Post-reveal linkage, method, and route-status audit
Run run:manual:074ced121fdbd82f1b729113d9477151 · 2026-09-23T06:26:50Z–06:29:18Z UTC · public, lawful, primary/read-only sources. Direct inference/access requests: n=0.
Route result
| Route | Source evidence | Direct evidence | Conclusion |
|---|---|---|---|
stealth/union-alpha | Historical page says reveal = Pareto by Unbiased; free stealth period ended. Current catalog has no old ID. | 0 requests. | Historical/retired. Catalog omission is not an old-ID HTTP result. R4-S01 · R4-S02 |
unbiased/pareto | Current paid listing. Canonical slug unbiased/pareto-20260917; declared 262,144 context, 131,072 max output, text/image-to-text, tools/tool choice, and per_request_limits: null. | 0 requests. | Separate current route. Metadata does not prove entitlement, fixed preview equivalence, behavior, capacity, or quota. R4-S03 |
| Former Go alias | Not found in inspected current Go docs, public model catalog, or changelog. | 0 requests. | Source absence only; not an HTTP, price, privacy, quota, or access conclusion. R4-S05 |
Limits, data terms, and direct observation
No inspected current primary source publishes a numeric route-specific cap for either former alias or current Pareto. OpenRouter’s generic account/key/in-flight policy does not supply one. per_request_limits: null is not “unlimited.” R4 made no request, so it has no rate-limit header, latency, capacity, routing, reliability, tool, image, privacy-response, or entitlement result. R4-S04
Unbiased terms describe Pareto as a proprietary blended model and allow the underlying composition to change without notice. Those terms apply to Unbiased services. They do not transfer historical OpenRouter preview retention wording or historical Go documentation to another route. R4-S06
Benchmark method and figures
| Item | Published material | Classification / missing fields |
|---|---|---|
| Headline scores | Provider model card: DeepSWE 74, Terminal-Bench 4.0 51, MMMU-Pro 78, HLE 49, ArXivMath 88. | Provider-published. No identified task IDs, sample sizes, run dates, harness commit, prompts, agent settings, retry policy, raw outputs, or task costs. Not an independent or Union Alpha result. R4-S07 |
| Preliminary DeepSWE | Provider card: Pareto 26.9, Sep. 20, n=30, 70%, USD 0.29/task. | Provider-published preliminary claim; screenshot-transcribed and not independently validated. Denominator and cost method need confirmation. R4-S07 |
pareto-evals | Public configurable harness: temp default 0, seed default 0, bounded transient 5xx/429 retries, configurable concurrency, independent HLE judge, and response-cost/configured-token-price calculation. | Method artifact only. Endpoint/key/model are user-configured; Pareto rejects sampling parameters. Inspected surfaces had no raw result bundle, release/CI artifact, former-alias link, or fixed-version pin. R4-S08 |
Method limits
- Public browser retrieval succeeded. Shell curl and python3 were unavailable, so saved captures are faithful excerpt notes rather than byte-for-byte HTTP bodies.
- Start/end public retrieval found no material within-run change.
- Repository release, commit, and action subpages could not be fetched from the browser cache. “No raw package found” applies to the public landing, README, environment example, and run code that were inspected.
Recommendation
Wind down routine watch unless an official trigger appears: an explicit former-alias-to-fixed-version mapping, an immutable raw benchmark package with settings, an independent reproducible result, or official re-listing of a former alias.
Current pages: summary · availability · benchmarks · attribution. Historical reports: R2 and R1.