# EXP-10: multi-mind per-contributor worth attribution in a mixed genuine/fake collaboration (program M2, additive)

- Date: 2026-10-08 (iter5 run, iter8 corrected-eval rerun) | Cadence version: 0.78.0 / `750c8ef` (both runs) | Program iter: iter5, iter8
- Script: `exp/EXP-10_vs_multimind_attribution.py` (current, #164 reset fix) and `.py.bak` (pre-fix eval) | Evidence: `exp/EXP-10.out` (iter5, pre-fix), `exp/EXP-10-fixed.out` (iter8, current)

## Summary
- Question: in one collaboration where some minds are genuine and some fake, with identical interior structure, does measured transfer localize worth to the specific genuine minds, and does aggregate transfer scale with the genuine fraction?
- Setup: K=8 contexts split into 4 disjoint blocks (minds) of 2 contexts, 2 actions, `modules=(16, 8)`, 1200-step witness, 12 seeds. A student brain ingests the composite map: genuine blocks teach T, fake blocks teach the flipped T on their contexts. Per-mind run: blocks 0,1 genuine, blocks 2,3 fake; transfer measured per block (2 contexts each). Proportional run: first n of 4 blocks genuine, n = 0..4, aggregate transfer over all 8 contexts.
- Prediction: written in the docstring before running: "genuine minds' blocks show positive transfer, fake minds' blocks ~zero/negative, WITHIN THE SAME collaboration".
- Result: current (iter8, fixed eval, 12 seeds): per block +0.292 (SEM 0.092), +0.458 (0.092) genuine; -0.500 (0.118), -0.458 (0.092) fake; genuine mean +0.375, fake mean -0.479, attribution gap +0.854. Proportional curve f = 0 / 0.25 / 0.5 / 0.75 / 1.0: -0.542 / -0.302 / -0.052 / +0.198 / +0.448 (SEM 0.037 to 0.040). Superseded pre-fix (iter5): genuine +0.250, fake -0.292, gap +0.542, block 0 +0.125 (SEM 0.120, within about 1 SEM of zero); curve -0.292 / -0.115 / -0.021 / +0.052 / +0.208.
- Verdict: CROSSED (M2 additive half), on the corrected-eval numbers; the iter5 numbers are SUPERSEDED by the iter8 rerun (verdict held and got stronger; block 0 is no longer within 1 SEM of zero). Scope: additive per-node attribution only. The super-additive synergy half of M2 was not tested here; it was later addressed separately (EXP-18 data-channel synergy, EXP-19 learned composition); XOR-style synergy is still open per the program log.

## Reflection
- What it means: within a single collaboration, transfer assigns positive credit to the minds whose blocks carry real T-information and negative credit to the ones that teach the opposite, and aggregate credit is linear in the number of genuine minds. That is the per-contributor property a Shapley-style split needs, shown in the simplest additive case.
- Process: (1) Eval bug #164 (no reset between probes) affected iter5; the fixed rerun enlarged the effects by between about 1.5x and 2x (gap +0.542 to +0.854; f=1.0 +0.208 to +0.448). (2) The fake minds are anti-correlated teachers (flipped T), not zero-information fakes, so their negative transfer is close to guaranteed; the test does not include a fake that teaches nothing or something orthogonal. (3) Per-block accuracy is over 2 contexts, so it is quantized to steps of 0.5 and per-block SEMs are 0.09 to 0.12. (4) Linearity of the proportional curve is largely by construction: blocks are disjoint and the bandit is additive across contexts, and the curve re-expresses EXP-09's gamma through block structure; the new information is per-mind localization. After the fix the curve also looks to be at ceiling: consecutive points rise by about one block (2 of 8 contexts) each, which is what a student that learns the composite map perfectly would show, so the slope measures block size more than learning. (5) The pre-fix log called f=1.0 (+0.208) a consistency match with EXP-08 (+0.219), but the runs differ in witness length (1200 vs 1600), seed count (12 vs 16) and task-rule seeds, so that check was loose. (7) Stale records: the EXP-10 entry in EXPERIMENT-RESULTS.md, the M2 status line in VS-TRANSFER-PROGRAM.md SOLVED-CRITERIA and the EXP-04 row of EXPERIMENT-QUEUE.md still carry the pre-fix +0.542 gap. (6) "Interior-identical" holds only at the level the script models (node count, acyclic, disjoint consistent sub-maps), not the full `wash_sim.rs` feature set.
- Lesson: an additive task can only demonstrate additive attribution; claims about collaborative worth need a task where worth exists only in the combination, and fakes should include a zero-information arm, not just an opposite-teaching arm.
- Next: moved the program to M2-synergy (iter6 design crux, EXP-11 substrate smoke, EXP-12 XOR probes, then EXP-18 and EXP-19 synergy results).