0085: Circle STARK is the default research workflow¶
Date: 2026-10-28. Status: owner-selected; local acceptance passed, delivery checks tracked on PR. Superseded by 0090: the pairing backend was removed.
The owner explicitly selects Circle STARK as the default and base case for commands, demos, headlines and documentation. This supersedes the operational default choice in decision 0083 and the completed comparison report. It does not change the recorded measurements or establish a security ranking.
Hypothesis: default classifier and composed workflows can run entirely with
Circle STARK proofs. Baseline
is f613c611cefa9037320082b73777f26c9ec45990. Preserve all 55 LLM verifier source
pins, the four classifier circuit/compiler sources and Rust engine, the model
exports, arithmetic, thresholds and immutable benchmark plans/data.
Mutable paths: Make dispatch, new native model-preparation/default controllers, the existing Circle orchestration driver, workbench adapters/catalog/result views, focused tests, current docs, STATE/LEDGER and new acceptance evidence. Old workflow IDs and historical results retain their original meaning.
Pass criteria: bare make e2e builds/enrolls Circle and generates/verifies nine
proofs with its binding negatives; default retained-input make e2e-composed
checks 23 SASH proofs and three new Circle classifier proofs for one request;
web/TUI choose Circle first and validate the actual new result contracts.
Local, distribution, app, docs, exact-head CI, main and Pages gates must pass.
Run locally, one proof workflow at a time, four prover threads, at most 30 GiB sampled worker RSS, 4 GiB per classifier campaign and 1,800 seconds for that campaign. The classifier controller is capped at two hours; keep 8 GiB free during proving. Retained LLM composition needs no new paid proof job; fresh LLM jobs keep their existing admission ceilings and source/parameter checks. No hiding, hardware, independent qualification or production-policy claim follows from the default.
New acceptance runs have new directories. Existing CPU/GPU measurements remain labelled with their original source, stateless harness and timing boundaries; they are not measurements of the new preparation/controller path.
Acceptance¶
The numeric record retains bare Make training/proving, retained-LLM default composition and an actual browser-launched nine-proof run, with source, enrollment and artifact hashes. The 55-file LLM identity and classifier engine/compiler identity are unchanged. Core checks pass 2,588 tests with 50 optional/platform skips; all 56 app tests pass. Strict docs and core distribution checks pass after correcting three old anchors. Initial fixture/menu/docs failures are retained, not proof samples. No new paid jobs or hiding/hardware/qualification claims are introduced.