# E124 Forensic Reconciliation

**Audit date:** 2026-11-21
**Entity:** FIRME AI CORPORATION · **System:** CAIN-42
**Scope:** E124 and its E125-A / E126 / E127-A / E127-B descendants
**Rule:** every discrepancy has a resolution or stays visibly UNRESOLVED. No failed experiment is rewritten as a success.

Machine-readable: `artifacts/e125/forensic/E124_METRIC_LINEAGE.json`

## What was inspected

- Branches/worktrees: `/root/cain-e124` (`2dee32731`), `/root/cain-e124-open-world`
  (`48aab80e3`), `/root/cain-e124-persistent-open-world` (`93c8332e1`), and the main tree at
  `466c73b9f` (the branch name `evolution-79` is historical; the content is E124–E128).
- E124 artifacts: capability, causal world model, collective, discovery, self-improvement,
  transfer, neural LoRA, 12-hour mission status/journal, SWE-bench reports, replay, public pages.
- E124.1 bundle, E125-A fresh-variant run, E126 (no usable result), E127-A CV selection,
  E127-B compute-matched control.
- The existing capability, memory, causal, transfer, collective, discovery and self-improvement
  modules; the LoRA run; publication scripts.

## Commit ancestry (as recorded)

| Commit | Claim |
|---|---|
| `15dde64bb` | E124 first $20; 23 official resolves; rank 136→131 |
| `48aab80e3` | E124 deploy save; 111/500 official resolves; rank 127 |
| `2c89deb1c` | E111→E124 evidence panel |
| `5ba528289` | E124/E125/E126 12-hour mission PASSED + replication |
| `42364d0fb` | E127-A CV selection positive on never-used variants |
| `55c77c8a6` | E127-B correction: gain dissolved under compute-matched control |

The working tree's HEAD at audit was `466c73b9f`; E125 was frozen at `be3494d964…` before the
confirmatory seed-126 run.

## The ten mandatory questions

### 1. Why did E124's 97.9% not replicate on E125-A fresh variants?

E125-A used **new data seeds and new parameter draws for the same eight paradigms and the same
primitive library**. That tests *interpolation within a known rule family*, not a new mechanism.
E124's 0.9786 (274/280) was therefore within-grammar performance. On fresh variants the full arm
scored 0.855 versus a compute-matched 0.925. E127-B later showed the apparent recovery was
compute. E125 replaces the question with an irreducible-rule challenge and separate A–E
categories.

**Status: RESOLVED.**

### 2. Why 85.5% versus 92.5% for the compute-matched control?

Selection variance. Several depth-2 hypotheses fit all 12 training rows exactly; fixed-order
search kept a spurious one, while the random-order control happened to keep a better one. The
paired difference was real in that run (6 vs 20, exact McNemar p = 0.00936), but it measures
search order, not understanding.

**Status: RESOLVED.**

### 3. Were the original held-out examples genuinely unseen mechanisms?

No. They were fresh instances of the same mechanisms. E125 defines A (new examples), B (new
combinations), C (new irreducible rules), D (new primitives), E (new environments), and admits a
C family only if **no depth ≤ 2 rule in the same vocabulary fits its training rows**.

**Status: RESOLVED.**

### 4. Why did the self-improvement system promote zero of 30 candidates?

The baseline sealed accuracy was 0.975. Candidates that gained on one slice lost on the
regression or canary slice; promotion required a held-out gain and every governance stage.
Zero promotion is the correct outcome for those candidates, not proof that self-improvement is
impossible. E125 repeats the test with 30 evaluated candidates, 4 refused protected targets, and
again **0 promoted**.

**Status: RESOLVED.**

### 5. Why did LoRA change weights but not improve held-out performance?

The E124 evaluation asked a 0.5B CPU model for multi-digit arithmetic in a raw continuation
format. Before and after, every prediction was the token `1`; training loss moved 2.1755 → 1.9227
over 160 steps. This is an evaluation-format and capacity failure, not evidence that parameter
updates cannot help. E125 diagnoses it and runs a representable lexicon task; LoRA (430,080
parameters) reaches 0.667 held-out, full fine-tuning reaches the high 0.9s, frozen is 0.25.

**Status: RESOLVED BY THE E125 ADAPTATION EXPERIMENT.**

### 6. Why was the E124 causal decision gain only 0.0026?

The active policy barely differed from passive (0.4453 vs 0.4427) in a discrete two-cause world.
The chosen interventions carried little information. E125 uses **Markov-equivalent graph pairs**
where passive observation provably cannot orient the edge: active 0.75 vs passive 0.625
(+12.5 points, p = 0.0018). Active is **not** better than random (0.75 vs 0.75, p = 1.0); that is
reported as a negative sub-result.

**Status: RESOLVED, WITH A NEW NEGATIVE (ACTIVE ≠ BETTER THAN RANDOM).**

### 7. Does the 0.0492 collective advantage survive stronger compute-matched controls?

**UNRESOLVED for E124.** The original run did not match total inference budget or independent
evaluation. E125's own collective uses paired agents at 12 calls/task: dynamic 0.70 vs best single
0.483 (+21.7 points, p = 6.5e-8). That does not retroactively validate E124's 0.0492.

**Status: UNRESOLVED FOR E124; NEW E125 RESULT REPORTED SEPARATELY.**

### 8. Is the 12-hour mission reconciled across surfaces?

E123's gate matrix says G11 (12-hour mission) BLOCKED because that page predates the E124 run.
E124's signed status says COMPLETED at **43,210.59 s**, 2,142 cycles, 2,136 useful, 0 failed,
214 checkpoints, journal verified, head `84f6d6ee…`. Both can be true at different times, but the
public pages do not display the timestamp distinction. The elapsed time is supported. “Useful” is
the engine's own counter, and the content is repetitive capability induction, not a 12-hour
diverse research mission.

**Status: PARTIALLY RESOLVED.** Elapsed time supported; productive diversity not demonstrated.

### 9. Why do public SWE-bench counts differ?

`clawx-site/evidence/e124-1-2026-11-21/evidence.json` (05:14:53Z) records **189 resolved /
486 evaluated / rank 110**. Later official report files bring the union to **257 unique resolved
/ 500 evaluated / rank 75**. We verified the union directly: 89 official report files, 279
report-level resolutions, 257 unique IDs, and **zero** IDs in `full_results.json` outside the
official union. The 189 figure is a stale snapshot, not a contradiction of the raw reports.

**Status: RESOLVED AS A STALE SNAPSHOT.** The E124.1 bundle must be labelled historical.

### 10. Are the published results reproducible from a clean checkout?

A replay on 2026-11-21 found transfer, collective, discovery and causal results byte-identical;
**self-improvement was not** (`artifacts/e124/replay_2026-11-21/replay.json`). The capability
challenge was not rerun from a clean checkout in that replay. E125 ships a standalone verifier
that recomputes its own numbers from raw rows.

**Status: PARTIALLY REPRODUCIBLE.**

## What this audit does not claim

- It does not claim E124 fabricated data. The files exist and most sibling artifacts replay.
- It does not claim the E124 12-hour run was fake. The elapsed time and journal are supported.
- It does not claim the E124 SWE-bench 257 is an official leaderboard position. It is a self-run
  aggregate, and the official leaderboard was not verified.
- It does not claim the E125 result is independent. The same team wrote the engine and generator.
