# Evolution 80 — Public Evidence Reconciliation (Evolution 79)

Inspected 2026-11-15 (UTC): the repository copies of the three homepages at commit `f4449aa79`,
the live `clawx.click` homepage and evolution history, and the E79 artifacts. `cainstudio.online/`
and `mcpgate.online/` timed out when fetched (see the baseline audit), so their live HTML could
not be read; the repository copies they are served from were read instead.
`audit.cainstudio.online` does not resolve.

## Verdict

**The conflicting numbers are two separate, real experiments, both run on 2026-11-15, and both are
published on the same pages without saying which one is current.** Neither set is fabricated.
Neither was deleted. The second run (v2) is the accepted Evolution 79 result; the first run (v1) is
preserved under `artifacts/e79/v1/` and is superseded.

| Reported conflict | v1 — first run (superseded) | v2 — accepted run | Explanation |
|---|---|---|---|
| Transformer calls | **1,186** | **986** | Different experiments. v1: 1,000 benchmark calls + 186 other calls. v2: 986 attributed calls. |
| Tokens | **115,622** | **167,172** | v2 prompts are longer (eight domains, code and tables). |
| Benchmark size | **100** synthetic arithmetic tasks | **104** tasks in 8 domains; collaboration arms on a **32**-task subset | Different datasets. "32-task" is the v2 subset, "100-task" is v1. |
| Single-agent score | **1%** (direct answer, 1/100) | **34.4%** (11/32; 35/104 = 33.7% on the full set) | Different tasks and prompts. The v1 direct-answer prompt almost never produced a parseable answer. |
| Four-agent score | **29%** (29/100, four expressions + calculator) | **25%** (8/32, four specialists) | Different arms on different tasks. v1's "four agents beat one" compared against a 1% baseline. |
| Grade | **B-** (composite 7.04) | **C-** (composite 5.78) | Different scorecards. v2 is stricter. v2's composite also contains a stale memory value (F80-P0-3). |
| Verification checks | **22/22** (20-field bundle, v1 verifier) | **144/144** (22-field bundle, v2 verifier that recomputes from raw rows) | Different verifiers on different bundles. Both verify. |

## Why the v1 conclusion was replaced

v1 reported that four agents (29%) beat one agent (1%). That comparison was against a single-agent
prompt that failed to produce answers, so it measured a formatting failure, not collaboration. v2
fixed the single-agent arm (34.4%) and found **no significant gain from collaboration**. v2 is the
honest result and is the one Evolution 80 builds on.

## Where each version is published

| Surface | v1 numbers | v2 numbers | Problem |
|---|---|---|---|
| `clawx-site/index.html` (live at `clawx.click/`, fetched) | yes (banner `#cain2027-e79-intro-banner`, labelled "EVOLUTION 79 LIVE") | yes (top block) | Two blocks, both called Evolution 79, the older one labelled LIVE |
| `cainstudio-homepage/index.html` | yes | yes | same |
| `platform-gateway/frontend/index.html` | yes | yes | same |
| `clawx-site/evolution-history.html` (live, fetched) | 1 mention | 1–2 mentions | both present |
| `clawx-site/evidence/evolution-79-collective-intelligence-2026-11-15/` | — | v2 bundle | consistent |
| `integrations/huggingface-space/.../evolution-79-real-intelligence-2026-11-15/` | v1 verifier | — | v1 only |

## Corrections made (branch `evolution-80` only — NOT deployed)

One label was changed in each of the three homepage source files. No number was changed and no
block was removed.

| File | Before | After | SHA-256 before → after (first 16) |
|---|---|---|---|
| `clawx-site/index.html` | `EVOLUTION 79 LIVE` | `EVOLUTION 79 · FIRST RUN (v1) · SUPERSEDED BY THE v2 RUN ABOVE — KEPT AS HISTORY` | `7d8a6785a2c71ec1` → `3d6faed9ae784840` |
| `cainstudio-homepage/index.html` | same | same | `14d933a5edbeb6ce` → `b4c024180e8e4535` |
| `platform-gateway/frontend/index.html` | same | same | `071bb3a5437b8c76` → `8d0aa2999fb591a7` |

The live sites still show the old label. Publishing the correction is a production change and
needs the owner's authorization.

## Corrections NOT made (recommended to the owner)

1. The Hugging Face Space still ships only the v1 verifier under an "evolution-79" path. Add a
   "superseded by v2" note next to it.
2. `clawx-site/evolution-history.html` should mark the v1 line as superseded.
3. The E79 scorecard memory value (F80-P0-3) should be recomputed in a new, clearly versioned
   scorecard. The signed v2 bundle must not be edited.
4. The E79 replication should be described as "replication on a partially overlapping set
   (14 shared questions)" (F80-P0-2).

## Canonical evidence index

Machine-readable: `artifacts/e80/canonical_evidence_index.json`. It records, for every E79 and E80
experiment: evolution, full commit, experiment id, dataset identity, benchmark version, model and
revision (weight-file SHA-256), seeds, raw artifact hashes, verifier, verification result,
publication status and supersession.
