dissent-board

Dissent Board

A five-director adversarial board that tears your proposal apart before you ship it. Forces real dissent, evidence-backed opinions, and a mandatory skeptic role — built so the board surfaces flaws, not so it manufactures consensus.

🇮🇱 קרא ב-עברית — README.he.md

---

Who This Is For

You are about to ship a decision that you will live with. An architecture choice, a schema migration, a production deploy, an ADR you're about to commit. You want a real critique — not LLM cheerleading that approves whatever you wrote.

The typical user is a senior engineer, tech lead, or staff engineer working with Claude Code (or any agent that supports the Agent Skills standard) who has a non-trivial plan and wants it stress-tested before code is written, before a migration runs, before a deploy ships.

If your decision is reversible in five minutes — this skill is not for you. If undoing your decision in six months will cost a week of work — this is exactly for you.

---

What You'll Get From a Session

A single Markdown resolution that tells you:

1. Action — what to do next (one sentence at the top).
2. Verdict — APPROVED, APPROVED WITH CONDITIONS, APPROVED WITH REVIEW, REJECTED, or ESCALATED.
3. Blocking concerns — what must change before you proceed, each with a file/line citation.
4. Conditions — operative checks that close each concern.
5. Dissent — full reasoning from any director voting against majority.
6. Auditor findings — calibration, evidence, veto-validity, and hard-rule checks.
7. Skeptic report — three concrete risks the assigned skeptic found in your proposal.

A real session ends with one of three feelings: "I'm shipping with confidence" (APPROVED), "I almost shipped something broken — glad I checked" (REJECTED), or "This needs human attention, not auto-resolution" (ESCALATED).

---

The Problem This Solves

When you ask an LLM to "act as a board of experts and review this proposal," you get a predictable failure pattern:

• Convergence — all five "directors" approve, because parallel sub-agents share training and bias
• Sycophancy — the board tells you what you want to hear, because the proposal frames the question
• Theater — votes look serious, but scores are arbitrary and not tied to cited findings
• Domain crossover — every director comments on everything; no one goes deep on anything
• Tie collapse — when the board disagrees, an automatic tiebreaker (usually "the architect wins") quietly defeats the disagreement

Dissent Board exists to eliminate all of them.

---

What It Enforces

Pressure	Counterweight
Convergence	Phase 2 cross-examination — each director must find a concrete weakness in another director's reasoning, or the round is void
Sycophancy	Every finding requires file:line or quoted-passage evidence; unsupported claims are dropped by the Auditor
Theater	A separate Auditor pass validates that scores are calibrated against cited concerns
Domain crossover	Each director has an explicit Out-of-Bounds section; out-of-domain veto is flagged invalid
Tie collapse	No automatic tiebreaker — ties escalate to the human with a polarization report
Skeptic absence	One director is assigned the Skeptic role and must surface ≥ 3 concrete risks, or the session is invalid
Hard-rule bypass	Director profiles declare hard rules (e.g., CPO cannot APPROVE without a named user role); Auditor enforces them

---

Before & After

Without Dissent Board	With Dissent Board
"All five experts agree — looks great."	One director finds a buried flaw the proposal hid; vote becomes 4-1 with veto
Generic concerns ("scaling risk", "needs more testing")	Concerns cite src/api/items.ts:42 and quote the line that fails
Score 8/10 with three serious concerns	Auditor flags calibration failure; score must be re-defended
"Should we ship it?" → "Yes"	"Should we ship it?" → "Not until these three blocking concerns are closed"
Tie quietly resolved by the loudest voice	Tie escalated to a human with full polarization report
Security flagged as "watch out for vulnerabilities"	Security vetoes with named CWE, input path, and remediation file
LLM hallucinates a commit hash that doesn't exist	Auditor runs git log, drops the unsupported claim, documents the drop

---

The Protocol

Four phases. No phase is optional in Full mode.