Website and repository reconstruction with contracts checked in a browser; 0 false completions.
Oracle7-Chameleon 2.0 (internal)
Status: public release. Prometheus7 Research Institute, 2026-09-25.
Source digest (SHA-256 over path + bytes of every .py/.mjs/.js/.ts file under packages/, scripts/ and
apps/api in the project-chameleon working tree, 348 files; node_modules, dist and caches excluded):
cb48b3e527b490b8f1d70071042914dc39b08f58eae45cb96bc11d6c71992822. The working tree is ahead of its last commit
(cac9772); the digest identifies the code that produced the evidence below.
Release digest (2026-10-13): 823787a94793d1d1fef99600339834e512b9be8a9fb1faae54b0011adb9262d5, recipe: for each of the
same 348 files in byte-sorted repo-relative path order, SHA-256 update with the UTF-8 path and then the file bytes (no
separators). The exact recipe of the 2026-09-25 digest above was not recorded and could not be reproduced; no covered file
has a modification time after 2026-09-25 12:00 UTC, so the code is the code that produced the evidence. SHA256SUMS in
the repository lists every file.
What it is
Chameleon is a website and component reconstruction engine run as an Oracle7 specialist.
- Its stores are captures and receipts.
- Its operators are MetaFold repair operations. Each operation acts only when its admissibility preconditions hold.
- A clone counts as COMPLETE only when contracts declared before execution pass in a supervised browser.
- The delivered package is re-verified by an independent audit.
trafficcop is vendored in as the decision service. It chooses repair operators and triages captures, and runs in shadow mode for captures.
Results
Repository reconstruction, release panel (results/release-panel-current-20260921b.json)
| Count | |
|---|---|
| Registered targets / families | 10 / 7 |
| Complete on first pass (no repair) | 3 |
| Complete at latest attempt (development repair allowed) | 7 |
| Held-out targets complete on first pass | 1 / 4 (25%) |
| False completions (COMPLETE with a failing contract or failed audit) | 0 |
| Release gate (24 targets across 6 families) | NOT ESTABLISHED |
The four held-out targets:
| Target | Outcome | Stage reached |
|---|---|---|
| robinparisi/tingle (modal) | COMPLETE | Fresh acquisition, then button-close, Escape-close and footer-cancel contracts all PASS; independent audit of all three 13-file packages PASS |
| autocomplete | Stopped honestly | Before behaviour, at REPOSITORY_PACKAGE_ENTRY_OR_LICENSE_MISSING |
| validator | Stopped honestly | Before behaviour, at REPOSITORY_LICENSE_FILE_REQUIRED_FOR_PACKAGE |
| Pickr (colour picker) | Behaviour verification FAILED | COUNT_ASSERTION_FAILED:0. The failure was reported, and no repair was counted toward the held-out score |
Evidence: results/metafold-tingle-heldout-first-pass-20260921.json and its audit.
Website quality (clone catalogue: 1,493 captures)
Current state: 33 approved, 31 needs_repair, 1,429 needs_review.
Website quality is scored by the same sealed 60-site blind panel as trafficcop (see the trafficcop card). Chameleon's figure from that panel is the clone approval rate: the fraction of the 60 random needs_review captures that Ben scores 4 (ship as is), with the mean score alongside. Reviewer check (2026-10-13): the project's owner looked through all 60 sampled sites and judged them good. This is a stated overall judgement, not the pre-registered blind 0–4 score per site, so no panel number is claimed and the scored panel stays open.
Failures, stated plainly
- The release gate is not met. It needs 24 targets across 6 families; 10 targets across 7 families are registered. A 1/4 held-out first-pass rate is an inventory, not a generalisation estimate.
- Held-out targets were chosen by the evaluator, who read their documentation to write contracts. That is source- informed testing, not externally blinded evaluation.
- Two of the three held-out non-completions stopped at licence and packaging admission, before any behaviour was tested. They measure acquisition breadth, not reconstruction ability.
- One held-out target (Pickr) failed behaviour verification. The cause, missing initialisation versus an unsupported transformation, has not been established.
- Recording trafficcop advice in the durable MetaFold engine is not done. That work is owned by Codex in
project-chameleon and logged in
docs/progress.md.
What the evidence supports
The completion discipline holds: no COMPLETE has ever been issued with a failing contract or a failed audit, and every stop names the stage and the reason. Reconstruction breadth is early.
Reproduce
In project-chameleon: scripts/clone_benchmark.mjs and the release-panel audit (see docs/BROAD_RECONSTRUCTION_EVALUATION.md,
docs/TINGLE_TRANSFER_FREEZE.md).
Files not included
18 deployment and demo scripts for the builders' own servers are not in this repository (they name private hosts). None is part of the method or of the evaluation that produced the numbers above. SHA-256 of the sorted list of their paths: 059eff217257a94a4435f9ca13d9e6da513a47b786067cf09056b6bcd472a1c1.