# Evolution 85 — Baseline Audit

**FIRME AI CORPORATION · CAIN-42 · Evolution 85** — branch `evolution-85`, baseline Evolution 84 (`2c419a042`). Development result. Not deployed. Not released. Built by the three-person team.

## What was checked before anything was built

| Check | Result |
|---|---|
| Evolution 84 commit `2c419a042` is an ancestor of this branch | True |
| Evolution 84 tests | 6 passed, 0 failed |
| Evolution 84 receipt verified from a clean copy with `python3 -I` | exit 0, 62 checks: `VERIFIED_COMPLETE  (exit 0)  effect=EFFECT_VERIFIED  witnesses=1 independent=0` |
| Tampered Evolution 84 receipt | rejected, exit 40 |
| Evolution 84 adversarial lab as recorded | 24 passed, 0 failed, 9 not run, of 33 |
| Evolution 84 acceptance as recorded | 27 PASS / 3 PARTIAL / 4 BLOCKED / 2 NOT EVALUABLE / 1 FAIL (G38) |
| Lines changed in reused code since Evolution 84 | 0 (`cain/receipt`, `cain/action_commit`, `cain/identity_federation`, `cain/mcp_proxy.py`, `scripts/evolution84`, `cain/cain2027/e81`, `cain/cain2027/e80`) |

Evolution 85 did not exist before this session: no branch, no worktree, no `docs/evolution-85`. The branch was created
from the Evolution 84 commit in its own worktree and the coordination file was written before any code.

## The gap Evolution 84 left open (its G38)

Evolution 84 reported one FAIL: no kill switch in the receipt path, and a process that writes the resource's SQLite
file directly never passes the enforcement point. Evolution 85 addresses both; see `KILL_SWITCH_ARCHITECTURE.md` and
gate `E84-G38` in `ACCEPTANCE_MATRIX.md`. The second half is only partly closed and is reported that way.

## Defects found and fixed while building

1. **Kill-switch cache.** SQLite's `data_version` does not move for a connection's own writes, so the process that
   engaged the switch kept reading its stale cache. Caught by the first scope test; the cache is now invalidated on
   every control write.
2. **Evaluation contamination.** The causal-reasoning family had only eight distinct prompts, so the same prompt
   appeared in the training and held-out splits. Caught by the split-disjointness test; event names are now drawn per
   task and the run refuses to start if any held-out prompt appears in train or validation.
3. **Contamination screen false positive.** The screen flagged the word "Override" in legitimate access-rule tasks
   and quarantined all of that family's verified examples. The pattern was narrowed to instruction-like phrases.
4. **Resource transport.** A dead resource process raised a raw `BrokenPipeError` instead of a refusal. It now fails
   closed as `RESOURCE_UNREACHABLE`.
5. **Kill benchmark.** The first stage finished its work before the kill landed, so it measured idle workers. Tasks
   were made heavy enough that the kill arrives mid-run, and the time of the engage transaction is now reported.
