# E126.a Baseline Audit

**Date:** 2026-11-21 · **Entity:** FIRME AI CORPORATION · **System:** CAIN-42
**AGI status:** NOT ACHIEVED

## 1. Repository and deployment map

| Surface | Source file | Served by |
|---|---|---|
| CAIN Studio `cainstudio.online` | `platform-gateway/frontend/index.html` | platform gateway / Traefik |
| MCPGate `mcpgate.online` | `platform-gateway/frontend/mcpgate_home.html` | platform gateway / Traefik |
| ClawX `clawx.click` | `clawx-site/index.html` | ClawX site server / Traefik |

Evidence mirrors are written to `platform-gateway/frontend/proof/bundle/`, `cainstudio-homepage/evidence/`, `clawx-site/evidence/` and `docs/`. **No live deployment was performed in this session**; deployment is BLOCKED.

## 2. Current first-screen architecture

Before E126.a, all three pages carried a generic E126 section plus older E125/E124 panels. The
three first screens were effectively identical, which violates the one-ecosystem, three-purpose
requirement. The older panels are preserved below the new first screens.

## 3. E125 status

- Final commit `4df98c0ea`; prereg commit `be3494d96`.
- Offline verifier **81/81**; tests 10/10.
- Gates **15 PASS / 7 PARTIAL / 1 FAIL / 2 BLOCKED**.
- B+C 112/120 vs 98/120, exact p=0.00131, Holm 0.00262.
- Category E **0.000**; continual **0.6905 vs 0.6857, p=0.77**; transfer no accuracy advantage;
  adaptation catastrophic forgetting; causal active tied random; mission 301.77 s.

## 4. E126 status

E126 is **PARTIAL**, not a success. Its evidence is available and is integrated with its
limitations:

- Novel environments **87.0% vs 78.0% retrieval (p=0.0636, H1 NOT met)**, but **87.0% vs 61.0%
  E125 engine (p=3e-8)**.
- Continual learning **69.6% vs 83.3% no-memory — FAIL**.
- Causal active **0.0127 MAE vs random 0.0002 — FAIL**; counterfactual implemented (0.0191 vs
  passive 0.1120) — PASS.
- Transfer **+12.5 points on applied pathways, 0 false accepts** — PASS.
- Collective **69.0% vs 51.4% best single, 0/500 Byzantine flips** — PASS.
- Discovery **8 cycles, 2 independent reproductions** — PASS.
- Self-improvement **40 evaluated, 0 promoted, 4 refused** — PASS.
- Adaptation target **0.5417 vs frozen 0.25**, but control **0.375 → 0.0** — FAIL on retention.
- Mission **~15 minutes**, not six hours — PARTIAL/RUNNING at registry build.
- Governance 3 processes, **0 unauthorized, 0 post-fence** — PASS on one host.

## 5. Evidence inventory and claim mapping

The canonical registry is `cain/evidence/public_registry/e126a_registry.json` (20 records).
Every record has a stable claim ID, scientific status, integrity status, raw paths and SHA-256,
bundle URL, verifier, preregistration, sample size, limitations and supersession status.

## 6. Contradictions and stale claims

| Claim | Audit |
|---|---|
| SWE-bench 189/486 vs 257/500 | E124.1 189/486 is a **stale 05:14 snapshot**; current official-report union is 257/500. Marked INTERIM, not leaderboard. |
| 12-hour mission | Elapsed time is supported by the journal; productive diversity is PARTIAL. E123 G11 BLOCKED predates the run. |
| 48-hour soak | The 2026-11-15 PBFT safety incident and quarantine mean it must not be claimed as passed. Marked FAIL/SUPERSEDED. |
| "AGI infrastructure" / "Proto-VAGI" | Must always carry **AGI NOT ACHIEVED**. E126.a does not upgrade the status. |
| Verifier success = scientific success | Explicitly separated: `scientific_status` and `integrity_status`. |
| E126 = success | E126 is PARTIAL. Persistent learning and active causal selection failed. |

## 7. Broken or missing links

The old generic E126 section linked to a bundle that was not yet built. E126.a links point to the
E126 and E126.a bundles after they are built; the verifier checks the local bundle manifest and
the registry paths. Live external links were not fetched (browser disconnected).

## 8. Mobile and desktop

Static responsive checks pass: `clamp()` typography, `auto-fit` grids, `overflow-x:auto` tables,
single-column mobile fallback, semantic headings and aria labels. Browser/Lighthouse screenshots
could not be captured because no desktop browser is connected; **G18/G19/G20/G28 are BLOCKED or
PARTIAL**, not PASS.

## 9. Implementation

- `cain/evidence/public_registry/` canonical schema and registry.
- `scripts/e126a/build_registry.py` and `build_frontpages.py`.
- Three distinct first screens, one per site.
- `tests/test_e126a_frontpages.py` (12 passed, 1 skipped pending bundle).
- E126.a offline verifier and evidence bundle.

## 10. Risks and rollback

- Risk: editing production homepages can break layout. Rollback: `git checkout --` the three
  homepage files; the generator is idempotent and preserves the old panels below.
- Risk: a stale registry could publish a wrong metric. Mitigation: registry is generated from
  raw result files; verifier recomputes hashes and key metrics.
- Risk: the 20-record registry is larger than first-screen space. Mitigation: first screens show
  priority records; full registry is linked.
