Coalition-stable equilibria in multi-instrument autonomous materials discovery (e.g., phase-change memory) will emerge when instrument-specific data streams are modeled as cooperative agents with thermodynamic admissibility constraints, reducing false-positive material candidates by ≥40% (Bridges: Real-time Multi-instrument Autonomous Discovery × Computing Equilibrium beyond Unilateral Deviation × Discovering Thermodynamically Admissible Dissipation Potentials).

In a multi-instrument autonomous materials discovery loop (≥3 heterogeneous characterization instruments, e.g., XRD, Raman, resistive-switching electrical probe, applied to phase-change memory candidate library), modeling each instrument's data stream as a self-interested agent in a cooperative game — with candidate-material proposals accepted only if they lie in a coalition-stable equilibrium (no subgroup of instrument-agents can profitably deviate to a different consensus) AND satisfy thermodynamic admissibility constraints (non-negative dissipation potential, Onsager-consistent flux-force relations) on the proposed phase-transition pathway — will reduce the false-positive rate of candidate materials (defined as candidates that pass initial autonomous screening but fail full-panel confirmatory characterization) by ≥40% relative to a baseline unilateral/majority-vote fusion pipeline, at equal or lower true-positive rate (≤5 percentage point TPR loss), measured over ≥200 candidate materials per condition.

Minimum viable test: retrospective + prospective hybrid on a phase-change memory (PCM) candidate library.

1. Retrospective arm: mine existing autonomous-lab datasets (or generate synthetic-but-instrument-realistic data via simulated XRD/Raman/electrical noise models) for ≥500 previously screened PCM candidates with known confirmatory outcomes.
2. Prospective arm: deploy on a physical or high-fidelity simulated autonomous lab (e.g., A-Lab-style or synthetic digital twin) for ≥200 new candidates, split randomly into baseline-pipeline vs coalition-stable-pipeline arms (or run both pipelines on identical raw sensor streams for direct paired comparison — preferred, removes randomization noise).
3. Compare FP rate, TP rate, time-to-decision, and compute cost between: (a) baseline unilateral/majority-vote fusion, (b) thermodynamic-filter-only ablation, (c) coalition-stability-only ablation, (d) full combined method.
4. Confirmatory ground truth via full-panel high-fidelity characterization (e.g., synchrotron XRD + TEM + full I-V switching cycle testing) on all candidates in all arms (blind to which pipeline flagged them).

# Pseudocode: Coalition-Stable Thermodynamically-Gated Fusion

for candidate in candidate_stream:
    signals = {instrument_i: get_reading(instrument_i, candidate) for i in instruments}

    # Step 1: Thermodynamic admissibility gate
    pathway_hypotheses = propose_phase_pathways(signals)
    admissible = [
        h for h in pathway_hypotheses
        if dissipation_potential_surrogate(h) >= 0
        and onsager_reciprocity_check(h) == True
    ]
    if not admissible:
        reject(candidate); continue

    # Step 2: Cooperative game formulation
    def characteristic_function(coalition_S, hypothesis_h):
        joint_evidence = combine_likelihoods(
            [signals[i] for i in coalition_S], hypothesis_h
        )
        return joint_evidence  # v(S)

    # Step 3: Compute (approx.) core / Bondareva-Shapley stable set
    core_set = compute_approx_core(
        instruments, characteristic_function, admissible_hypotheses=admissible,
        method="linear_programming_core_approx", timeout_s=10
    )

    # Step 4: Decision rule
    if core_set.nonempty() and core_set.best_hypothesis.stability_margin > 0:
        accept(candidate, hypothesis=core_set.best_hypothesis)
    else:
        reject(candidate)

    log_metrics(candidate, latency, thermo_score, core_stability_margin)

# Evaluation harness
for pipeline in [baseline_majority_vote, thermo_only, coalition_only, full_method]:
    run_on(candidate_set, pipeline)
    compare_to_ground_truth(confirmatory_panel_results)
    compute(FP_rate, TP_rate, latency, McNemar_p, bootstrap_CI)