Market title abbreviated for character limit; the full operative title and all binding definitions are in this description.

Audit Motion 002

Title

Will at least one frontier AI company publish recurring organizational governance metrics alongside capability benchmarks, in a single document, on or before 31 December 2026?

Research Thesis

This prediction tests whether frontier AI governance will evolve from policy disclosure toward performance disclosure.

Frontier AI companies now routinely publish governance frameworks, responsible scaling policies, and regulatory compliance documents. This motion asks whether governance itself will become measurable through publicly reported operational performance indicators — rather than remaining a collection of qualitative principles, or being conflated with model-level safety evaluations.

The single-document requirement is a theoretical design choice, not a technical formality: the market tests not merely whether governance metrics appear somewhere, but whether governance performance begins to enter the same public accountability surface as capability evaluation — whether a company is willing to let its governance record be read in the same sitting, by the same audience, as its capability claims.

The underlying hypothesis: governance maturity eventually requires operational metrics in the same way engineering maturity requires benchmarks.

Hypothesized progression:

Governance Principles → Governance Frameworks → Governance Processes → Governance Metrics

This market tests whether the final step begins within the resolution window.

Definitions

Frontier AI company means any of the following named organizations: OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, Microsoft AI, Amazon (AGI/Titan), Mistral AI, DeepSeek, Alibaba (Qwen); or any other organization that qualifies as a “large frontier developer” under California SB 53 / TFAIA thresholds (≥10²⁶ FLOPs training compute and applicable revenue threshold) at the time of publication.

Capability benchmark information means at least one quantitative capability metric — a score, rate, ranking, or equivalent numerical result on a model capability evaluation (e.g., MMLU score, SWE-bench pass rate, internal capability evaluation results reported numerically). Mere mention of benchmark names, qualitative capability descriptions, or hyperlinks to capability results published elsewhere do not satisfy this element.

Organizational governance metric means a quantitative indicator measuring the performance of the company’s frontier-risk governance processes — not the behavior of its models, and not platform content-moderation operations. Examples (non-exhaustive):

• safety review completion rate or throughput

• frontier-safety incident count and incident response time statistics

• internal escalation / override frequency in capability-deployment decisions

• RSP / preparedness-framework execution statistics (evaluations conducted, thresholds triggered, mitigations applied)

• governance or safety audit frequency and findings statistics

• policy exception counts within the safety governance process

• time-to-remediation for identified compliance gaps

• external audit statistics

• governance SLA attainment or equivalent operational indicators

Single document (publication unit) means one self-contained document: a single PDF, a single report, or a single continuous webpage. A website, hub, or portal comprising multiple pages or downloadable documents is not a single document; the co-existence of governance metrics on one page and capability results on another page of the same site does not satisfy the criteria. Both required elements must appear within the same self-contained document.

Recurring — two qualifying types, and the Resolution Memorandum must identify which applies:

• Established Recurrence: the qualifying document is at least the second installment in a series of such disclosures published on or before the deadline.

• Committed Recurrence: the qualifying document is the first installment, but the document itself or its official announcement explicitly commits to a fixed publication cadence (e.g., quarterly, semi-annual, annual).

Resolution Criteria

This market resolves YES if, on or before 31 December 2026 at 23:59 UTC, at least one frontier AI company (as defined above) publicly releases a single document that reports both:

1. capability benchmark information (as defined above); and

2. one or more recurring organizational governance metrics (as defined above).

Explicit Exclusions — This Does NOT Count

The following alone do not satisfy the criteria:

• Model-level safety evaluations of any kind — refusal rates, jailbreak resistance scores, harmful-output rates, red-team pass rates, alignment benchmark scores, or any quantitative measurement of model output behavior. These measure the model, not the governance process, and are already standard system-card content.

• Platform trust-and-safety / content-moderation enforcement statistics — banned account counts, user appeal volumes, appeal overturn rates, CSAM/NCMEC reporting counts, government data request statistics, and DSA-style transparency reporting. Rationale: this disclosure genre has existed at major platforms since 2011–2013 and measures platform operations, not frontier-risk governance. Its presence does not test the progression hypothesis. (This exclusion is decisive at T0: Anthropic’s Transparency Hub already publishes such statistics on a recurring basis; see T0 Baseline Audit.)

• Cross-document aggregation — governance metrics in one document and capability results in another, even if published simultaneously, by the same company, or on the same website/hub.

• Governance principles, values statements, or safety commitments

• Responsible Scaling Policies, Preparedness Frameworks, Frontier Safety Frameworks, or equivalent policy documents, regardless of sophistication

• Regulatory compliance mappings (e.g., TFAIA / EU AI Act Code of Practice alignment documents) that describe processes without reporting quantified process performance

• Organizational charts or governance structure descriptions

• Qualitative descriptions of governance activity (“we conducted extensive reviews”)

• One-off retrospective statistics with no recurrence commitment

• Capability benchmark reports without governance metrics

• Media reporting, third-party analyses, or leaked internal documents

Evidence Hierarchy

Primary (resolution-determining):

• Official company documentation, technical reports, system cards

• Official governance or transparency reports

• Publicly accessible official regulatory filings

Secondary (supporting only):

• Company blog posts and official announcements

Media reporting alone will not determine resolution unless it directly links to official published materials satisfying the criteria.

Resolution Authority & Conflict-of-Interest Protocol

Resolution relies on publicly available official documents published on or before the deadline. Where ambiguity exists, the conservative interpretation is adopted: ambiguous cases resolve NO.

