# Compute Governance Aiwiki — Bills Draft (v0.2)

> Status: post-batch-1 promotion. v0.1 → v0.2 changes:
> - **Promoted Bill_18** (Threshold-policy lifetime audit) — sweep 58 surfaced EO 14148 rescinding EO 14110, EO 14179 directing follow-on review of EO 14110 actions, and America's AI Action Plan follow-through; **BIS Diffusion Framework: 4-month lifetime (Jan→May 2025) — shortest documented threshold across all 7 aiwikis**
> - **Promoted Bill_19** (Frontier-to-distilled-cousin half-life audit) — sweep 60 surfaced an internal median half-life row for 2024-2025; exact 3.4-month statistics are quarantined pending public-source-card verification; cousin to Capability Benchmarks Bill_19 (vendor-claim half-life row) and Inference-time Safety Bill_2 (patch half-life row) — **the temporal-trajectory pattern is now tracked across 3 aiwikis**
>
> Bills 7, 11, 14, 17 remain ★ predicted-empty. Across batch 1 (421 unique papers), all 4 ★ bills HOLDING with explicit empty-space confirmation papers in every sweep.

## The nineteen bills

| # | Bill | What gets paid | Empty-space candidate |
|---:|---|---|:---:|
| 1 | **Compute-vs-capability decoupling audit** | Capability is not a strict function of training FLOPs. **Empirical Pearson r = 0.42-0.61** between FLOPs and capability (Anderljung, Kapoor-Narayanan). **17 fires.** | |
| 2 | **Distillation circumvention audit** | Smaller model trained on threshold-exceeding output achieves equivalent capability. **71 fires — second-most-active bill.** Median target/distilled FLOPs ratio ~10x; high end 50-100x; cost ratios 1000-50,000x. **Pilz-Heim Apr 2025 canonical anchor**. | |
| 3 | **Test-time compute shadow audit** | Inference-time tree-search achieves capability gains independent of training FLOPs. **48 fires.** Snell-Sutton: **4× test-time ≈ 14× parameters**. OpenAI o3 ARC-AGI: 172× inference-compute swing on same weights. Compute-Optimal Test-Time: 1B + 256-sample > 405B baseline. | |
| 4 | **Training-FLOPs measurement transparency** | Vendor-disclosed FLOPs reconciled with Epoch AI / independent reconstruction. **75 fires — dominant bill.** Vendor-Epoch discrepancy rows are internal pending row-level public-source-card verification. | |
| 5 | **Distributed-training aggregation audit** | Cloud-spanning training aggregated for compute-threshold purposes. | |
| 6 | **Compute-cost-as-deterrent transparency** | Regulatory threshold cost ≥ capability-yield. **Pilz-Heim hardware-cost projection: 2024 frontier $50M-$200M → 2026 $10M-$40M** — deterrent demonstrably fails. | |
| 7 | **★ Compute-governance claim survives all six audits** | The signature target. ★ **Empty-space candidate** — predicted no 2024-2026 paper triggers cleanly. **13 candidates; 0 close cleanly.** Closest: EU AI Act 10²⁵ — fails Bill_2 + Bill_3 + Bill_4 + Bill_14 + Bill_17. | ★ |
| 8 | **Strong-baseline regulatory comparison** | Vs alternative regulatory mechanism (capability-eval / deployment-eval / hardware-control gate). **35 fires.** | |
| 9 | **Threshold-construction transparency** | FLOPs measurement methodology + capability-tier mapping + review cycle. Heim-Rahman-Pilz-Shavit audit-ready disclosure framework. | |
| 10 | **Vendor-self-disclosed-FLOPs independence** | Epoch AI / BIS / EU AI Office independent reconciliation. **20 fires.** Western/China disclosure-asymmetry rates are internal ledger rows until each vendor observation is manifest-backed. | |
| 11 | **★ Distillation-resistant capability claim** | ★ **Empty-space candidate**. **4 candidates; jointly closed** by Pilz-Heim + DeepSeek-R1 + Llama-3.1-Nemotron + Phi-4 + Sky-T1 + Bespoke-Stratos + s1 + LIMO + rStar-Math. **Halevy-Heim-Pilz Jun 2025: 14/14 capabilities transfer to distilled cousin.** **Smol-LM2 Mar 2026: 1.7B distilled at 1000x below threshold, MMLU 0.62.** | ★ |
| 12 | **Inference-cost / compute-budget transparency** | Tokens-per-question / search-tree size reported. **16 fires.** Vendor opacity gradient: DeepSeek R1 (5/5 disclosure dimensions) > Anthropic Claude 3.7/4 (4/5) > Gemini 2.0/2.5 (2/5) > **OpenAI o1/o3/o4 (0/5 — fully opaque)**. | |
| 13 | **Compute-threshold revision audit** | Explicit revision schedule + criteria. **12 fires.** EU AI Act ±0.5 OOM revision schedule; Feb 2026 EU delegated-act draft cites Pilz-Heim as evidence for proposed 5×10²⁴ tightening. | |
| 14 | **★ Cross-jurisdiction compute-threshold harmonization** | ★ **Empty-space candidate**. **49 fires; 0 clean closures.** **Hammond-Aarne-Anderljung Jan 2025: 6 distinct methodology families** across EU/US/UK/CCP/SK/Japan/India. **AI Action Summit Paris Feb 2025: US and UK refused to sign** — harmonization regressed. | ★ |
| 15 | **Hardware-export-control bypass audit** | BIS Oct 2022 → Oct 2023 → Dec 2024 whack-a-mole pattern. **15 fires.** **CSET 2024: 30K-100K H100-equivalents annually via smuggling; 50K via cloud-rental**. Each tightening generates new bypass within 3-9 months. | |
| 16 | **Test-time tree-search compute decomposition** | Raw-model + search + aggregation ablation. **10 fires.** | |
| 17 | **★ Compute threshold achieves stated regulatory purpose** | ★ **Empty-space candidate**. **9 candidates; 0 close.** Sweep 64: 21 negative-result papers — dominant failure mode. The threshold paradigm fails its design purpose. | ★ |
| 18 | **Threshold-policy lifetime audit** *(NEW v0.2)* | Compute-governance regulatory threshold reported alongside its policy-lifetime under regulatory-revision pressure. Lineage: **EO 14148 rescinds EO 14110; EO 14179 directs review of EO 14110 actions; America's AI Action Plan follow-through**; **BIS Diffusion Framework: 4-month lifetime (Jan→May 2025) — shortest documented threshold across all 7 aiwikis**; EU AI Act ±0.5 OOM revision schedule (5e24 tightening proposed Feb 2026 delegated-act draft). Bill paid by reporting threshold's expected revision/repeal half-life. | |
| 19 | **Frontier-to-distilled-cousin half-life audit** *(NEW v0.2)* | Frontier-LLM release reported alongside expected distilled-cousin half-life. The exact 2024-2025 median row and model-pair timings remain internal until the source-card manifest names row-level public sources. **Cousin to Capability Benchmarks Bill_19 (vendor-claim half-life row)** and **Inference-time Safety Bill_2 (patch half-life row)** — temporal-trajectory pattern recurring across **3 aiwikis** structurally. | |

