SSKIM LAB — Study-1 public aggregate v1
Published: 2026-11-03. Experiment: 2026-10-13.

This is a five-column projection of the closed fixed-model control experiment's
POLICY_SUMMARY.csv. It contains four aggregate policy rows, not episode/task logs.
source SHA256: acf96b4a7dc35191b690238f285d90a5a1afd389c5b568a786b8ed6506adf655

Column mapping:
policy -> policy
completed -> episodes
successes -> normalized_successes
strict_terminal_success -> strict_successes
recovery_api_calls -> recovery_calls

Each policy completed 864 episodes. Total = 3456 episodes, 288 common blocks,
24 task instances from six families, 12 source models, three fault types.
Do not treat 864 states as 864 independent tasks. Do not infer paired/cluster
significance or equivalence from these marginals. M4 superiority was not shown.
M5 vs M4: 51 fewer recovery calls = 48 local bypasses + 3 independent retry
differences. Observed local false accepts were zero among 48 local accepts;
124 correct candidates were rejected. Zero observed errors is not a safety proof.
Recovery calls are neither total collection attempts nor an invoice amount.

Recompute, offline, with Python 3's standard library:
python recompute_policy_summary.py study1-policy-summary.csv

The JSON file supplies the same public rows, definitions, scope and limitations.
The web tool uses this snapshot by default; custom CSV stays in browser memory.
No account, upload or API payment is needed. Do not paste secrets into shared PCs.

These public aggregates and the newly authored verification script may be reused
with attribution to SSKIM LAB and this version. This grants no rights to private
logs, external models, unpublished manuscripts or third-party material.
Contact/corrections: sum3231@gmail.com