# Evolution 80 — Reproduction of the Evolution 79 Results

Reproduced 2026-11-15 (UTC) from commit `f4449aa7913623bc034fa4434ca96c2a2919a889` in a clean
worktree. No reported number was accepted before it was reproduced.

| Reported E79 result | How it was reproduced | Result | Class |
|---|---|---|---|
| 144/144 offline checks | `python3 scripts/evolution79/verify_e79_v2_offline.py artifacts/e79 --artifacts artifacts/e79` in the clean worktree | 144 passed / 144; Merkle root `d42ff7744cfb9ab66149f8d891cac771bcfa950dd53b924f0e48133a18aadcc8` | PASS |
| Clean-room verification | Bundle copied to an empty directory; `python3 -I verify_e79_v2_offline.py . --artifacts .` (isolated mode, no CAIN imports) | 144 / 144, same Merkle root | PASS |
| Second-host verification | Bundle copied to `45.76.169.191`; same command | 144 / 144, same Merkle root (2026-11-15T16:51:50Z) | PASS — **same administrative domain, not an independent witness** |
| Raw benchmark reconstruction | Independent recount of `artifacts/e79/bench/*.jsonl` (not using the verifier) | A 10/104, B 35/104, C 12/32, D 8/32, E 10/32, SC2 11/32, T 14/104; replication A 5/32, B 11/32, T 3/32 | PASS |
| 986 calls, 167,172 tokens | Count of rows and token sum in `bench/model_calls.jsonl` | 986 rows; 167,172 tokens | PASS |
| Single agent 34.4% on the 32-task subset | B on the subset = 11/32 | 0.3438 (the full-set figure is 35/104 = 33.7%) | PASS |
| Adversarial results 57/57 | Value in the signed security lab, checked by the verifier (`security:bypassed_count`) | 57 blocked, 0 bypassed | PASS (record reproduced; the lab was not re-attacked) |
| Worker recovery | Signed drill record: 8 OS worker processes, 39 completions, 0 duplicate effects | matches | PASS (record reproduced; the drill was not re-run) |
| Research replication | Signed record: seed 791, 32 tasks, 64 calls | matches; but see F80-P0-2: 14 questions overlap the main set | PARTIAL |
| E79 test suites | `pytest tests/test_cain2027_e79.py tests/test_cain2027_e79_v2.py` | **25 passed**, 0 failed (265.6 s; includes one real local inference) | PASS |

## What "reproduced" means here

- The benchmark **numbers** were recomputed from raw rows three ways (verifier, clean room, an
  independent recount). The model calls themselves were **not** re-run: re-running 986 CPU
  inferences would only reproduce a deterministic record at a cost of about 100 minutes of shared
  production CPU.
- The security lab and the worker-kill drill were verified as signed records, not re-executed.
  Evolution 80 runs its own lab and its own drill.

## Discrepancies found while reproducing

1. **Replication overlap** (F80-P0-2): 14 of the replication questions also appear in the main set.
2. **Scorecard memory value** (F80-P0-3): the composite uses 0.0 → 0.0 where the raw rows say
   0.10 → 0.70.
3. **Two public E79 blocks** (F80-P0-1): see `PUBLIC_EVIDENCE_RECONCILIATION.md`.

None of these changes a raw measurement. All three change how the E79 result should be described.
