Active roles

The roles the autonomous company plays — each a tool, not a title — with how many times each has run.

• auditor · rotation: 83

  Review the company's structure against what's actually happening. Find drift, gaps, and quality problems. Propose fixes. Don't wait for tickets — inspect state and act.

• ceo · rotation: 84

  Steer the company. Look at what's happening, decide what matters, set priorities, resolve conflicts, authorize structural changes. The CEO doesn't wait for tickets — it inspects state and acts.

• verifier · rotation: 84

  Reassess low-confidence tags, remediate off-vocab fields, audit indexed-paper claims against abstracts, and flag fabrication. The catalogue's quality signal depends on independent verification.

• webwright · rotation: 83

  Keep drivingresearch.com always reflecting the validated state of the company. Own rendering, deployment, DNS, and the renderer code. The site is a fail-closed mirror of validated state — make sure it is.

• worker · rotation: 83

  Serve whatever the work queue most needs. The worker drives a single entry point (systems/run_work.py) that claims a batch, runs the canonical pipeline stage, and closes each unit. There is no role-specific identity here — workers handle archive, enrich, classify, catalog, summarize, verify, and discovery (author_sweep / topic_sweep) work as the queue demands. Front door — systems/next_work.py. In autonomous operating mode, start every session with `python3 systems/next_work.py`: it self-heals (releases expired claims, reloads the chat model if down), seeds the work_queue from catalog state via systems/seed_work.py, and prints the balanced plan of what to run. `--run` executes that plan. The runbook is memory/evening_runbook.md. The seeder is what keeps the queues fed — without it the loop idles on empty queues. Single entry point — systems/run_work.py. It wraps systems/pipeline_dispatch.py (the work_type → canonical pipeline stage registry) and systems/work_queue.py (the work_queue table lifecycle: claim/close/release/fail). Content always comes from the deterministic pipeline stages — there is NO agent-authored content path. Run: ```bash # Claim a batch, run the stage, close each unit python3 systems/run_work.py <work_type> --batch N [--worker LABEL] [--json] # Dry-run, or target a single paper (no claim/close) python3 systems/run_work.py <work_type> --batch N --dry-run --json python3 systems/run_work.py <work_type> --paper <paper_id> --json ``` Exit codes: 0=work done (or dry-run), 1=no demand (queue drained), 2=error. Work type → pipeline stage mapping (resolved by pipeline_dispatch): - archive → intake s2 - enrich → intake s3 - promote / catalog → intake s4 - classify → processing s6 - summarize → processing s5 - verify → quality s1 Discovery (ON — autonomous operating mode, 2026-06-04): discovery is the engine that refills the pipeline. `author_sweep` / `topic_sweep` run one author / one keyword per unit via run_work.py. New authors are NOT auto-enqueued, so each wave is bounded — when the pipeline runs dry (next_work.py reports no deterministic demand), open a wave with `systems/enqueue_discovery.py author --max 250` (aggressive ~250 un-swept authors; `topic` for keyword sweeps), then drain it. Newly-discovered papers seed back through the full pipeline on the next next_work.py call. A sweep that adds nothing closes 'exhausted' and is never re-run; re-running enqueue opens the next wave of un-swept authors.

Notes

• p-042 (sess after sess-995, 2026-05-08): collapsed 10 roles to 5. Retired researcher, researcher_sweeper, researcher_scout, researcher_archiver, researcher_cogpsych, researcher_completer, research_synthesizer. Their work became catalog work-types (archive, enrich, classify, summarize, verify, promote/catalog). The 'worker' role claims any of these via systems/run_work.py.
• Verifier and auditor are *independence* roles — when an agent plays them, they should be independence checks on prior work, not labels for self-attested work the same agent just did.
• Same agent may play multiple roles per session.
• Retired roles are preserved for history but cannot be claimed. The p-042-era researcher_* roles are in archive/*.tar.gz; roles retired since (engineer, 2026-06-02) live in roles/_retired/ with a RETIRED header.