Date: 2026-08-29 Repository issue: #79
The project owner asked coding agents to work professionally in parallel on ten unclaimed issues, solve them, push the work, and integrate only after repository review and merge safeguards are satisfied.
Issue #79 was unassigned and had no open PR or matching remote branch. It was claimed publicly before implementation. This work implements E025; the issue’s initial experiment label collided with the canonical ACO experiment. This work does not self-merge; evidence needs independent review at the exact PR head SHA.
Required work includes deterministic calibration/held-out separation, learned reliability and dependence baselines, matched review cost, classification and calibration metrics, correlation-estimation error, effective evidence, history sensitivity, distribution shift, harmful regimes, uncertainty-bearing JSON, documentation, and CI coverage.
[7,1,1,1,1] geometry with heterogeneous group accuracy.(votes, frozen_model) and has no truth input.N_eff formula as an explicitly unsafe baseline and add a
calibration-observed alternative.A later review questioned whether the effective-size helper compared a
majority-correct probability with group error. Direct evaluation showed the
summation is over k < majority correct votes, so it already computes majority
error: for n=1, p=.75 it is .25, and for n=3 it is .15625. Inverting
the comparison would have introduced the reported bug. The existing
independent_error error-to-error comparison was retained, its meaning was
documented, and the two known cases were added as regression tests.
A separate exact-diff review then found that empirical_neff selected only the
nearest odd integer, whereas E015 defines effective size by continuous linear
interpolation between bracketing odd sizes. E025 now reuses E015’s canonical
effective_n, fails closed to one when a short-history reference accuracy is
not above chance, and tests an interpolated value between n=1 and n=3. The
committed artifact is now validated against the strengthened schema and
reproduced in full by the focused test suite.
In the 20-seed reference run, combined learning improves on naive majority in
all six stable cells and hurts in six shift cells. At 200 history claims under
stable item difficulty, error falls from .2312 to .0597; under reliability
reversal it rises from .1071 to .2182.
When dependence emerges after calibration, combined learning slightly improves decision error but worsens Brier score and produces roughly nineteen times as many high-confidence errors. Better classification alone does not establish safer confidence.
The CLI, schema, committed artifact, tests, and experiment document make the result reproducible with local/public CI and no project spend. The claim is explicitly synthetic and negative regimes remain discoverable. Follow-up work must determine how to gate learned weights under domain shift, represent shared blind spots beyond pairwise correlation, and test the method on bounded software tasks with independent delayed ground truth.