Red-Team Protocol

Adopted 2026-06-10 as a hard doctrine change (see changelog-2026-06-10-hard-changes). A buy verdict cannot be issued without a completed, answered red-team artifact.

The problem this fixes: the kit currently self-referees. The same agent writes the bull thesis, the "steelman of consensus," the pre-mortem, and the calibration tracker. Steel-manning your own thesis is structurally weak — you cannot fully argue against a conclusion you have already reached, because the framing, the inputs, and the emphasis are already chosen. The 04-intellectual-virtues file calls for skepticism of one's own narrative; this protocol gives that skepticism an independent operator instead of asking the author to police themselves.

Rule 1 — The red-team is a separate, blind pass

For any name reaching a buy or pass-with-trigger verdict, a red-team brief is produced by a fresh agent that has not seen the bull thesis. It is given only:

• The ticker and the entity-confirmation block.
• The raw sourced figures (the data-integrity input table), not the author's interpretation.
• The mandate: build the strongest case that this is a value trap, a permanent-loss risk, or simply not cheap. Assume the bull author missed something. Find it.

The blindness is the point. A red-team that reads the bull case first inherits its framing and produces agreement dressed as scrutiny.

Rule 2 — What the red-team brief must contain

<TICKER>-red-team.md, ~800–1,500 words:

1. The bear thesis in one paragraph — the single most convincing reason to lose money here.
2. The permanent-loss path — the specific sequence (balance-sheet, fraud, obsolescence, moat collapse, dilution) that takes equity toward zero, with a probability estimate. See permanent-capital-loss.
3. The value-trap path — why "cheap" could stay cheap or get cheaper: deteriorating economics the screen doesn't show, accounting that flatters earnings (run the quality-of-earnings checks independently), a melting asset base.
4. Attack on the load-bearing input — independently re-derive or challenge the single input that carries the most value (usually the growth runway or the normalized margin). State where the bull number is most likely wrong and in which direction.
5. Disconfirming evidence the bull case would avoid — short reports, activist letters, customer/channel signals, insider selling, comparable-company failures.
6. A "what breaks this" probability — the red-teamer's estimate that the thesis is materially wrong within 24 months, with the one observation that would most cheaply confirm the bear case.

Rule 3 — The verdict cannot issue until the red-team is answered

The bull author (or thesis-builder) must respond to the red-team in the thesis §9 Risk and pre-mortem, point by point:

• Each red-team point is either rebutted with evidence, absorbed as a kill criterion, or conceded (which may downgrade the verdict or the confidence tag).
• A buy verdict with unanswered red-team points is invalid.
• If the red-team's "what breaks this" probability is ≥ 35%, the name cannot be Core 1/Core 2 regardless of margin of safety — it routes to Mid/Probe or pass-with-trigger pending resolution. <span class="tier-cal" title="35% threshold, 2026-06-10; revisable against backtest.">AS-cal</span>

Rule 4 — The red-team is preserved and scored

<TICKER>-red-team.md joins the canonical bundle. At each calibration checkpoint, the tracker records whether the red-team or the bull author was closer to what actually happened. Over time this scores whether the red-team is adding signal or rubber-stamping — and whether the bull author systematically under- or over-weights specific risk types.

Operational note (how the blind pass is run here)

The red-team pass is run as a separate sub-agent invocation with a clean context, fed only the sourced-figures packet and the mandate above. It does not receive thesis.md. In a session without sub-agent capacity, the red-team must at minimum be written in a separate pass before the author re-reads their own thesis, explicitly adopting the bear seat — but the blind sub-agent is the standard and the in-session version is the fallback, flagged as such.

Bundle + frontmatter additions

• New canonical file <TICKER>-red-team.md — required for buy and pass-with-trigger, optional for pass/avoid.
• Frontmatter: red_team: complete | pending, red_team_break_probability: <pct>.
• The thesis-builder may not set a buy verdict while red_team: pending.

Linked

• 04-intellectual-virtues — the skepticism-of-own-narrative principle this operationalizes
• pre-mortem-and-inversion — the in-house inversion technique the red-team extends with independence
• permanent-capital-loss — the path the red-team must trace
• quality-of-earnings — run independently by the red-team
• short-and-pessimism — the bear-side toolkit the red-team draws on
• thesis-bundle-standard — where the red-team file sits in the bundle
• changelog-2026-06-10-hard-changes — adoption record