# Founder guide — operating AutoScholar.ai

One person operates the foundry. Everything happens in the admin console at https://autoscholar.ai/admin (LAN fallback: http://<M3U96b LAN IP>:8470/admin).

# 1. First sign-in

• Username admin, bootstrap password admin123 — valid only until the first sign-in. You are forced to choose a strong password (≥ 12 chars, three character classes, not the bootstrap one, never reused). All other sessions are revoked. The rotation was already performed on 2026-10-30; the current password is in the operator's secrets folder.
• Sessions last 12 h (2 h idle). Five failed logins lock the account (1 min, doubling up to 1 h). All events are in Settings & security → Audit log.

# 2. Write a research brief

Start from Templates (general, empirical, theoretical, computational, review) — "Use this template" prefills New research. Follow docs/RESEARCH_INSTRUCTIONS_GUIDE.md: one precise question, mode, verified-source rules, data/proof foundations, pre-registered procedure, robustness, acceptance criteria. Briefs that target the platform (sudo, SSH, credentials, "ignore the rules") are rejected at upload; briefs that require regulated activities fail at preflight with the reason shown in the run timeline.

# 3. Launch

New research → upload .md/.txt or paste → Validate & freeze (exact bytes + SHA-256 stored as an immutable version; warnings list missing sections) → choose discipline, mode, secondary modes, compute profile, budget cap (USD), runtime cap (minutes), attempts, auto-publish → Launch on MacLustr. The button carries an idempotency key: refreshing or double-clicking cannot create a second run. You get a run ID immediately; close the browser whenever you like.

# 4. Monitor

Runs → run: live timeline (reconnects automatically), stage progress (preflight → framing → literature → design → execute → robustness → manuscript → review → reproduction → quality → publication), cost vs cap, attempts, policy denials, artifacts (download any file). MacLustr: worker health, heartbeat age, sandbox status, running jobs; drain a worker before maintenance, quarantine a misbehaving one. Cancel stops a run cooperatively (its logs and frozen input stay); Retry grants one more attempt resuming from the last checkpoint.

# 5. Review quality

Quality → run: every deterministic gate with evidence (hard gates block publication; soft gates warn), the automated referee's findings and resolution states, the rubric scores and the publication decision. Hard scientific blockers (unverified citations, untraceable numbers, missing PDF, synthetic data, secrets, simulated worker) cannot be overridden. Other failures can be overridden with a written rationale, which moves the run to awaiting approval.

# 6. Publish, correct, retract

• Auto-publish: when enabled on the run and permitted for the mode/discipline, a run that passes all hard gates is published by the scheduler within seconds — public page /research/<slug>, PDF, LaTeX source, code/data when present, citations, provenance.
• Manual: Publications → Publish for runs awaiting approval; "publish as correction of …" creates version n+1 of an existing paper (history preserved).
• Retract writes a permanent notice; Withdraw artifacts stops file delivery (HTTP 410) while keeping the notice and the private audit copy; Edit metadata records a versioned change.

# 7. Emergency controls (Settings & security)

• Pause dispatch — no new research starts; running jobs continue.
• Halt publishing (global or per mode) — passing runs wait in awaiting approval.
• Worker drain/quarantine; provider availability; env-level switches DISPATCH_ENABLED=0 / PUBLISHING_ENABLED=0 (restart required).

# 8. Costs

Each run's budget cap is enforced per stage by the Agent SDK and across stages by the runner; the scheduler enforces the runtime cap. Spend shown is the SDK's client-side estimate (not the invoice). Typical small study: USD 5–25.

# 9. When something fails

• Preflight INFEASIBLE → the brief needs changes (see reason). No retry is attempted.
• quality_failed → open Quality, read the hard failures, retry (a new attempt revises from the checkpoint) or fix the brief and relaunch.
• Worker offline → queued runs wait; running runs resume on the next heartbeat window from their last checkpoint.
• Operations, backups and restore: docs/OPERATIONS.md.