# Cross-Ledger Bridges Meta-Aiwiki — Batch 1 Synthesis (2026-05-09)

## Stage 3 SWEEP complete

5 sweeps delivered, 104 total records aggregated.

| Sweep | Theme | Records |
|---|---|---:|
| 701 | Intrinsic bridge records | 7 |
| 702 | External corroborating audits | 41 |
| 703 | External rebutting evidence (adversarial) | 32 |
| 704 | Audit cycle re-poll plan | 3 |
| 705 | 14th-ledger inheritance prediction | 21 |

## Bridge support / rebut tally (normalized)

| Bridge | Supports | Rebuts | Net | Status |
|---|---:|---:|---:|:---:|
| **B1** — Causal mechanism empty across 7 domains | 12 | **10** | +2 | ⚠️ thinnest |
| **B2** — Closure cycle compressed 18mo → 3-4mo | 14 | 4 | +10 | ✓ holds |
| **B3** — Caps transfer; mitigations don't | 15 | 3 | +12 | ✓ strongest |
| **B4** — Distillation = arch-portability = scaling-portability | 7 | 3 | +4 | ✓ holds |
| **B5** — '0/N' audit pattern | 13 | 6 | +7 | ✓ holds |
| **B6** — Anti-saturation is only working closure | 7 | 5 | +2 | ⚠️ thin |
| **B7** — Western-vs-Chinese disclosure inversion | 6 | 3 | +3 | ~ marginal |

## 14th-ledger inheritance predictions

For the 3 next-Queued ledgers (Robotics / Multilingual / RAG):

| Bridge | Extends | Fails | Inverts | Unclear | Pre-audit verdict |
|---|---:|---:|---:|---:|:---|
| B1 | 2 | **1** | 0 | 0 | ⚠️ Robotics may FALSIFY: RL-from-physics-grounded-reward could provide causal mechanism LLM-domains lack |
| B2 | 2 | 0 | 0 | 1 | ✓ Likely holds; Multilingual "two-speed closure" is unclear case |
| B3 | 3 | 0 | 0 | 0 | ✓ Robustly extends to all 3 |
| B4 | 1 | **1** | 0 | 1 | ⚠️ Robotics substrate fractures (flow matching vs autoregressive tokens) |
| B5 | 3 | 0 | 0 | 0 | ✓ Strongest — extends cleanly to all 3 |
| B6 | 3 | 0 | 0 | 0 | ✓ Anti-saturation methodology generalizes |
| B7 | 1 | **1** | 0 | 1 | ⚠️ RAG INVERTS: Western open-source frameworks dominate |

## The honest signals

**Bills 9 ★ + 12 ★ are looking PREDICTIVE:**

- **Bill 9 ★** (single-counter-example falsification): predicted-empty PARTIALLY FALSIFIED already by Anthropic Tracing Thoughts attribution graphs on Claude 3.5 Haiku (2025), Goodfire Ember production API, "Lie to Me" 89.9% CoT faithfulness on DeepSeek-V3.2. The 10 external rebuttals on B1 are exactly the kind of counter-example construction Bill 9 ★ predicted. B1 is the weakest bridge.

- **Bill 12 ★** (universal cross-ledger pattern): predicted-empty PARTIALLY CONFIRMED. Inheritance check shows 3 bridges (B1, B4, B7) fail to extend cleanly to the next 3 Queued ledgers. The 7-way star-mechanism alignment is *probably* a 7-LLM-domains-only finding, not domain-invariant in the strict sense — Robotics with grounded-reward likely closes Bill 4-equivalent.

- **Bill 7 ★** (empirical-only / no-narrative test): partially supported. B6's +2 net and B1's +2 net are thin enough that prose framing is doing meaningful work. B3, B5 are robust to narrative stripping.

## Implications for the synthesis preprint

The harness has surfaced what we predicted it would surface — that's the audit working, not failing. Specifically:

1. **B1 must be renamed and rescoped**. Current: "Causally-faithful mechanism is empty across 7 domains." Revised: "**Causally-faithful mechanism is empty across 7 LLM-centric domains**; predicted to fail extension to grounded-reward RL domains." This is a stronger, more falsifiable claim, not a weaker one.

2. **B4 needs substrate-conditional framing**. The distillation-portability-scaling-portability equivalence holds for autoregressive-token substrate; flow-matching substrate (robotics action prediction) may not inherit.

3. **B7 needs temporal qualifier**. "As of 2026-05-09, Western-vs-Chinese open-weight disclosure is inverted." Mistral Large 3, Llama 4 Apache 2.0 are mid-flip; RAG-framework Western dominance inverts again.

4. **B3, B5, B6 hold robustly.** These three are the load-bearing findings of the synthesis. The 6-way / 4-way star-mechanism alignment is real; the 0/N audit pattern is real; anti-saturation as constructive closure is real.

5. **B2 needs operational discipline**. "Closure cycle compressed 3-4 months" should be reported as a *range* (30-100 days half-life across 3 distinct metrics) rather than a point estimate. The "10× faster than policy" framing should be retired in favor of "3-30× depending on metric × policy choice."

## Synthesis preprint revision plan (pre-2027 audit cycle)

Rather than wait until 2027-Q3 lock, the batch-1 sweep has already produced an action list for revising the synthesis abstract:

- **Title (unchanged):** *"Domain-invariant closure-pattern failures in frontier ML claims (2024–2026): a 13-ledger meta-audit"* — but with explicit qualifier: "across LLM-centric domains."
- **Reframe B1 / B4 / B7** per above.
- **Hold B2 / B3 / B5 / B6** as load-bearing.
- **Add Section 7 (Predicted Falsifiers):** explicitly list the 3 Robotics / RAG predictions where bridges are expected to break.

## What this aiwiki did, in one sentence

The harness was pointed at our own synthesis, and within one batch surfaced 3 specific weaknesses in our 7 bridges + 3 specific predictions where 14th-ledger work will falsify them — **before** the synthesis was submitted to arXiv. The discipline worked.

## Next stage

Stage 4 CLASSIFY — build `scripts/bill_classifier.py` for this meta-aiwiki + 30-case hand-curated benchmark. Then Stage 5 POPULATE wiki/ with per-bridge / per-evidence / per-prediction markdown. Then Stage 6 REVIEW.

Lock target unchanged: **2027-Q3 audit cycle complete**. The synthesis preprint is parked until then.