# E102 Phase 0: forensic baseline and evidence reconciliation (2026-11-19)

## Lineage
- `evolution-100` = `a33b5fab1` (E100 implementation). The deploy commit `af4537b17` sits on `evolution-79` and publishes the same `artifacts/e100/campaign_summary.json`. The bundle published at `clawx.click/evidence/e100-integrated-organism-2026-11-19/` is byte-identical to the repo copy, and its verifier passes 32/32.
- Before this session, `evolution-101` pointed at the same commit as E100; the only E101 commit was a proposal document (`f1edab7a1`). The E101 implementation sat uncommitted in `/root/cain-e101-worktree` and was finished by another session at 03:13Z. It is preserved unchanged as `cb848cebd` on `evolution-101` and merged into `evolution-102`.
- E102 is built on branch `evolution-102`, worktree `/root/cain-e102-worktree`.

## Why the public sites show older evolutions
The three homepages mention Evolutions 81, 82, 89 and 91 only. Results from E98 to E100 were published only to the proof-factory lane and the clawx evidence directory; no homepage text was updated after E91.

## E100 internal inconsistency
`docs/evolution-100/E100_ACCEPTANCE_MATRIX.json` (zero-shot 99/100, few-shot 52/100, skill 21/50, F1 0.6452) comes from an earlier run than `E100_REPORT.md` and the published bundle (53/60, 19/60, 45/50, 0.6667). The matrix was never regenerated. The summary also labels the delayed family `BLOCKED_GENERATOR_DID_NOT_YIELD_SOLVABLE_SET` while reporting per-family results for it (family 6 is the delayed family).

## Defects found (affect E99, E100, E101)
1. **Oracle contamination.** `make_family_envs` / `make_solvable_envs` run the solvability oracle on the real environment, so every environment handed to an agent is already in the oracle's end state (all 60 of 60 in the E100 sealed set start at step 1-10).
2. **Shared environments across arms.** E100 evaluates GRU, then MLP, then random, then few-shot on the same list of environment objects; each arm starts where the previous one stopped.
3. **The oracle's beam search is dead code.** It computes a 3-step beam and then takes a one-step greedy action with true dynamics, so "solvable" means "solvable by a greedy planner that knows the rule". This biases sealed sets toward greedy-friendly tasks.
4. **The E100 verifier does not recompute from raw rows** (its own comment says so), and several checks restate the result ("few-shot negative vs zero-shot", "MLP within 5 of GRU").
5. **E101 natural-failure F1 0.9831 is degenerate.** 29/30 rows predict, and observe, the first failure at step 0, so a constant predictor scores the same.
6. **E98-E101 governance is in-process.** The four validator keys live in the same Python process as the agent, the grant is issued by the wrapper the agent calls, and the kill switch is a boolean flag on that wrapper.

## Reproduction (committed E100 models, sealed seeds 130000+, n = 60; `artifacts/e102/phase0/oracle_contamination.json`)
| arm | E100 exact order (defects 1+2) | defect 1 only | corrected (fresh env per arm, oracle on a copy) |
|---|---|---|---|
| GRU (recurrent) | 53 | 53 | **53** |
| MLP (non-recurrent) | 48 | 55 | **57** |
| random | 21 | 28 | **4** |
| few-shot GRU | 19 | - | **56** |

- The exact E100 numbers (53/48/21/19) are reproduced, so the defects are confirmed as their cause.
- E100's "few-shot adaptation hurt by -56.7 pp" (a FAIL gate) is an artifact. Corrected few-shot is 56/60 against 53/60 zero-shot (+3, one seed, not significant).
- E100's "recurrent beats non-recurrent" is **not supported**: corrected MLP 57/60 is at least as good as GRU 53/60.
- The GRU's advantage over random is real and larger than published (53 against 4).

Historic E99-E101 evidence is preserved unchanged. These corrections are additive records, not rewrites.

## E99 re-score (`artifacts/e102/phase0/e99_rescore.json`, `scripts/evolution102/phase0_e99_rescore.py`)
E99 has the same generator defect and evaluated GRU-greedy, GRU-CEM, MLP-greedy, MLP-CEM and random one after another on the same environment objects. The MLP model was not saved, so both models were retrained with E99's exact settings on corrected data.

| greedy, sets 1/2/3 (n = 100 each) | published | exact E99 order, committed GRU | corrected |
|---|---|---|---|
| GRU, committed model | 93/84/88 | 93/84/88 (300/300 pre-played) | **88/86/86** |
| GRU, retrained on corrected data | - | - | 88/87/86 |
| MLP, retrained on corrected data | 68/63/70 | - | **87/86/86** |
| random | 20/24/23 | - | 8/16/12 |

E99's world model is real and strong against random. Its headline ("the recurrent model beats the non-recurrent model by ~20 pp on every set") does not survive: scored correctly, the two are tied.
