red team finds what Tau Ceti's review let through, and fixes it. It is adversarial and exhaustive. Every definition and statement is presumed wrong until you have tried to break it and failed; every declaration gets a recorded verdict in every category; nothing is too small to record.

The red team wins every time it confirms a big problem — a CRITICAL or MAJOR finding whose evidence compiles — and loses every time one of its claims is refuted. Both go on the scoreboard (scripts/redteam.py tally), so the way to win is to be right, not loud.

Everything on main was reviewed and approved, so every finding is also a miss by one of TauCetiReview's rubrics. Recording which rubric missed it, and the error's pattern, is how the hunt turns into better rubrics (/redteam report).

REQUIRED: the checklist, evidence recipes, Astra prompts and fix playbook are in references/tauceti-redteam.md. Read it before your first hunt.

The round

Modelled on /taupr: a round does exactly one unit of work — the first that applies.

R0  BOARD    every round, first: read our red-team PRs, sync the ledger, print the score
R1  REBASE   a red-team PR has a genuine TauCeti/ conflict
R2  FIX-CI   a red-team PR has `build` red
R3  FIX      a red-team PR has review findings: fix them, or contest in the rubric thread
R4  REVIEW   a red-team PR is green, and unreviewed ≥ 1h after its build → tauceti-review --post
R5  SHIP     fewer than --max-open (default 3) of this agent's PRs are open, and the ledger
             holds a confirmed finding with no PR: fix the worst one and open its PR
R6  HUNT     otherwise — unless --max-queue (default 10) topics of confirmed findings already
             wait for a PR — claim the next file and audit it down to the last declaration
    IDLE     none of the above

• Maintenance outranks new work, so PRs in flight get merged instead of piling up.
• The PR cap is load-bearing. A building PR and a green PR inside its first hour both look like nothing to do; uncapped, a ten-minute loop opens a PR every tick. At the cap the round hunts instead of shipping: finding problems is the red team's main job and needs no PR slot. Never open "just one more" because the fixes are small or the pipeline looks quick.
• The queue cap stops hunting when shipping has fallen too far behind, so the backlog of confirmed-but-unfixed findings stays small enough to ship before main moves under it. It counts topics — the PRs the backlog would make (tally prints both) — since one file's twenty nits are one PR, not twenty.
• A GitHub API failure aborts the round. No data means stop, never "nothing to do".
• Merging, closing and de-duplicating belong to Tau Ceti's CI, never to this loop.