[Tristan-Stoltz-ERC]

Summary

Implements the first increment of MINE-ADAPTIVE-001 #2051: a standalone Rust trace auditor for causal admissibility and non-anticipative counterfactual replay.

• Verifies every declared policy input was delivered no later than its decision timestamp; collection/observation time alone is not enough.
• Blocks declared use of evaluator-only truth, unqualified/rejected or missing results, missing provenance, QC failures, unknown observation references, duplicate references, duplicate IDs and out-of-order decisions.
• Pairs two counterfactual futures that share the same pre-fork delivered-event history; demands identical pre-fork actions and evidence references while requiring a genuine post-fork outcome divergence.
• Rejects a planted pair where policy actions diverge before outcomes become observable.
• Emits exact-corpus SHA-256 + source-commit receipts and reason codes.
• Contains 11 frozen authored synthetic traces and a focused Rust workflow; no third-party mine data or physical measurement actions.

Paradigm shift

A score from a policy that can see future assay results is not merely a slightly optimistic estimate—it is a causally invalid experiment. The counterfactual fork check makes that failure falsifiable by changing only the future outcome while holding earlier observable history fixed.

Research anchors

• Armstrong et al. (2021), Adaptive open-pit mining planning under geological uncertainty: rolling-horizon planning updates the geological model as information arrives; it improved mean outcomes on studied cases but did not win on every realization. https://doi.org/10.1016/j.resourpol.2021.102086
• Khalifi et al. (2026 preprint), Adaptive mine planning under geological uncertainty: A POMDP framework for sequential decision-making: formalizes sequential action and belief updates. Its reported benefits remain claims to reproduce, not fixture values. https://arxiv.org/abs/2605.13702
• Jeuken & Forbes (2024), value of drilling under chance-constrained blending: https://doi.org/10.1007/s42461-024-01061-8

Independent semantic check / CI status

An independently implemented JavaScript checker matched 11/11 expected dispositions and exact reason-code sets, counted five causal violations, accepted the valid counterfactual pair, and rejected both branches of the planted pre-fork disagreement.

Exact head 22dd5e0a5b0bdd46cda177cbb2caee4e18b7295d passed the focused Mine Adaptive Causal Replay Benchmark workflow 39165916229: rustfmt passed, 8 Rust unit tests passed, and the receipt matched all 11 controls, five causal violations, two counterfactual groups, and one rejected pair. Exact-head repo CI run 39165916263 subsequently completed successfully, in addition to focused run 39165916229. This verifies the Rust trace-auditor contract and repository CI at head 22dd5e0a5b0bdd46cda177cbb2caee4e18b7295d; its claim ceiling remains synthetic trace semantics only.

Acceptance targets / limitations

The frozen contract targets zero future-event use, zero use of evaluator-only truth, and identical actions for identical pre-fork observable history. It audits the declared input-event list; by itself it cannot prove a policy executable did not read an undeclared field. A later evaluation harness must capture the complete delivered observation snapshot, hash it, bind policy binary/configuration, and independently verify runtime logs or enforce a sandboxed API boundary.

The public fixtures are not a blinded predictive test. This PR audits trace semantics only—it does not score mining utility and says nothing about real geology, reserve quality, worker safety, or industrial performance.

Next stage: a separate policy runner/evaluator that compares static, two-stage, rolling-horizon, and decision-aware measurement policies against the same held-out synthetic realizations, with clairvoyant plans labeled only as upper bounds. Reuse MINE-SYSTEM, ORE-VOI, and blend contracts rather than duplicating those owners.

Exact-head verification update

Exact head 22dd5e0a5b0bdd46cda177cbb2caee4e18b7295d passed focused causal replay run 39165916229 and repository CI 39165916263. Focused Rust validation passed formatting, 8 tests and the 11-scenario receipt gate; repository CI completed successfully. Claim ceiling remains synthetic trace semantics; no field-performance claim.

[Tristan-Stoltz-ERC]

Cross-domain follow-up from the shipyard recovery work: draft PR #2730 checks multi-scenario information-set consistency at each decision tick, with the expected scenario inventory and source-owned event logs separated from candidate-only action traces. This complements rather than replaces this PR's stronger domain-specific event-delivery/QC/evaluator-visibility checks and paired counterfactual receipt. I have explicitly noted that shared canonicalization/receipt/non-anticipativity primitives should be extracted as the contracts mature, rather than merging two independent implementations without review. #2730's exact-head Rust workflow remains pending, so the new code is not yet validated.