# E110 baseline audit (Part I)

Prepared 2026-11-19 by session b5 (infrastructure, evidence, public surfaces) in coordination with session c1
(organism). FIRME AI CORPORATION, a three-person team. Read-only audit: nothing in E100-E109 was modified.

- **Start commit (infra line):** `4ad5e3e345528bad1ebab924cedf3cd0a2655ce3` (E108 sealed results), branch
  `evolution-110-infra`, worktree `/root/cain-e110-infra`.
- **Organism line (c1):** branch `evolution-110-organism` from `42bbff35a` (E108 + founder-only names fix).
- **Status vocabulary:** PASS / PARTIAL / FAIL / BLOCKED / NOT TESTED.

## 1. What actually exists, E100-E109

| Evolution | Branch head | Public bundle | Verifier re-run 2026-11-19 | Status | Notes |
|---|---|---|---|---|---|
| E100 integrated organism | `a33b5fab1` | e100-integrated-organism | 1/1 | CONTEXT | Headline claims corrected by E102 Phase 0 |
| E101 generalization/adaptation | `cb848cebd` | none published | — | NOT TESTED here | Few-shot adaptation negative (per AGENTS.md) |
| E102 persistent organism | `967f08c68` | e102-persistent-organism | 1/1 | RECEIPTS | Loopback govnet 27/27, KILL 0 effects; 6 h mission **not finished** (0.5545 h) |
| E103 transfer and skills | `0d2c3b78d` | e103-transfer-and-skills | 1/1 | NEGATIVE | Skills failed sealed test |
| E104 strategy switch | `a2cee0f92` | e104-precondition-discovery | 1/1 | NEGATIVE | Strategy switch failed sealed test |
| E105 hidden preconditions | `55be10303` | e105-precondition-location | 1/1 | MIXED | Learned (33.0% lift on 10% cells), not yet useful (24 vs 27) |
| E106 invalid gate | `17943a474` | none (stopped at development) | — | NEGATIVE | Permutation gate inverted; nothing claimed |
| E107 autonomous experimentation | `bc7efdb8d` | e107-organism-experiments | 1/1 | SUPPORTS | edge-lock 91 vs 64 of 200 (27 won, 0 lost); no per-action receipts |
| E108 multi-rule discovery | `4ad5e3e34` | e108-hidden-rules | 1/1 | SUPPORTS | 4 laws, 100 incidents won / 0 lost; no per-action receipts |
| **E109 open-world causal discovery** | `evolution-109` = `4ad5e3e34` (no commits) | none | — | **NOT COMMITTED** | See section 2 |

All published verifier results come from `clawx.click/evidence/vagi-audit-index-2026-11-19/` (each bundle's own verifier
re-run on a fresh copy; manifest SHA-256 pinned there).

## 2. E109: actual state

- The `evolution-109` branch has **no commits of its own**; it points at the E108 head.
- An **uncommitted** E109 ("open-world causal discovery") exists in `/root/cain-e109`: `cain/cain2027/e109/`
  (engine, grammar, memory, metacognition, collective, governance, transfer), `scripts/evolution109/`,
  `docs/evolution-109/`, `tests/test_e109_core.py`, `artifacts/e109/`. Files are dated 14:08-15:16 UTC on 2026-11-19.
- **Author unknown.** It is not session b5's or c1's work (the only two Claude sessions listed). Flagged to the owner.
- Its own `summary.json` (uncommitted, unverified, quoted only for the audit): e109 arm **68/80** (1 false discovery),
  e108 arm 69/80, fixed baseline **73/80**. **No gain over baselines.** E110 credits nothing to E109 unless its owner
  commits, publishes and verifies it.

## 3. Dependency map: what participates in a mission

Import graph (AST, including relative imports) of `cain/cain2027`:

```
E98 → E99 → E100 → E101                      (integrated-organism line; ends at E101)
E102 → E103 → E104 → E105 → E106             (persistent-organism line)
E102 → E107
E102 + E104 + E105 → E108                     (current head)
```

- **The two lines never join.** E102-E108 import nothing from E100/E101.
- **The E102 line imports no other CAIN subsystem.** It does not use `cain/mcp_proxy.py` (MCPGate), the AOS mission
  runtime (`cain/aos/`, E63-E75), the distributed workforce (`cain_workforce/`, E68-E72), the collective
  (`cain_collective/`, E69), AETHER-33 memory, or the PBFT cluster.
- **The organism core that actually runs a mission today** is `cain/cain2027/e102/`:
  `organism.py` (Percept, WorldModel with particles, plan, Organism), `cognition.py` (CausalMemory, Metacognition,
  Brier/ECE), `govnet.py` (CainService decision + WorldService resource + signed hash-chained log; loopback),
  `worlds.py` (constructed grid worlds), `harness.py` (Wilson, McNemar, sign test), `endurance.py` (journaled mission
  with resume); plus `e107/experiment.py` and `e108/{rules,gate}.py`.

