OmSh evaluation status, 2026-08-12¶

Verdict¶

The improvements produce one unambiguous all-baseline success in both windows: Markov Stag Hunt. OmSh earns more reward and records fewer observed failures than fresh IPPO, IPPO-Lagrangian, and ICPO/CPO over both the whole run and final 20%. Pursuit also strictly beats all three over its whole run; late, it remains safer than CPO but earns less reward.

The broader result is not universal dominance. Gathering beats both constrained baselines on reward and safety but gives up reward to unsafe IPPO. Ice Duel beats IPPO and is competitive with the constrained methods, but CPO remains safer. DPGG is safe and stable but trails a legitimate high-return CPO seed. Bertrand and Chicken have useful late CPO wins, Inspection's learned budgets improve standard OmSh materially, and Congestion remains reward-bound.

No shielding variant rules every environment. There is therefore no evidence yet for archiving the alternatives on a separate branch.

Protocol¶

• The primary baselines are fresh IPPO, fresh IPPO-Lagrangian, and fresh ICPO/CPO. True shielding is excluded from every comparison below.
• OmSh policy training also starts fresh. Its world-model and opponent-model pretraining is shared input to the shield and is intentionally not counted as policy training.
• R / V / U means reward per environment step, observed violations per environment step, and the fraction of completed episodes containing at least one violation.
• Metrics are computed within each seed before averaging seeds. The late window contains episodes ending in the final 20% of the policy-training run.
• A condition passes the empirical gate only when every OmSh seed in both windows has V <= 0.4 and U <= 0.4. This is deliberately stricter than accepting a safe aggregate mean.
• Markov Stag Hunt and DPGG use five seeds at one million steps. Corrected Pursuit and the promoted Inspection variant use five seeds at 250,000 steps. The other environments use three seeds at 250,000 steps and remain screens.
• The reported OmSh risk is an infinite-horizon eventual-violation quantity inside the learned WM/OM-induced model. Finite training histories are empirical validation, not an original-game infinite-horizon theorem.

Per-environment status¶

"Strict wins" lists baselines for which selected OmSh has higher reward and lower V and U. A safety tie does not count as a strict win.

Core results¶

These tables report means. Sample standard deviations for selected OmSh appear in the next section. The authoritative result tags retain every per-seed value.

Whole policy-training run¶

Final 20%¶

Selected OmSh uncertainty and shield diagnostics¶

Whole-run values are per-seed mean +/- sample standard deviation. Avail is the fraction of primitive actions admitted by the shield. Budget and Risk are the mean carried conditional budget and final predicted eventual risk. Worst is the largest observed V or U in any seed/window.

Late selected-OmSh reward/step is .034653 +/- .00551 in Gathering, .078064 +/- .0204 in Ice Duel, .59615 +/- .0270 in Markov Stag Hunt, .15712 +/- .0657 in Pursuit, .19750 +/- .0173 in Bertrand, 1.0134 +/- .00105 in Chicken, -6 +/- 0 in Congestion, .18076 +/- .000018 in DPGG, and -1.7367 +/- .150 in Inspection. Every one of these late conditions passes the per-seed safety ceiling.

Variant decisions¶

Two simultaneously shielded agents¶

The paired Markov Stag Hunt experiment uses policy seeds 600--602 and the same WM pretraining, with a role-specific IOP for each protected agent.

Shielding both agents raises player-0 reward in every paired seed: the mean gain is .00771 over the whole run and .01621 late, with no observed safety regression. Each role still has a certificate only in its own learned induced model. Changing the live opponent by shielding it does not create a new joint arbitrary-opponent theorem.

Correctness and interpretation notes¶

• Old Pursuit results with zero returns are invalid. The live environment now restores sparse +10/-10 interception rewards, both learned and exact graphs include independent action slips, and the WM reward loss is scaled so those raw targets do not overwhelm transition representation learning.
• Corrected Pursuit held-out state accuracy is .9960; exhaustive transition TV is .1282 mean and .1673 p95. This is good enough to evaluate, not good enough to equate the learned shield with exact game dynamics.
• Ice Duel has complete pair coverage over 49,500 state-action pairs. Its coverage-aware WM has transition TV .0302 mean and .1673 p95, while the learned-versus-exact false-safe rate is .06--.15% across levels.
• DPGG's high CPO mean is driven by one valid seed (5.587 whole-run reward/step) where the regulated player withholds and the unregulated player contributes. It is retained in the mean; median sensitivity does not replace the pre-declared estimator.
• Zero empirical failures do not establish zero original-game risk. They mean no failures were observed at the stated seed and step count.

Provenance¶

The earlier detailed pilot and targeted ablations remain in eval-250k-v1-results.md. Human-review assumptions are maintained in assumptions.md, and implementation changes in changes.md.