E-002: What crosses the callosum, and does the fold help it?
Status: DECLINED 2026-09-09. E-001 found the waves lose (ΔR = −0.333, ΔΦ ≈ 0); there is no callosum in the build to measure. Kept as the record of what would have been asked. Decides: whether Fano projection earns a place as the callosum's transfer grammar, whether phase should survive recall, and whether ChiralScale positions should drive dimensionality.
The questions
ADR-0021 gave the corpus callosum four properties (selective gate, bandwidth limit, asymmetric rates, balance-seeking) and one projection grammar (Fano folds). The four properties were built. The grammar was measured only as a ranking prior, where it never improved a rank and was removed (ADR-0047 erratum, 2026-08-29). It was never isolated as what it was designed to be: the projection between hemispheres.
Three sub-questions, each its own arm against the E-001 winner as baseline:
| arm | change | measure |
|---|---|---|
| F | callosal transfer projects through a Fano line (three groups rotate, phase flips) instead of a plain copy | recall@10 on zero-overlap-33 after 30 dreams; callosal throughput; cross-hemisphere Kuramoto order |
| P | recall returns cos(Δφ) alongside the scalar and the merge keeps phase (ADR-0053) instead of collapsing to a real | rank changes on the 83 probes; count of probes where phase opposition suppressed a correct answer |
| S | ChiralScale::positions sets the dimension budget per hemisphere (ADR-0021 §Dimension Group Assignment) instead of a fixed D | recall@10; memory footprint; whether bilateral scale jumps correlate with consolidation events |
Ten runs per arm, same corpus and probes as E-001, same fresh-process measurement.
Decision rules
- Arm F is kept only if recall does not fall (interval includes zero or better) and cross-hemisphere order rises with the interval excluding zero. A projection that costs recall to buy coherence is recorded and declined.
- Arm P is kept only if it changes at least five ranks across the 83 probes and the net change is positive. Fewer than five means the phase is real but inert at this corpus size, and the scalar stays.
- Arm S is kept only if recall does not fall and footprint falls with the interval excluding zero. Otherwise the scale stays a per-facet observable and
Dstays fixed.
Known trap
The chiral merge once masked the stronger hemisphere's score and collapsed small stores at k ≥ 4 (kannaka-memory #716). Arm F's merge keeps the stronger score and is tested with the recall_score_is_top_k_invariant fixture before any run is scored.