check-tension — the init-time anti-bias gate

Anti-bias is enforced only at initialization (owner decision 2026-06-15): the guarantee is that the agents of a dispatch are tensioned by design, checked here before the human confirm. There is no post-dispatch realization check — the Dissent:-line apparatus was retired (.claude/skills/anti-bias-vector-composition/reference/validator-check.md v0.3.0). This skill is the gate; it owns no field and no type judgment.

When it runs

Between Propose and Confirm in the lifecycle (router §3), for any sheet that has a subject group — a group with n ≥ 2 and role investigate or evaluate. A sheet with no subject group has nothing to tension; the gate passes trivially.

The two agents (independent)

Both read the proposed sheet — the groups, each agent's angle, each group's anti_bias (field meanings owned by register-dispatch) — and judge it against the rubric below. Read-only: neither writes to the source tree.

agent	does
checker («aponta»)	applies Tests 1–4 to every subject group; returns PASS, or per-pair / per-group apontamentos naming the exact failing test + a concrete fix
reviewer («revisa»)	forms its own verdict on the same sheet, then marks agree / disagree on each of the checker's apontamentos — guards against both laxity and over-strictness in the checker

Spawn both in ONE message (parallel). Draw agent_names from telemetry/agents/agent-pool.yaml (auditor / skeptic role_fit); never the same name for both — the reviewer's independence is the whole point.

The rubric — validator-check Tests 1–4 (not duplicated here)

Owner: .claude/skills/anti-bias-vector-composition/reference/validator-check.md Items 1–8. Per subject group:

• Test 1 — axis. anti_bias names one canonical axis (methodology / source-corpus / attack-vector / temporal-prior) or a declared composite. Outside the vocabulary → REJECT.
• Test 2 — clone. No two angles share the same core noun phrase → else REJECT.
• Test 3 — spread. Not all agents share one methodology/corpus (investigate) or one attack-gate (evaluate) → else REJECT.
• Test 4 — evidence. Every unordered pair carries its predicted-disagreement sentence ("a_i runs X, a_j runs Y on the [axis] axis; a bias in a_i would be exposed by a_j") → else REJECT.

Outcome — only "both PASS" goes forward

• Both agents PASS → the sheet proceeds to the human Confirm.
• Either agent reproves → return to the strategist with the consolidated apontamentos; revise and re-run the gate.
• The two disagree (one passes, one reproves; or they contradict on a point) → also return to the strategist. Ambiguity is not good enough — the strategist revises until both pass cleanly. The human never adjudicates the gate.

The human Confirm therefore only ever sees a sheet that passed an independent double-check.

Infrastructure, not a dispatch

The gate's two agents are infrastructure — like the appender:

• they are not registered as a dispatch in the ledger (no row); they are reported in the parent's narration;
• they are not themselves subject to the gate (no infinite regress).

How the router uses it

The router runs this gate at the end of Propose, before presenting the sheet to the human (router §3 / Pointers). It does not change registration, running, or close — those are unchanged.