Rollout modes and operations

Autoresearch 1.0 ships three staged session modes. The mode is pinned by the confirmed plan — a session cannot change mode after creation, and every mode runs under the same pinned safety policy and the same deterministic Supervisor epoch frame.

Mode	What runs	Terminates	Default
assistant	Bounded generation + staged reviews (initial, then evidence-grounded full)	Completes when the batch is reviewed	Yes
tournament	Assistant, plus proximity landscape, adaptive matches (simple + debate), Elo history, evolution, meta-review	Completes on work exhaustion or stops at a budget ceiling	No
continuous	Tournament, re-armed by new seeds and signals	Never auto-completes — only budget ceilings or operator stop/pause	No

A plan that names no mode is an assistant plan. Continuous is never a default: the plan parser refuses a continuous draft without budgets, and even a hand-crafted or migrated plan cannot become a session unless at least one enforceable ceiling (tokens, costUsd, wallTimeMs, or maxEpochs) is set — session creation throws continuous mode refused otherwise.

Cost controls

Plan budgets are enforceable ceilings, not annotations. At session creation the plan vocabulary maps onto the Supervisor stop policy:

• budgets.tokens → maxTokens (ledger tokens)
• budgets.costUsd → maxCost (ledger cost)
• budgets.wallTimeMs → maxDurationMs (since session creation)
• budgets.maxEpochs → maxEpochs

Ceilings are evaluated before dispatch at every claim boundary. A tripped ceiling terminates the session with a durable stopReason (max-tokens, max-cost, max-duration, max-epochs) after reconciling the last checkpoint — a nearly exhausted budget never skips reconciliation and never leaves a session falsely running. Spend lands on the session ledger exactly once per task outcome; replays are budget-neutral.

Recovery procedures

Every failure mode has a typed, inspectable recovery path:

• research.supervisor.inspect returns an operational block: attempt-exhausted tasks (with their child run ids), safety-held artifacts, stale lease counts, and a human-readable recoverable action for each.
• Crash or restart mid-epoch: the next claim re-leases unreported content-addressed tasks; re-ingesting an already-processed child run is replay-neutral. No duplicate hypotheses, reviews, matches or rating rows can result — effect identity is content-addressed end to end.
• Paused sessions (operator pause or safety breaker): resume re-arms the trigger chain. Sessions paused by the safety breaker keep their pauseReason.
• Idle continuous sessions: seeding a hypothesis into a running session re-arms the chain; the session page mirror shows the current epoch, budget, landscape, meta-review and safety state at every checkpoint.
• Failed work: attempt-exhausted tasks terminate as error-carrying holds and never block the epoch; re-seeding or plan refinement re-plans equivalent work under new task identity.

Provider setup

Under PROVIDERS=fake (the only mode CI exercises) every specialist, embedder and parser is deterministic, and no external calls occur. Live runs require model provider credentials for the agent engine and are opt-in per deployment; the benchmark suite additionally requires an explicit spend ceiling before it will run any arm. External execution (MLOps launches) and any outreach remain separate, explicitly approved lanes — no research mode performs them.

Migrating from the 0.6 loops

The 0.6 heartbeat/Elo-sweep loops keep working; 1.0 is additive:

• Legacy pages are adopted in place — aliases keep old slugs resolving, and legacy research.status vocabulary remains valid beside the 1.0 eligibility/selection fields. In-flight 0.6 workflows finish under their original definitions.
• The research.elo.sweep baseline stays untouched; 1.0 session ranking writes an advisory research.elo snapshot with full versioned history in its own collections.
• The autoresearch-mlops adapter consumes both vocabularies: legacy open hypotheses by status/goals refs, and 1.0 hypotheses only once a scientist selects them. Empirical outcomes feed back as exact-ref evidence and an empiricalStatus derived from the human-confirmed observation directions.
• N-1 record shapes are journal-compatible: sessions written by earlier bundles keep claiming and finishing epochs after an upgrade.

Limitations and non-claims

• Autoresearch is inspired by Co-Scientist-style multi-agent loops; it does not claim to reproduce any published system or its results.
• Internal Elo is an advisory scheduling signal. It never gates empirical status or external execution, and it is not evidence of hypothesis quality, novelty or truth.
• Novelty verification is bounded by the evidence a session retrieved; a review that could not search marks itself incomplete rather than guessing. Human preference and empirical-outcome correlation are ground-truth-pending in the benchmark report — no automated proxy stands in for them.
• Safety screening is deterministic and conservative (D2/D3 domains refuse Autoresearch modes entirely); a quarantined artifact keeps its page for audit but never reaches reviews, ranking or selection.