check-tension

name

check-tension

description

Init-time anti-bias gate. Before the human confirm, two independent agents verify a proposed dispatch sheet is genuinely tensioned against the rubric this skill owns — the four anti-bias tests plus a cross-group coherence check; the sheet reaches the human only if BOTH pass. Any reproval or disagreement returns it to the strategist. Gate infrastructure: not itself a dispatch, not subject to its own gate.

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

Anti-bias is enforced only at initialization: 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. This skill is the gate and owns the rubric below; it owns no dispatch field and no type judgment.

When it runs

Between Propose and Confirm in the dispatch lifecycle — the step the domainspec-subagents-strategy skill defines — 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 the register-dispatch skill) — and judge it against the rubric below. Read-only: neither writes to the source tree.

agent	does
checker («aponta»)	applies tests 1–5 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) — two distinct, independent agents; never the same identity for both, since the reviewer's independence is the whole point. As gate infrastructure they carry no agent-pool identity and write nothing to it.

Each apontamento is structured, not prose, so the strategist can act on it mechanically:

• target — the failing subject group (and the unordered pair, for tests 2 / 4).
• test — which test fired (1 / 2 / 3 / 4 / 5).
• fix — one concrete revision (e.g. "re-axis a_j from precedent-attack to definitional-attack").

The checker returns PASS or a list of apontamentos. The reviewer returns its own PASS/list plus an agree/disagree flag on each of the checker's apontamentos.

The rubric — the four anti-bias tests + a cross-group check

Tests 1–4 are the four anti-bias tests every subject group must survive; test 5 is the fifth, cross-group coherence check applied once per sheet.

The axis vocabulary (closed — test 1 checks anti_bias against it):

• methodology — empirical / formal / adversarial / historical / computational
• source-corpus — e.g. arXiv-categorical / physics-journals / dissent-literature / textbook-canon / backward-citation-tree
• attack-vector (skeptics only) — precedent / vacuity / definitional / scope / counter-example
• temporal-prior — modern-only / historical-lineage / mixed-with-decade-bins

Per subject group — REJECT if any fires:

• Test 1 — axis. anti_bias names one canonical axis above, or an explicitly declared composite of them. Outside the vocabulary → REJECT.
• Test 2 — clone. No two angles share the same core noun phrase — tokenize, drop stopwords, and the pair must yield ≥ 2 distinct primary verbs or nouns → else REJECT.
• Test 3 — spread. investigate: ≥ 2 distinct axes across the angles; all sharing one methodology or one corpus → REJECT. evaluate: no two skeptics share an attack-vector (precedent + vacuity + definitional = three; three "find problems" = one) → 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. Writing this per pair is what forces real tension instead of asserting it.

Once per sheet (not per group):

• Test 5 — global coherence. When ≥ 2 groups have n ≥ 2, each subject group's anti_bias must be a plausible specialization of the sheet's anti_bias_global theme, not an unrelated axis → else REJECT. (Uncoordinated axes across groups drift.)

A group passing tests 1–4, and a sheet passing test 5, PASSES — no residual judgment beyond the evidence sentences.

This gate checks only the tension half of the anti-bias guarantee. The partition half — angles non-overlapping AND covering the goal — is checked earlier by the domainspec-subagents-strategy skill. A sheet must pass both; partition first, then this gate.

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.