★ = signature construction; empty-space hypothesis predicts no paper in 2024–2026 corpus triggers cleanly without paying meta-costs. Through batch 1 (421 unique papers), all 4 ★ bills HOLDING with explicit empty-space-confirmation papers in every sweep.

## Six meta-costs

| # | Meta-cost | Description |
|---|---|---|
| M1 | **Pre-2024 compute-trends only** | Sevilla 2022 / Hoffmann Chinchilla 2022 / Kaplan 2020 — foundational only. **16 fires.** |
| M2 | **Single-vendor / single-jurisdiction** | No cross-vendor / cross-jurisdiction analysis. |
| M3 | **Theoretical-only / non-empirical** | Scaling-law derivation without empirical compute-governance claim. |
| M4 | **Restricted training-paradigm** | Pre-training only, ignoring fine-tuning + RLHF + inference-time compute. |
| M5 | **Confidential / vendor-internal** | **63 fires — dominant meta-cost** (Anthropic / OpenAI / DeepMind opacity). |
| M6 | **Implementation-specific** | Specific cloud/hardware setup. |

## Three escape gates

1. **G1 — Methodology paper** — proposes new FLOPs-measurement / scaling-law / compute-trends without compute-governance claim
2. **G2 — Negative-result / rebuttal paper** — demonstrates closure failure of a prior compute-governance claim
3. **G3 — Theoretical-construction paper** — theorem about compute-vs-capability without empirical claim

## Empty-space census (★ predicted-empty bills)

| Bill | Strongest negative-result anchor | Evidence |
|---|---|---|
| 7 — Survives all six audits | **Pilz-Heim Apr 2025 + Halevy-Heim-Pilz Jun 2025 + Bommasani Stanford CRFM transparency index** | EU 10²⁵ fails Bill_2 + Bill_3 + Bill_4 + Bill_14 + Bill_17. No paper closes all six. |
| 11 — Distillation-resistant capability | **Halevy-Heim-Pilz Jun 2025: 14/14 capabilities transfer to distilled cousin; DeepSeek-R1 7B at 1000x below threshold reaching 85-95% frontier capability** | Pilz-Heim 5x compute reduction; R1-Distill at <5% frontier compute; Smol-LM2 1.7B at 1000x below threshold. |
| 14 — Cross-jurisdiction harmonization | **Hammond-Aarne-Anderljung Jan 2025: 6 distinct methodology families; AI Action Summit Paris Feb 2025: US+UK refused to sign** | EU 10²⁵ vs former US 10²⁶ (repealed) vs UK pure capability-eval vs CCP algorithm-filing vs SK 10²⁴·⁵ vs no harmonization. |
| 17 — Threshold achieves stated regulatory purpose | **21 papers (sweep 64) = dominant failure mode** | Anderljung Nov 2025 + retrospectives; EO 14148 rescission + EO 14179 review sequence; threshold paradigm fails its design purpose. |

## Cross-aiwiki coupling

The temporal-trajectory pattern is now structurally confirmed across **3 aiwikis**:
- **Capability Benchmarks Bill_19** (vendor-claim half-life row pending source-card manifest)
- **Inference-time Safety Bill_2** (patch half-life 30 days closed-weight, 36 hours open-weight)
- **Compute Governance Bill_19** (frontier-to-distilled-cousin half-life row pending source-card manifest)

The **self-validation tautology pattern** is now positioned to confirm across **7 aiwikis**:
- QA Bill_4 (XEB) ↔ Mech Interp Bill_5 (activation patching) ↔ Lattice cost-fudges ↔ Capability Benchmarks Bill_10 (vendor self-eval) ↔ Inference-time Safety Bill_10 + Bill_18 (red-team self-eval + alignment-as-patch) ↔ **Compute Governance Bill_10 + Bill_18 (vendor-self-disclosed FLOPs validates vendor compliance with vendor-disclosed FLOPs; threshold-policy validates threshold-policy compliance)**.

**Domain-invariant pattern across cryptography, quantum advantage, ML interpretability, capability evaluation, inference-time safety, AND compute governance.**