Failure modes · six_layer_coherence

2 of 3 modes published

• High

  judge_layer_skipped_biases_unresolved

  When the LLM-judge layer (S6) is disabled via skip_llm_judge=True, the sixth vote defaults to UNRESOLVED. The 4/6 supermajority becomes harder to reach, so the aggregate verdict drifts toward UNRESOLVED even on pairs where the mechanical layers agree.

  Trigger

  The conclusion was produced in offline batch mode, on a tight cost budget, or with skip_llm_judge=True. Watch for clusters of UNRESOLVED verdicts on pairs the firm previously considered decidable.

  Example

  In an offline batch run we evaluated 800 contradictory pairs with the judge disabled to save cost. 41% came back UNRESOLVED versus 8% with the judge on. The pairs were not harder; the threshold was simply unreachable.

  Mitigation

  Re-run the contested pair with the judge enabled, or treat UNRESOLVED outputs from skip_llm_judge runs as "needs human" rather than "indeterminate".

• Medium

  argumentation_layer_starves_without_neighbors

  The argumentation layer (S2) needs neighbour claims and precomputed pairwise contradiction scores to construct an acceptable extension. When those are absent — early in a project, after a corpus reset, or for an isolated claim — S2 emits a weak or null signal and effectively abstains.

  Trigger

  The conclusion is the first or one of very few claims in its cluster, the contradiction score table has not been backfilled, or the pair sits far from any neighbour in embedding space.

  Example

  A new founder uploaded a single transcript that produced four claims with no prior corpus context. S2 returned UNRESOLVED on every pair, leaving the verdict to rely on five layers instead of six.

  Mitigation

  Check the argumentation layer's neighbour count before trusting its abstention; backfill contradiction scores for the cluster before treating S2 outputs as evidence.