Process-Aware Observable-Only Backcasting Meta-Layer (POB-ML) is a deterministic, audit-ready protocol specification for autonomous agents operating under observable-only constraints—no privileged meta-view, no hidden state, and no undeclared I/O. The specification defines a complete evidence surface (observations, inbox entries, governance updates, candidate traces, evaluation artifacts, InputSet bindings, hold events, safety certificates, and optional audit/notary packets) that is schema-closed, integrity-chained, threshold-signed, and replay-verifiable. POB-ML formalizes backcasting as boundary-conditioned selection over committed candidate ExpectedFutureTrace objects. Decision-influencing FOQL programs are statically validated (parse/type/schema-closure, guard-safety for invalid-sensitive reads, and bounded evaluation cost) before they can affect decisions. Candidate processing is deterministic and budgeted: bounded scan, deterministic rejection with reason codes, exact deduplication, bounded evaluation, and deterministic selection with explicit tie-break rules. A central contribution is deterministic decision input binding via InputSet and bounded InputSetUpdate packets, including a fixpoint-safe exhaustion rule that forces safety-only behavior if update budgets are exceeded. Governance updates use delayed key activation to avoid brittle “immediate activation” failures in asynchronous environments. Safety is enforced by a runtime-assurance Action-Gate that dominates planning. If a committed safety-case envelope is violated, an emergency action is selected deterministically; otherwise the gate selects among feasible non-fallback actions under conservative non-stopping semantics, and falls back to a last-resort action if needed. Executed-action outcomes (including failures) are recorded as observable claims without breaking determinism. To preserve cross-architecture replay, POB-ML imposes a deterministic numeric boundary: platform-dependent floating-point transcendental intrinsics (e.g., libm sin/cos/exp/log) are forbidden on decision-critical paths. Optional MathKernel commitments provide deterministic approximation behavior with mandatory self-tests. Optional extensions include VRF-based public randomness with deterministic PRG receipts, causal-density metrics for auditing dependence on external observations vs internal artifacts, and a ZK proof packet pattern for privacy-preserving attestations while keeping the log replay-locatable.
K Takahashi (Wed,) studied this question.