OTIP: A Falsifiable Anti-Goodhart Protocol for Testing Ontological Commitments via Interventions in Reinforcement Learning Agents

Patch release of the OTIP Registry Pack (v0.2.1) containing corrected runnable artifacts for Phase-1, Phase-2R, and Phase-3B1.1, plus verification materials showing that patched reruns reproduce the archived FINAL result tables without changing the archived reference outputs. DOI: 10.5281/zenodo.18900300. Description (Zenodo, OTIP v4.0) This release (v4.0) consolidates the final, sealed OTIP toy-suite into a single, fully reproducible paper package. OTIP (Ontology Through Intervention Protocol) is presented as a proof-of-principle evaluation template for testing whether an agent’s reported/implicit ontological commitments remain stable under targeted interventions, while explicitly guarding against reward-proxy exploitation (Anti-Goodhart framing). The work is intentionally scoped to a small, interpretable reinforcement-learning toy domain. New in v4.0: A complete DEV vs FINAL (sealed holdout) reporting protocol: the main paper reports FINAL-only results, with DEV used solely for test design and diagnostics. Phase-1 suite consolidation (baseline vs OTIP) with planned ablations (A1–A3) and Negative Control, plus strict NC (nc_warmup=0) reported in the appendix. Phase-2 and Phase-2R “mid-episode mechanism switch” tests (H5) demonstrating a falsifiable calibration signature: conservative UNKNOWN behavior under distribution shift vs “confidently wrong” ablations. Phase-3 stress-test families: (B1.1) noisy interventions with “given-success” metrics to disentangle actuator failures from adaptation; (B2) a proxy-shift Goodhart test; and (B3) delayed-effect temporal perturbation (efficiency sensitivity without catastrophic failure). A bundled registry pack (tables + runbook + hashes) enabling independent re-runs and auditability. This release does not claim general ontology learning or AGI alignment. It provides an interpretable, falsifiable, and reproducible toy benchmark/evaluation scaffold that can be extended to additional domains and stress-test families. Ontology Through Intervention Protocol (OTIP) v3.0.1 is a proof-of-principle evaluation template for testing whether an agent’s claimed “ontology” (causal law hypothesis) is functionally grounded in interventions rather than proxy-optimized behavior. The package implements a compact toy-domain benchmark and a transparent reporting protocol aligned with an Anti-Goodhart stance: performance is not sufficient—claims must be auditable, falsifiable, and robust to negative controls. What’s new in v3.0.1 Phase-1 ablations (A1–A3) added and reported under a sealed holdout protocol (DEV vs FINAL seeds). Main results are reported FINAL-only; DEV is used strictly for development and debugging. Strict Negative Control (NC) reported as a supplementary robustness check with nc_warmup=0 (in addition to the standard NC setting). Restored/added reporting metrics for Phase-1 tables, including H2 (OSI) and calibration (Brier score) alongside H1/H3, shift performance, trap/Goodhart indicators, and NC outcomes. Clarified “scientific hygiene” commitments: each component change is linked to expected metric-level signatures (confirmable/falsifiable). Core evaluation signals (toy domain) Causal recovery and counterfactual correctness (H1/H3) Shift handling and switching time (D) Anti-Goodhart trap resistance (H4_goodhart_gap, trap_probe_rate) Negative Controls to detect hallucinated causal knowledge (HNC + strict HNC0) Auditability (tamper-evident logs / commit-chain checks) Scope and limitations OTIP v3.0.1 is intentionally scoped as a toy-domain proof-of-principle and an evaluation/reporting scaffold, not a claim of general intelligence, real-world causal inference, or “consciousness.” Results should be interpreted as demonstrating that an intervention-grounded evaluation pipeline can be made reproducible and falsifiable in a minimal setting. Reproducibility The archive includes an Overleaf-ready paper source and runnable minimal demo scripts to reproduce Phase-1 results, including the sealed DEV/FINAL split and strict NC reporting. DOI: 10.5281/zenodo.18203787 This record contains the v2.0 paper and reproducibility materials for OTIP — Ontology Through Intervention Protocol, a falsifiable, Anti-Goodhart evaluation protocol for testing whether an agent maintains functional ontological commitments: stable, testable claims about entities/relations/causal structure that (i) support interventional (“do-operator”) prediction, (ii) generalize under distribution shift, and (iii) resist metric gaming. What OTIP tests OTIP treats “ontology” operationally: the agent must commit, at the episode level, to: an ontology snapshot (entities + relations + a law hypothesis when applicable), probabilistic do-predictions for defined interventions, explicit falsifiers (conditions under which the current law/ontology must be rejected),all written to an append-only, hash-chained audit log to prevent post-hoc rewriting. Protocol structure The protocol includes four complementary conditions: Stable regime (L1) — evaluates ontology recovery and interventional prediction when the causal law is stationary. Law shift (L2) — evaluates distribution-shift adaptation and time-to-switch (T_switch). Goodhart trap (proxy-only reward) — checks resistance to proxy gaming under misleading incentives. Negative control (no causal link) — checks epistemic calibration (“knowing when it does not know”) when no stable law exists. Minimal reproducible demonstration (20 seeds) In the included runnable demo, an ontology-based agent: achieves near-perfect performance in the stable regime (H1 graph F1 ≈ 0.993 ± 0.014; H3 do-prediction accuracy ≈ 0.993 ± 0.014; Brier ≈ 0.004 ± 0.007), adapts rapidly under law shift (mean shift accuracy ≈ 0.856 ± 0.072; median T_switch = 3 episodes), shows no proxy exploitation in the Goodhart trap (trap probe rate = 0; Goodhart gap = 0), passes the negative control with calibrated epistemic humility (HNC pass rate = 1.0),and all runs pass the append-only log integrity audit. Notes on negative control methodology The negative-control phase in v0.3.8 uses longer NC evaluation with a short warm-up to ensure sufficient post-intervention evidence per episode (details in the paper). For transparency, stricter NC variants (e.g., no warm-up) are discussed as auxiliary metrics. Intended use OTIP is intended as a transparent template for evaluating agentic systems in a way that is: falsifiable (explicit rejection conditions), Anti-Goodhart (sealed-style evaluation logic + proxy trap + negative controls), auditable (commitment integrity via hash chaining),and extensible to richer environments, partial observability, stochastic causality, and stronger baselines. Author: Karel HrubecVersion: v2.0DOI: 10.5281/zenodo.18202077 OTIP (Ontology Through Intervention Protocol) is a falsifiable, Anti-Goodhart evaluation protocol for testing whether reinforcement-learning agents develop and maintain functional ontological commitments (entities, relations, and causal mechanisms) that remain stable under distribution shift, support interventional (“do-operator”) reasoning, and resist metric hacking. This release (v0.1) provides a paper-ready protocol specification with: an episode-level commitment mechanism (hashed, append-only logs) for ontologies, probabilistic predictions, and explicit falsifiers, a sealed holdout evaluation design to reduce leakage and Goodharting, an out-of-band auditor for tampering detection (hash-chain integrity, violation reports), two environment families (a ground-truth Causal Playground with switchable causal laws; and an Operational Sandbox scored via hidden holdouts and auditing), pre-registered hypotheses H1–H4, acceptance thresholds, baselines, ablations, negative controls, and reproducibility checklists. OTIP is intended as a methodological benchmark for “machine ontology” in the functional sense (testable commitments that guide action), not as a claim about consciousness.DOI: 10.5281/zenodo.18196498.