S-AI-Decision: Decision-Making as Certified Commitment

Abstract Background. Decision-making cannot be reduced to world representation, reasoning, planning, or valid action execution. In partially observable environments, an agent must distinguish feasible plans from preferred, admissible, and authorized commitments while controlling uncertainty, risk, and the cost of further analysis. Research problem. The central challenge is to determine when available evidence and prospective consequences justify action, clarification, deferral, or refusal. The distinctive empirical question is whether a Decision Commitment Gate improves outcomes beyond a robust constrained controller with equivalent observations, plans, and action rights. Methods. We specify S-AI-Decision as a sparse, versioned world-model architecture that separates partially observed belief construction, candidate planning, independent semantic checking, bounded Accept–Clarify–Reject regulation, commitment authorization, and context-bound Ed25519 certification. Its executed reference subset uses Gymnasium 1.2.2 and MiniGrid 3.1.0, trace-based native replay, persistent anti-replay controls, adversarial mutations, finite decision budgets, development-frozen policies, and scenario-matched comparisons. The gate is further evaluated on benchmarks in which commitment risk is hidden, deferred, misspecified or temporal (episodic Tiger, RockSample(5,7) with a POMCP planner, two MiniGrid extensions with hidden deferred damage and expiring permits, and Safety-Gymnasium with learned proposers), against gate-off ablations and equal-information constrained planners. A reference implementation of the hormonal regulator is compared with non-hormonal schedulers under sparse and dense activation on RockSample(7,8). An improvement campaign then tests uncertainty sets calibrated against revealed outcomes, a gate that returns its contract to the proposer, rules that change during operation, and an asymmetric hormonal regulator. The consolidated evaluation uses each campaign’s original eligible environments, seeds, and denominators once; repeated execution and partial replays provide reproducibility checks, not additional independent scenarios. The native adapters are not equated with a complete integration of the historical S-AI-WorldModel, S-AI-Planning, and S-AI-RLM engines. Results. P5-I established official-runtime execution over 64 episodes, 2,124 native transitions, 32 identical matched pairs, and 1,062 local attestations; an included 24-episode subset also passed 15/15 source audit checks and a native replay audit. P5-M exercised the integrated reference verifier for 24 episodes and 1,672 transitions, with 12/12 paired-identical executions and 836 verified semantic attestations; 20/20 contract regressions and 4/4 adversarial checks passed. In P5-V, each of three controllers succeeded on all 200 matched Dynamic Obstacles scenarios (600 controller episodes, 35,139 native transitions, 11,713 verified signatures, 200/200 equivalent paired trajectories), with 60/60 altered claims rejected. On P5-W’s 60 tighter-budget scenarios, C and its gate-off comparator D achieved the same 56/60 successes and 29.9617 cumulative reward, with identical paired actions. On P5-X’s 60 native LavaCrossing scenarios, all B/D/C controllers succeeded, but C generated 90 additional gate deferrals and received lower reward (53.38995) than D (54.187575) or B (54.804075); 1,493 local signatures and ten adversarial modifications were checked. No observed lava termination was prevented by those additional deferrals. On the additional benchmarks, the gate removed harmful commitments when the risk contract was stricter than expected-utility maximization, when the consequence model was misspecified within its declared uncertainty (RockSample: episodes with a bad sample fell from 36% to 0% and the discounted return rose by 1.49), under unreliable proposals (damaged deliveries fell from 106 to 0) and under expiring permits (violations fell from 220 to 0); it did not outperform equal-information constrained planners. In Safety-Gymnasium, an observation-based shield without a dynamics model increased the cost of unconstrained learned policies (by 16 per episode), showing that a gate is only as reliable as its model of consequences. In the regulation study, sparse activation halved end-to-end CPU time but lowered the discounted return by 0.49 and raised the rate of episodes with a bad sample from 0.2% to 2.0%, because the robust specialist was not activated when needed; the hormonal regulator reduced the oscillation of computation quotas about threefold relative to a memoryless scheduler but recovered more slowly from a transient perturbation (by 1.2 cycles). In the improvement campaign, calibrating the uncertainty set cut episodes with a bad sample under a wrong declared interval from 28.8% to 2.0%; returning the violated contract to the proposer brought the gate to parity with the integrated constrained planner (remaining gap in return −0.08); under rules changing during operation, a gate with local repair stayed compliant at 60 to 600 times less computation than full re-solving; and an asymmetric fatigue coordinate halved the recovery gap of the hormonal regulator without removing it. Conclusions. The combined evidence supports an executable and reproducible reference implementation of observation-grounded, locally certified commitments with model-relative semantic and cryptographic checks. An incremental decision advantage H8 is conditionally supported against gate-off ablations and unreliable proposers, and reaches parity with an optimal equal-information constrained planner, without exceeding it, when the gate returns its contract to the proposer; under rules that change during operation, it adapts at a small fraction of the cost of re-solving. The evidence does not establish computational parsimony at non-inferior decision quality, a uniform stability benefit of hormonal regulation, universal reliability, full historical-engine integration, independently delegated real-world authority, or independent external validity. The results distinguish successful certification from improved decisions and report the null and unfavorable controlled comparisons without double-counting re-executed data.