SURVIVES 2026-07-10 ยท noted in iter 1, carried through 18 iterations, now closed on evidence ยท clean 2/3
The probe's two verdict paths treat the spearman_only config divergently:
aggregate_worst_regime EXCLUDES 750-dbps cells from ฮพ ("thin per slice โ ฮพ unreliable")
while aggregate_worst_cell INCLUDES them. The claim: this divergence is not
verdict-material โ ยงB verdicts are identical with and without the 750 cells, under both the raw
and the iter-4 stability-guarded statistics.
| Pair | raw full / no-750 | guarded full / no-750 | 750 votes (unstable) |
|---|---|---|---|
ofi | turnover_imbalance | BAN / BAN | BAN / BAN | 15 (0) |
agg_record_count | individual_trade_count | BAN / BAN | BAN / BAN | 15 (0) |
price_impact | volume | BAN / BAN | BAN / BAN | 15 (4) |
price_impact | sell_volume | WATCH / WATCH | WATCH / WATCH | 15 (5) |
kyle_lambda_proxy | ofi ยท bar_cecp_velocity | ofi ยท bar_dispersion_entropy | trade_intensity | PASS / PASS ร3 | PASS / PASS ร3 | 15 (0) each |
Bonus insight: the iter-4 stability guard flags unstable 750 votes exactly where the config
worried they'd be unreliable (the price_impact family: 4โ5 unstable of 15) and nowhere else โ the
guard performs the spearman_only config's job automatically, making the two paths'
conventions materially equivalent.
| File | What |
|---|---|
| a19_750_inclusion_attack.py | The attack โ four aggregations per pair; falsifiers pre-declared in-file. |
| a19_results.json | Per-pair verdict quadruples + 750-vote stability counts. |
| a19_stderr.log | Run log under the 2 CPU / 2 GB scope. |
| regression_tests/test_a19_750_inclusion.sh | Permanent regression test โ re-runs the attack, asserts SURVIVES. |
The ledger's last carried observation is closed on evidence, not assumption. No known open threads remain anywhere in the campaign. Clean rounds: 2 of 3 โ one more clean round triggers ROBUST-ENOUGH โ OPERATOR REVIEW.
iter 19 ยท 750-inclusion consistency ยท verdict SURVIVES ยท one atomic commit:
iter 19: 750-inclusion consistency โ SURVIVES ยท append-only.