# EXP-35b: pedigree share as a control dial for discovery, at K=96

- Date: 2026-10-09 to 2026-10-10 | Cadence version: 0.80.0 (`19bee92`) for iter27/iter28; 0.78.0 for the void iter25 run | Program iter: iter25 (VOID), iter27 (NULL), iter28 (SUPPORTIVE)
- Script: `exp/EXP-35b_pedigree_controller.py` (pilot: `exp/EXP-35b_pilot.py`) | Evidence: `exp/EXP-35b-item-2026-10-10.out`, `exp/EXP-35b-K96-R30-item.json/.jsonl`, `exp/EXP-35b-pilot-item-P*.jsonl`, `exp/EXP-35b-full-2026-10-10.out`, `exp/EXP-35b-K96-R30-{orig,fixed}.json/.jsonl`, `exp/EXP-35b-K96-R30-iter25.jsonl`

## Summary
- Question: if a discovery's ORIGIN earns a share ALPHA of the pay whenever others teach it onward, do brains choose to discover more, and does the network learn more?
- Setup: 8 brains, K=96 bandit contexts, each brain starts knowing 4 (64 known to nobody). Each round a brain either TEACHES (its neediest peer, own or relayed items, pay = student's measured gain) or DISCOVERS (learns one unknown context from the world, D=40 graded trials). Move choice follows each brain's running estimate of what each move pays (epsilon 0.15). Notes + drift-check + rehearse (EXP-33 fix) on. 30 rounds. Arms: ALPHA 0, 0.9, and a PI controller. Three credit rules across three runs: orig (iter25), fixed (iter27), item (iter28: Discover is credited ALL pay on items the brain found, its own teaching of them plus royalty; each move's value = decayed pay / decayed move count, PRIOR = one pseudo-move).
- Prediction (iter28, written in the docstring before the full run): "per-move Discover pay ~2x (>3 SEM), Discover share in the last 10 rounds higher by >=.08, network-new discoveries higher by 1-4 (may not clear 2 SEM), final accuracy within +-.02 (no claim). PI should hold alpha well above 0."
- Result (iter28, CREDIT=item, PRIOR 0.01, 12 seeds, mean +- SEM), ALPHA 0 vs 0.9: Discover pay per move .0036+-.0005 vs .0071+-.0004; Teach pay per move .0091+-.0003 vs .0079+-.0003; Discover share (all rounds) .41+-.01 vs .46+-.02; last 10 rounds .51+-.01 vs .63+-.03; network-new discoveries 54.0+-1.0 vs 57.8+-0.8; final accuracy .790+-.006 vs .811+-.009; mean-over-time .688+-.004 vs .696+-.004. PI: final .798+-.007, discovery .41, share path .57 -> .13 (ends .14+-.04).
  Pilot (3 seeds, same conditions) at PRIOR .01 and .002: Discover pay per move .0034 -> .0071 and .0022 -> .0064; Discover share .39 -> .44 and .35 -> .42.
  Earlier runs: iter25 (orig, v0.78) VOID, royalty credited to the wrong move. iter27 (orig and fixed, 6 seeds) NULL: under fixed, discovery royalty was ~1/40 of direct pay and the Discover estimate never left PRIOR (.0100 at ALPHA 0, .0099 at 0.9).
- Verdict: SUPPORTIVE (iter28). With correct credit the share roughly doubles what discovering pays, brains discover more (+.12 late, ~4 SEM) and the network finds more new items (+3.8, ~3 SEM). Accuracy gain small (+.021, ~2 SEM, just outside the predicted +-.02; no claim). PI part of the prediction WRONG. iter27 NULL is superseded as a credit-design failure; iter25 stays VOID.

## Reflection
- What it means: the pedigree share CAN steer behaviour toward discovery in this small world, but only if Discover is credited with everything a discovery later earns, including the discoverer's own teaching of it. Royalty from other brains' relays alone is too small to see. The share is a transfer, not free money: Teach pay per move fell (.0091 -> .0079) and relaying fell (.57 -> .49) as discovery rose. Whether that trade raises total knowledge is only weakly supported (accuracy +.021, ~2 SEM).
- Process: three runs to get one valid answer. (1) iter25 credited royalty to whatever move was made that round, so the share never reached the choice (bug). (2) iter27 fixed that but booked a discoverer's own teaching of its find as Teach pay and scored Discover only when royalty arrived, so the Discover estimate stayed at its PRIOR: a design bug, visible in pilot-scale numbers. (3) A 2-round smoke test was mistaken for a pilot before iter27. iter28 ran a real pilot that printed every link of the chain (share -> per-move pay -> estimate -> choice -> outcome) at full length and two PRIOR values before the full run. PRIOR still shifts baseline accuracy (.796 at .01 vs .749 at .002 in the pilot), not studied further. The PI controller's set point (fraction of contexts nobody knows) falls by construction as the network learns, so the controller lowers the share late: a controller design flaw, not evidence about the share.
- Lesson: a smoke test is not a pilot. Name the mechanism chain and print each link per arm before the full run; an estimate stuck at its prior is a design bug, not a result. And credit a move with everything it causes, not just the most visible payment.
- Next: a PI controller with a set point that does not shrink by construction (e.g. target a discovery rate, or the marginal value of discovery); test the share at a second world size; design notes may now cite "in a small simulation, a pedigree share raised discovery" with these caveats.