anti-bias-vector-composition — designing tensioned groups

Arcanum library copy (public). Definition owner of the anti-bias principle, referenced by formulae/dispatch-spec (anti_bias_composition technique, rules R26–R29). The consuming repo also runs an executable gate, check-tension, that enforces this principle's Tests 1–4 before the human confirm; check-tension is not part of this formulae package.

This skill is the design layer: it states why and how to pick the angles of a multi-agent group so bias cancels. The executable gate that checks a sheet against this principle is check-tension (two independent agents, Tests 1–4, before the human confirm) when the consuming repo provides it. Author here; verify there.

The principle (one sentence)

When N agents share a macro goal, their micro vectors (per-agent angle, methodology, prior, source corpus) must be deliberately tensioned into structurally opposed directions — so any bias internal to one is forced into the open by at least one other — not merely made non-overlapping. Diversity of surface presentation is necessary but insufficient: agents on the same corpus, same operator, same artifact produce correlated errors, and the synthesizer reports a biased finding with falsely high confidence. More agents do not break the correlation; structural opposition does.

The clean test. For each pair (a_i, a_j): what bias internal to a_i would not survive a confrontation with a_j's output? If the answer is "none — they'd just produce two compatible findings", the pair is diverse but not tensioned. If it names a specific failure mode, it is.

Where it applies

Group-locally, to a subject group: n ≥ 2, group role investigate (explorers) or evaluate (skeptics). It constrains the per-group anti_bias axis and each agent's angle. It does not apply to synthesize (single writer) or meta-evaluate (single auditor).

• investigate / explorers — tension along source-corpus, methodology, or temporal-prior. Bad: four arXiv keyword searches differing only in vocabulary. Good: arXiv-categorical + physics-journals + dissent-corpus + backward-citation-tree.
• evaluate / skeptics — tension along attack-vector. Bad: three "find problems" agents. Good: precedent-attack + vacuity-attack + definitional-attack.

The four canonical axes (closed vocabulary for per-group anti_bias)

• Methodology — empirical / formal / adversarial / historical / computational
• Source-corpus — arXiv-categorical / journals / dissent-literature / canon / citation-tree
• Attack-vector (skeptics) — precedent / vacuity / definitional / scope / counter-example
• Temporal-prior — modern-only / historical-lineage / mixed-with-decade-bins

A composite is allowed only if explicitly declared. anti_bias_global (dispatch-wide theme, present when ≥2 groups have n≥2) is free-text and never vocabulary-checked.

Tensioned-pairwise is stronger than the partition check

The subagent-strategy chain already checks the angle set is non-overlapping (no two agents cover the same concern) AND covering (together they span the goal). Those are partition conditions — which subset each angle covers. Tensioned-pairwise is a direction condition: a group can pass partition and fail tension (four explorers cleanly partitioning the goal, all using one arXiv-keyword methodology). The gate runs both.