# Eve delivery fault game — Task 11.5

**Admission date:** 2026-09-16  
**Contract:** `eve.delivery-fault-game.v1` with
`eve.delivery-fault-game-admission.v1`  
**Decision:** admitted for Task 11.5; Phase 11 and G12 remain open

## Decision

Task 11.5 is complete. The retained 2026-09-13 campaign observed every named
fault, preserved the operator-owned uncommitted bytes in every scenario, refused
success while the fault was active, and ended in a recovered, reverted, or exact
resumable state. Its twenty semantic and document controls each went red before
the final green regression.

The campaign was originally and correctly held behind Task 11.11. That
dependency completed later on 2026-09-13 with authenticated goal/plan revision
bindings, real PostgreSQL persistence and restart coverage, stale-work
invalidation, history preservation, tenant isolation, and independently
attributed confirmatory and adversarial QA. Admission therefore joins two
immutable snapshots instead of rewriting the earlier measurement:

- preparation commit `efffeb7263eae354ca6e8b2b26f37e6e9cbb9fc1`;
- Task 11.11 implementation commit `b6e39f097e81d869439ca4220ecfcc1400868459`
  and evidence snapshot `dbd1bfdd3a4cf6a8ed2a738f5fbe3289ef0b6332`.

The admission verifier loads bound preparation sources from the first Git
snapshot and Task 11.11 artifacts from its evidence snapshot. It fails closed if
either commit disappears, any digest changes, the dependency reopens, the ledger
checkbox reopens, a scenario/control count shrinks, or the limits are hidden.
This is deliberately different from resealing old evidence against the current
tree.

## Covered failures

The admitted matrix covers conflict, moving `origin/main`, failed hooks,
provider-sick/flaky tests, pre-existing failures, dependency outage, partial
commit, agent crash, budget and operator kill switches, revert, and exact-once
resume. The detailed injection and recovery facts remain in the original
[preparation audit](EVE_DELIVERY_FAULT_GAME_PREPARATION_2026-09.md) and retained
measurement.

## Exact limits

- Git faults used real isolated clones and bare remotes, not GitHub.
- Process faults used real Linux process groups, but deterministic workers
  rather than paid coding-model agents.
- Provider and dependency faults used deterministic doubles. No external
  provider was disrupted.
- Task 11.11 proves the persisted fleet binding at a local PostgreSQL and
  in-process Fastify boundary, not a production multi-region deployment.
- Completing Task 11.5 does not close Task 11.6, Task 11.7, Phase 11, or G12.