Left and right are randomized. Model, provider, track, and Epicure condition appear only after a valid vote.
Epicure-powered evaluation
Taste the reasoning, not the model name.
Two anonymous culinary answers. One prespecified comparison. The service supports model comparisons under matched tool access and paired evaluations with or without Epicure.
The arena remains disabled until independent judgment, task-quality, privacy, abuse-control, provider-budget, and release gates pass. Existing real outputs are not exposed as public battles.
16 exact endpoints produced 211 gross complete Epicure-off/on pairs. After quarantine and source-verified coverage recovery, the blinded workload contains 186 uplift pairs and 915 arena comparisons over 188 Epicure-on answers; 4 further answers have no same-task peer. Quality judgments remain 0; these are workloads, not scores.
Read the protocol and release statusFree-form public battles and sealed controlled evaluations receive separate snapshots and are never pooled.
Controlled runs bind the admitted endpoint, task schedule, tool release, analysis and rater plans, and spending cap before collection. Cost and latency are reported afterward.
Research artifacts
The paper, arXiv source, privacy-safe review-workload identities, frozen endpoint evidence, exact Epicure application source, reconstruction kit, and execution receipt are served with their SHA-256 identities. The exact runtime data bytes are hash-verified at the immutable public commit named in the packet. Technical public-input reconstruction has passed; payload rights, a signed immutable image, recovered training lineage, independent reproduction, and official benchmark release remain open.
Current paper PDFPublic
The August 2026 retrospective measurement audit and exact-frontier development study.
sha256 1bbbfa8cff6f2ef4b1d50e2294d08461123e3237021e00c3e5b5e0e930fbee27Read paperarXiv source packagePublic
Self-contained LaTeX source, vector figures, aggregate plot data, and provenance manifests.
sha256 9df3d5ae7e732b288b281d52a8652508ece7098e0283314ec7ef97a0ca431659Download sourcePaper release manifestPublic · independently frozen checksum set
Binds the exact 25-page PDF, arXiv source archive, and full paper artifact inventory distributed by this release.
sha256 f854fe9e19ff8c39e95a7b44a9f857727495aa111f429c2244ea820b5562aa28Verify release hashesPaper artifact inventoryPublic · SHA-256 inventory
Lists the manuscript sources, figures, aggregate plot data, provenance records, and readiness artifacts bound into the frozen paper package.
sha256 20eade7d90ed1623969e275711b8c0fc91a61df729feae35a017ba9897d61176Inspect artifact inventoryCurrent-study manifestPublic
Content-addressed roster, evidence boundary, retained workload counts, and same-origin artifact map.
sha256 7a2b5bb7dcb6df414d1b07dd73eecd37013dd391947df2442fcb642bc230e594Inspect manifestDerived blinded-review workloadPublic · hash and identity records only
915 arena comparisons, 192 Epicure-on answer identities (188 compared and 4 unpaired), and 186 uplift pairs. It contains no prompt or answer text, raw tool traces, credentials, ballots, or quality scores and is not the governed raw research dataset.
sha256 a50f7edfe24b995b94771b4df75a1154e3a76e19519edd177a4539d3631b4622Inspect review identitiesQwenCloud post-freeze operational projectionPublic · real calls · permanently unranked addendum
2 real arms, 6 provider responses, and 2/2 successful Epicure calls. Provider cost is unknown, the alias is mutable, and the record changes no season count or quality fit.
sha256 b2f7790b3eb18d1df083397ce02b5296c549e5ed3ddb3d3f32ea776db3ddca04Inspect Qwen addendumQwenCloud zero-call recovery receiptPublic · append-only recovery metadata
Records that terminalizing the complete successor source required no provider call, Epicure call, source mutation, replay, or new reservation.
sha256 7bb5f1392a2422437edc138b14940cd92736caa6bc6328acbf4b2dd73e8d479aInspect recoveryProspective task-validation campaignFrozen instrument · zero human ballots · no generation authorized
1,052 attributed public human-authored questions yielded 418 provisional strict passes and 505 blind-review-eligible candidates. The frozen slate contains 180 candidates for a 120-task target. Source records remain undistributed while rights and contamination review is incomplete.
sha256 76b248477b3adc81b6eb198666a93538534db8e945567e2a99fc69085f709709Epicure public reconstruction packetSource and data verifiable · release gates open
Version 3 binds the 28-file exact study application, 11 runtime-data files at an immutable public commit, dependency lock, SBOM, same-operator rebuild receipt, and lineage audit. It does not establish payload rights, an immutable executable image, exact training lineage, or independent reproduction.
sha256 554ddcede820860a3150e621e9834c41bf959dee4878cd13cb01eb2c2512d52dDownload packetEpicure public-input reconstruction kitPublic · executable recipe and exact hashes
Deterministic source archive, immutable 11-file data map, 41-wheel dependency lock, SBOM, tool catalog, parity fixture, and a standard-library reconstruction program. The kit contains no model/data payload, dependency wheels, credentials, or training material.
sha256 bc5b109840c3c7d468fdc050eff7b907b66a0bc3a31899bd4c39ea8bcb1ae4b6Download reconstruction kitPublic-input reconstruction receiptVerified by study operator · independent replication pending
A clean CPython 3.12/Linux x86-64 run fetched and hashed all declared public inputs, installed every locked wheel offline, verified wheel RECORD hashes, reproduced the runtime payload identity, and passed fixed tool probes. It is not an independent-operator receipt.
sha256 0d28bb310614e707edf678977b2a6e548bdafb0bdf7adea185a7e46cff0a4519Inspect execution receiptEpicure release-candidate boundaryTechnical reconstruction passed · official release blocked
Binds the verified reconstruction while keeping independent reproduction, payload redistribution clearance, training lineage, immutable OCI identity, and benchmark rank eligibility false.
sha256 878d93f8872f3f948b9a125dcb6084f01bbf06cc204e461d409611d96946f384Inspect release boundaryExact Epicure study application sourcePublic source archive · not an executable runtime
28 exact source files (116,668 bytes) under the observed MIT code license. The archive contains neither runtime data/model payloads nor credentials.
sha256 d08fb475e9c325a8c41daf5b789e6b4bca547228139eece4578f9b06c324703cDownload exact sourceExact-frontier protocol evidencePublic · unranked
The 16-endpoint main-stratum roster, canonical identities, fixed provider routes, and operational pair counts used by the paper and public models route. Post-freeze QwenCloud evidence is kept in its separate addendum. This record contains zero quality judgments and zero rankable comparisons.
sha256 5d834be66ec8fca611a298c2cbed75315b1f2cfc6bcd135fc465adf93c57a47dInspect frozen rosterOpenRouter catalog snapshotPublic snapshot
338 discovery records observed 3 August 2026. This is not a live availability, compatibility, eligibility, or quality claim.
sha256 0a9cbb38cc519cc90c337b12a55d30c5e1f931c0696eb3d1b4791af68500d0baInspect raw snapshotCatalog acquisition receiptPublic receipt
Unauthenticated request boundary, selected response headers, raw byte count, content hash, and claim limitations for the OpenRouter snapshot.
sha256 2d831746e833a7c3c2e04e3d124adbfbdd07319d5cbcc47bed6ca69257404cf6Inspect receiptMethods and release gatesPublic
Protocol, task holds, inference boundaries, governance, privacy, and limitations for the current study.
Read methodsBenchmark research dataRelease held
Raw prompts, outputs, and tool traces remain private pending source-rights, provider-output, privacy, and identity-leakage review.
Persistent research archiveDOI/deposit pending
No public archival deposit or DOI is claimed on this site yet.