Orchestrate Research to Publication As One Workflow

Coordinates all four skills (research to writing to review to revision to finalization) as one non-stop workflow with integrity verification at decision boundaries instead of treating each phase as a separate task. It enforces Material Passport handoff contracts, runs mandatory 7-mode AI failure checklist at Stages 2.5 and 4.5, performs 100% claim verification before publication, detects hallucinated citations deterministically, and produces a Process Summary with Collaboration Quality Evaluation so you have a complete audit trail and confidence in the output.

Once set up, run /ars-full to start the 10-stage pipeline: (1) RESEARCH (deep-research skill, full/quick/socratic/systematic-review etc.), (2) WRITE (academic-paper skill with outline approval), (2.5) INTEGRITY GATE (7-mode AI failure checklist covering implementation bug, hallucinated citation/result, shortcut reliance, bug reframed, methodology fabrication, frame-lock) plus claim sampling (30% min 10 claims), (3) REVIEW (academic-paper-reviewer full review, decision checkpoint), (3→4) COACHING (Socratic revision coaching, max 8 rounds, user can skip), (4) REVISE (revisions incorporated), (3') RE-REVIEW (narrow verification panel), (3'→4') RESIDUAL COACHING (trade-off dialogue, max 5 rounds), (4') RE-REVISE (content frozen, hard cap 2 revision loops total), (4.5) FINAL INTEGRITY (deep Mode 2 re-run with zero tolerance, 100% claim check), (5) FINALIZE (output format selection plus AI disclosure), (6) PROCESS SUMMARY (collaboration quality evaluation plus self-reflection report). Output: publication-ready manuscript plus Material Passport (traces all decisions, audit trails, repro-lock declarations).

Requires Material Passport (Schema 9 YAML file, generated at Stage 2). Optional flags: ARS_PASSPORT_RESET=1 (context reset at FULL checkpoints, enable resume_from_passport=<hash> mode for cross-session continuation), ARS_CLAIM_AUDIT=1 (opt-in claim-faithfulness audit at Stage 4→5, default OFF, requires cited excerpt retrieval), ARS_CROSS_MODEL (OpenAI or OpenAI-compatible endpoints for cross-model verification), ARS_SOCRATIC_READING_PROBE=1 (optional reading comprehension honesty probe at Stage 1). Write-scope guard (v3.10): 23 single-phase agents fenced to their phase directories, no Bash access (use Grep/Glob + structured editors). Claim audit (v3.8, opt-in) emits 5 HIGH-WARN classes consumed by Stage 5 formatter REFUSE rules; baseline FNR<0.15 + FPR<0.10 on 20-tuple gold set. Terminal policies (v3.10, opt-in strict mode) on citation_existence verification allow default-advisory pass (user /ars-mark-read-ack-able) or strict refusal.