# E82 Assurance Report — beyond hashes

This report covers the assurance work added after the initial E82 evidence bundle:

- **Negative controls** — remove one security check, prove the attack then executes.
- **Property-based tests** — Hypothesis-generated attacks on the Gate.
- **Formal model** — Z3 proves 15 properties of the gate decision model.
- **Repeated host reboots** — three real reboot cycles with an active mission.
- **6-hour endurance** — started; **not complete at the time of this report**.

Every artifact is stored in this worktree under `artifacts/e82/` and published under the public
proof directory. This remains same-project evidence; no third-party audit or witness exists.

## 1. Negative controls — 9/9 valid

For each control the same attack is run twice: once against the real `Gate` (must be blocked) and
once against a `WeakenedGate` with exactly one check removed (must execute). A control is valid only
when `original_blocked=true` **and** `weakened_executed=true`.

| Check removed | Original result | Weakened result |
|---|---|---|
| Signature | `GRANT_SIGNATURE_INVALID` | executed |
| Signed DENY decision | `NOT_ALLOW` | executed |
| Expiry | `GRANT_EXPIRED` | executed |
| Revocation | `IDENTITY_REVOKED` | executed |
| Replay | `GRANT_REPLAYED` | executed |
| Fencing | `STALE_FENCING` | executed |
| Tool substitution | `GRANT_TOOL_MISMATCH` | executed |
| Scope substitution | `GRANT_SCOPE_MISMATCH` | executed |
| Budget exhaustion | `NO_BUDGET` | executed |

Result: **9/9 valid**, 0 invalid. This proves each check is load-bearing for its attack. It does
**not** prove that no other bypass exists.

Artifact: `artifacts/e82/negative-controls.json` · harness:
`scripts/evolution82/negative_controls.py`.

## 2. Property-based tests — 3,000 generated attacks

`tests/test_e82_properties.py` (pytest + Hypothesis 6.169.0):

- 1,000 random scope substitutions → denied
- 1,000 random negative TTLs → expired
- 1,000 random signature tamper strings → denied
- valid grant executes exactly once
- revocation, fencing, agent, tool and budget checks are load-bearing

Result: **5 passed**, **3,000 generated cases**, 0 executions. Artifact:
`artifacts/e82/property_tests.json`.

Limit: it tests the Gate, not the MCP transport or the host kernel.

## 3. Formal model — 15/15 proved

`verification/e82_formal/gate_model.py` encodes the gate conditions in Z3 and proves the negation
unsatisfiable for each:

- no signature, non-ALLOW decision, agent/tool/scope mismatch
- expiry, stale fencing, revocation, cross-tenant
- replay, budget exhaustion, kill switch
- unknown or error
- single-use executes at most once
- child scope expansion does not execute

Result: **15/15 PROVED**. Artifact: `artifacts/e82/formal_model.json`.

**Limit:** this is a proof about the model, not a refinement proof of the Python implementation.
The negative controls link each model condition to the code by removing it and showing the attack
succeeds.

## 4. Repeated host reboots — 3/3 pass

Three independent cycles on the development replica, each with a **fresh** state file:

| Cycle | Boot id before | Boot id after | Downtime | Tasks | Duplicates | Stale commit |
|---|---|---|---|---|---|---|
| 1 | `4a5b981b…` | `eb40c04f…` | 38.568 s | 20/20 | 0 | STALE |
| 2 | `8261f085…` | `1e5ff28a…` | 37.615 s | 20/20 | 0 | STALE |
| 3 | `1e5ff28a…` | `97e9a0d3…` | 37.477 s | 20/20 | 0 | STALE |

Durable identity survived each cycle; checkpoint chain valid; event prefix preserved.

**Defect found and fixed during this work:** the first attempt at cycle 2 read a stale
`state-before.json` and rebooted before the new mission started. That cycle was **discarded**. The
harness now deletes the state file, records `t_start`, and refuses any state older than ten seconds.
A second defect (stale SQLite WAL/SHM after a prior cycle) caused a disk I/O error; the mission
script now removes `mission.db`, `-wal` and `-shm` before creating a fresh store.

Artifact: `artifacts/e82/host_reboot_repeated.json` plus `host_reboot/`, `host_reboot_2/`,
`host_reboot_3/`.

## 5. 6-hour endurance — RUNNING, not complete

A real wall-clock campaign was started with `run_endurance.py 21600 artifacts/e82/endurance-6h`.
At the time of this report it is **RUNNING** (not completed). The status file is updated every ten
cycles. Do not describe this as a completed 6-hour campaign until `endurance.json` says
`COMPLETED` and `target_reached=true`.

Artifact: `artifacts/e82/endurance-6h/status.json` (live).

## What this adds

| Claim | Evidence class | Status |
|---|---|---|
| Each of nine security checks is load-bearing | same-project fault injection | 9/9 |
| Random attacks do not produce execution | same-project property test | 3,000 cases, 0 bypass |
| Gate conditions are necessary in the model | formal model proof | 15/15 |
| Active mission survives real reboots repeatedly | L1 cross-host | 3/3 |
| Six-hour endurance | not completed | RUNNING |

## What it still does not add

- No third-party audit.
- No independent witness.
- No external reproduction attestation.
- No remote model call.
- No customer production deployment.
- No AGI claim; “governor” remains an architectural role.