Because the market creator is also the author of the underlying hypothesis, the following procedural commitments apply:

1. Full resolution memorandum. At resolution, a complete written justification will be published, listing every candidate document considered — including all documents evaluated and excluded, with reasons — and, for a YES resolution, identifying whether recurrence is Established or Committed.

2. Pre-registered criteria freeze. The definitions and exclusions above are frozen as of market creation (T0). No post-hoc reinterpretation in either direction.

3. Challenge window. Traders may submit candidate documents at any time; each will be logged and receive a written accept/reject determination in the resolution memorandum.

T0 Baseline Audit (Completed)

Baseline audit conducted prior to launch, confirming that no existing publication satisfies the criteria as of T0. Verdicts:

All documents examined were found Non-qualifying:

1. Anthropic Transparency Hub (platform security / system trust pages; disclosures covering the 2024 H2, 2025 H1, and 2025 H2 reporting periods) — Recurring quantitative statistics exist (banned accounts, appeals, overturns, NCMEC reports, government requests), but (a) these are platform T&S / content-moderation metrics, explicitly excluded; (b) the Hub is a multi-page portal, not a single document — capability/safeguard assessment content resides on separate pages.

2. Anthropic EU DSA Transparency Report (Feb 2026) — Content-moderation enforcement reporting; excluded genre; contains no capability benchmark information.

3. Anthropic RSP v3 and capability/safeguard assessment overviews — Policy and process documents; no recurring quantified process-performance metrics.

4. OpenAI Frontier Governance Framework (May 2026) — Process and compliance-mapping description; no quantified process performance.

5. Google “A Pragmatic Approach to AI Governance in America” (June 2026) — Policy proposal, not performance disclosure.

6. Anthropic Advanced AI Framework (June 2026) — Proposed regulatory regime, not self-reported metrics.

7. Representative current-generation system cards and model reports examined at T0 — Capability benchmarks plus model-level safety evaluations only; no organizational governance metrics. Coverage is representative, not exhaustive; the challenge window remains open for any unexamined document.

T0 Evidence Preservation Protocol

For every document listed in the baseline audit table, the following are preserved as of the audit date: official URL; downloaded PDF where available; screenshot; Internet Archive snapshot; and date of examination. The T0 baseline is thereby fully reproducible and held to the same Evidence Hierarchy this motion imposes on resolution. The preservation index will be published alongside the companion essay.

Boundary watch item (disclosed): Anthropic currently publishes its Transparency Hub and its frontier-compliance documentation as separate publication surfaces. As of T0, no single document satisfying this market’s definitions has been identified. If a future disclosure consolidates, within one self-contained document, (i) quantitative capability results and (ii) recurring non-T&S governance process metrics, it would qualify. Consolidation of this kind is flagged at T0 as the most plausible YES pathway and will receive symmetric scrutiny.

Falsification

The hypothesis (governance progression reaching the Metrics stage under external scrutiny) is weakened if frontier AI companies continue publishing increasingly sophisticated governance frameworks and compliance mappings throughout the resolution period while systematically avoiding quantified operational governance disclosure within their capability-reporting surfaces.

The hypothesis is not confirmed merely by a YES resolution driven by regulatory compulsion alone; the resolution memorandum will distinguish voluntary disclosure from legally mandated disclosure, as the two carry different theoretical weight.

Confidence

Probability: 35%

Reasoning:

Downward pressure (dominant):

• The 2026 regulatory calendar does not compel this specific disclosure. TFAIA requires incident reporting to California OES but does not require public release of governance performance metrics. The NY RAISE Act takes effect 1 January 2027 — after the deadline. EU AI Act GPAI Code of Practice enforcement (August 2026) requires documentation and process compliance, not published quantified operational metrics.

• YES therefore requires at least one company to voluntarily exceed compliance. Frontier governance metrics are structurally negative disclosures (incidents, overrides, exceptions) with no competitive upside.

• Industry-wide pattern to date: capability benchmarks are universally reported; governance reporting remains qualitative or confined to the excluded T&S genre.

• The tightened criteria (single document, quantitative capability element, T&S exclusion) narrow the qualifying surface relative to looser readings.

Upward pressure:

• Three major governance frameworks published within one compressed mid-2026 window indicate active competition on governance credibility, which could extend to metrics as a differentiation move.

• Recurring quantitative reporting infrastructure already exists at multiple companies (transparency reports, compliance documentation); the remaining step to a qualifying document is editorial consolidation rather than new measurement capacity, so the marginal cost of a YES-triggering publication is low.

• Enterprise procurement and insurance markets increasingly demand quantified assurance; one company may find first-mover value.

Resolution-date risk: the most likely YES pathway is a Q4 2026 annual compliance or transparency document positioned ahead of the RAISE Act. This is a narrow window and a specific bet.

Calibration commitment: this probability estimate is itself pre-registered and will be evaluated in the post-resolution Calibration Audit, alongside the divergent estimates recorded during drafting (initial author estimate 65%; adversarial audit estimate 30–40%; final registered estimate 35%). The spread is retained as data.

Publication Chain

• T0: This motion published on Manifold (@EntropyClinic) + companion essay on Entropy Clinic (Substack), bilingual, with Internet Archive timestamp.

• Companion essay (full T0 evidence preservation index with IA snapshots and hashes): Audit Motion 002: When Governance Stops Being a Promise and Becomes a Metric — https://open.substack.com/pub/entropyclinic/p/audit-motion-002-when-governance

• Resolution (on or before 31 Dec 2026): Full resolution memorandum published regardless of outcome, identifying recurrence type if YES.

• Post-audit: Retrospective essay evaluating both the prediction and the prediction process, including the Calibration Audit, per Audit Motion 001 format.