### Spec component → reality

| Spec component | Exists in repo | In the organism's mission path today |
|---|---|---|
| Cognitive kernel / world model | e102 WorldModel (particle beliefs) | YES |
| Causal reasoning | e102 CausalMemory, e107 interventional attribution, e108 rule gate | YES |
| Episodic / semantic / procedural memory | AETHER-33, cain/aos memory, e102 CausalMemory | PARTIAL: only e102 CausalMemory (file-backed, chained) |
| Mission engine / goal hierarchy | cain/aos (E63-E75) mission graph | NO (disconnected) |
| Strategic planning | e102 `plan` (single level) | PARTIAL: no Mission→…→Action hierarchy |
| Skill synthesis / validation | e103 (failed sealed), cain/aos/v75 | NO (e103 negative) |
| Autonomous research / experiment selection | e107 practice-mode planner, e108 | YES (constructed worlds) |
| Metacognition / reflection | e102 Metacognition (calibration) | YES (calibration only) |
| Multi-agent orchestration | cain_workforce, cain_collective | NO (disconnected) |
| Identity and authority | e102 govnet keys + permits | YES (loopback) |
| BFT / consensus | PBFT clusters (cain-mr-*) | NO (not in mission path) |
| MCPGate execution | cain/mcp_proxy.py | NO (the organism uses e102 govnet, not MCPGate) |
| Kill switch | e102 govnet `/kill` | YES (loopback) |
| ClawX evidence | published bundles | YES (after the run, not streamed) |
| Public research APIs | cain/argus, /api/v1/* | NO (not in mission path) |

**Consequence for E110:** G1 ("components operate together in one mission") starts at **PARTIAL**. The honest E110
work is to connect memory, hierarchical planning, the collective and MCPGate-grade enforcement to the e102 organism,
not to add new disconnected modules.

## 4. Tests and reproduction

- Tests present in the tree for this line: `tests/test_e100_organism.py`, `tests/test_e107_experiment.py`:
  **6/6 pass** (clean worktree at the start commit, 4 GB memory limit).
- E102-E106 and E108 have **no unit tests in the tree**; they are evidenced by their published bundles and verifiers.
- Clean-checkout verifier re-runs: see section 1 (17 of 19 bundles with a verifier pass; E79 fails 2 hash checks by
  design after a disclosed IP redaction; E92 is a corrected bundle whose own verifier fails 68/77).

## 5. Historical limitations that must not be silently upgraded

- **Constructed benchmarks:** E95, E100-E108 run on constructed task worlds, not real-world tasks.
- **Loopback, one host:** e102 govnet is three OS processes on one host; not the production MCPGate; not independent
  failure domains.
- **Signed ≠ true:** every bundle key is self-generated; a valid signature proves integrity, not truth.
- **Endurance:** no 6 h governed mission has completed (E102: 0.5545 h). The 72 h multi-region soak was never
  established; the 2026-11-15 PBFT safety incident (conflicting certificates on cain-mr-02) and the 2026-11-18 view
  incident on mr-03 stand.
- **Corrections on record:** E92 (not demonstrated), E99/E100 headline claims corrected by E102 Phase 0.
- **No external reviewer, no third-party audit, no external AGI benchmark.**
- **VAGI Agents (12):** deterministic rule engines with signed records. They are governance products and **do not count
  as vertical intelligence** for E110's Part IV.
- **Receipts scope:** E107/E108 sealed runs called the world directly (no per-action receipts). Receipts exist for the
  E102 boundary and the VAGI Agents only.

## 6. Public files that name team members other than the founder

Owner directive: public pages name only the founder; otherwise "a three-person team". Still naming the other two:

- Pages (not hashed; can be fixed with owner approval): `founder.html`, `changelog.html`, `changelog-2026-10-05.html`,
  `changelog-e18-e28.html`, `evolution-79.html` (each on both cainstudio and clawx).
- **15 files inside hashed evidence bundles**: editing them breaks their verifiers. Owner decision needed (withdraw and
  re-issue with a disclosed redaction, as done for E79's host address, or leave as historical record).
- Fixed today: the Hugging Face Space front page; E107/E108 bundles (by c1, re-pinned).

## 7. Gate baseline (before E110 work)

| Gate | Baseline | Basis |
|---|---|---|
| G1 unified architecture | PARTIAL | e102 core runs end-to-end; memory/planning/collective/MCPGate disconnected |
| G2 held-out performance | NOT TESTED | no E110 sealed challenge yet |
| G3 continual learning | PARTIAL | E95 harness improvement; E105 learned-not-useful |
| G4 causal reasoning | PARTIAL | E107/E108 sealed wins on constructed rule worlds |
| G5 skill acquisition (≥5) | FAIL | E103 skills failed sealed test |
| G6 transfer (≥3 of 5) | NOT TESTED | no five-pathway study |
| G7 metacognition | PARTIAL | e102 Brier/ECE; no abstention gate |
| G8 6 h governed mission | NOT TESTED → running | E110 run-1 started 15:38 UTC (see LONG_HORIZON when done) |
| G9 collective value | FAIL | E79: no significant gain from collaboration |
| G10 recursive improvement | PARTIAL | E95 (one harness change, held-out) |
| G11 resource-side enforcement | PARTIAL | e102 loopback 27/27; not MCPGate, not separate domains |
| G12 kill switch | PARTIAL | e102 loopback: 0 effects after KILL |
| G13 identity | PARTIAL | self-generated Ed25519 keys; no external issuer |
| G14 evidence | PARTIAL | bundles per evolution; not one mission ledger |
| G15 memory integrity | NOT TESTED | for the organism's memory |
| G16 independent verification | PARTIAL | clean-copy verifiers; no external reviewer |
| G17 endurance | NOT TESTED → running | 6 h run-1 in progress; 24 h / 48 h NOT TESTED |
| G18 operational independence | FAIL | one host, one admin domain for the organism path |
