# E92 Correction — 2026-11-18

**Corrected status: NOT RELEASED. PROCEDURE GENERATION NOT DEMONSTRATED.**
Gates 3, 4 and 5 are now **FAIL**. This was the owner's decision on 2026-11-18, following a code review.
The original E92 reports, rows and signed bundle are kept unchanged as the historical record. This document supersedes their conclusion.

## What E92 published

"RELEASED — PARTIAL EVIDENCE"; gates 3–5 PASS; G2 120/120 vs D/G/F 70/120 (p≈0); replication 120/120 vs 64/120;
"3 procedures generated, 0 designer-supplied".

## Why that conclusion does not hold

1. **The designer supplied the solutions.** `cain/cain2027/e92/synthesis.py::candidates_for` enumerates four
   hand-written program templates: `majority3`, `probe_argmax(n)`, `check_branch(a,b)` and `retry(a,b)`.
   Each one is the complete solution workflow for exactly one of the three families in `world.py`.
   The family's tool list (`read` / `probe` / `check+apply+verify`) picks its template one-to-one.
   The search only chooses 0–2 integer parameters within the correct template, and the space holds 12 candidates in total.
   The E92 brief prohibits exactly this ("Do not allow a designer to supply the solution workflow in advance").
   Calling these "0 designer-supplied" was wrong: the solutions were supplied as schemas.
2. **D, G and F are one program.** `experiment.py` runs `world.default_program(family)` for all three arms. That program returns
   source 0, returns index 0, or applies action 0, so the three arms score the same 70/120.
   - G has no repertoire, so "synthesis beats selection" (gate 4) could not fail.
   - F was never fitted on development data, so it is not a "best fixed strategy".
3. **The task prompts state the method.** For example, `reconcile` says "Return the majority value across the three sources."
4. **No model-proposed candidates and no counterexample-guided revision.** The adapter returns BLOCKED. No candidate was ever revised
   from a counterexample; candidates were only filtered.

## What E92 does show (unchanged)

- A typed JSON-AST DSL with a bounded interpreter, tool allow-lists and call budgets.
- A procedure store with lifecycle, provenance, and a distinct approver identity (Ed25519). Self-promotion is refused.
- The 15/15 in-process governance lab.
- Recomputation of the E91 baseline from raw rows (D 114/160, G 153/160).
- A signed bundle that checks its own integrity. Integrity is not the same as a valid experiment.

## Corrected gates

| # | Gate | Published | Corrected |
|---|---|---|---|
| 3 | Procedure synthesis without supplied solution | PASS | **FAIL** — templates are family solutions |
| 4 | Generation beats selection | PASS | **FAIL** — G is a no-op copy of D |
| 5 | Generation beats D and F | PASS | **FAIL** — D/F are an untuned strawman; the gain measures the templates |
| 16 | Claim discipline | PASS | **FAIL** — the public pages claimed synthesis was demonstrated |

The other gates are unchanged. The published header (8/3/5/0) also disagreed with its own gate list (9 PASS / 3 PARTIAL / 4 BLOCKED). Corrected counts: **PASS 5 / PARTIAL 3 / BLOCKED 4 / FAIL 4**.

## Source exposure

The public bundles at `clawx.click/evidence/e92-procedure-synthesis-2026-11-18/` and
`cainstudio.online/proof/bundle/e92-procedure-synthesis-2026-11-18/` shipped the `cain/cain2027/e92/*.py`
source and `tests/test_e92.py`, which goes against the owner's code-stays-secret directive. Those files were removed from both public copies.
Their SHA-256 digests remain in the signed manifest, so the offline verifier now reports them as missing. That is expected.

Measured after removal (2026-11-18): `python3 -I verify_e92_offline.py .` → **68/77**. The signature verifies, and the only failures are the 9 withdrawn files (8 under `cain/cain2027/e92/` and `tests/test_e92.py`).

## What a real test requires (carried into E93)

- An open search over DSL primitives with no per-family templates and no prompt hints.
- A G arm with a real repertoire of supplied procedures, and an F arm fitted on development data and frozen.
- Task families written separately from the search, including families the search could plausibly fail.
- The kill criterion applied as written.
