PF research library · 3 min read
Autonomous ISAES implementation contract (r34)
r34 documentation · isaes-autonomous-contract.md · Formatted reading copy
On this page
Baseline archive SHA256: 1987acf5712a7d2311744089bfb350885501bd925e18a2c31ee43447329d6a74. Preserve r34 and existing non-AI/report-discussion behavior. This is an integrated opt-in autonomous investigation, not a dialog. No fabricated executions or scientific validation.
Integrated module responsibilities
- service.py, contracts.py: controller; durable owner-scoped jobs and standing policy, budgets, trigger, BYOS/lifecycle integration.
- provider.py: provider; existing providers, autonomous system prompt, tool orchestration, checkpoint/replay, claims and decision tools. Uses the dedicated autonomous purpose through the existing provider clients.
- science.py and scientific worker: scientific tools; analysis/re-analysis, isolated generated code, visualization/export, child runs.
- literature.py: literature tools; safe online contextual discovery/retrieval, associated-source reading, extraction, source identity/version and provenance.
- http.py and GUI: application controls; API integration and standalone/SaaS enforcement.
- documentation: operational guide, acceptance matrix, developer guide and end-user manual.
Shared service API
AutonomyService(assistant): policy(owner,workflow_id), set_policy(owner,workflow_id,policy,expected_revision=None), enqueue(owner,run_id), get(owner,id), list(owner,workflow_id), process(owner,id), pause(owner,id), resume(owner,id), cancel(owner,id), artifact(owner,id,path). Internal interfaces: persist(owner,job,files=None) -> job; guard(owner,job=None); check_control(owner,job); consume(owner,job,kind,amount=1). artifact returns verified bytes. persist files are relative-path: bytes and registered in job artifacts. Use rp_investigation records and rp_autonomy_policy (kind <24 chars), workflow parent IDs. Root directory workflows/investigations/. All external artifacts retained and included in BYOS manifest and lifecycle dependency rules.
Job contract
Required: id, workflow_id, run_id, status, stage, policy, policy_revision, artifacts, usage, decisions, sources, claims, analyses, presentation, errors. context contains immutable base report_id/report_sha256/input_sha256 and retained workflow/run metadata. sources/claims/analyses are lists of typed objects with stable IDs. Provider uses service methods for durable checkpoints, control boundaries, cumulative budget. Modules expose definitions() and dispatch(service,owner,job,name,args). dispatch returns finite JSON, may modify job, writes bytes through persist. Tool naming prefixes literature_, science_; provider owns investigate_read, investigate_plan, investigate_assess, investigate_finish. Unknown tools fail closed. Provider entry: investigate(service,owner,job) -> job. Modules invoked from provider dispatcher. Never self-call old review_proposal to impersonate authenticated user.
Standing policy
schema_version=1; enabled, online_research, allow_generated_code, allow_followup_runs, share_with_provider are booleans. Enabled+share_with_provider is explicit one-time delegation; policy records actor and revision. budgets: max_steps, max_provider_calls, max_tokens, max_seconds, max_sources, max_download_bytes, max_child_runs. All configured limits are cumulative, including retry/resume/children. All jobs snapshot policy; revocation must stop at next boundary. Optional domain allowlist for online retrieval. Defaults must be visible and conservative but enabling capabilities must work end-to-end.
Scientific acceptance
Autonomous context -> plan -> computations -> claim/evidence assessment -> presentation -> refine or finish. Automatic run trigger, no user prompt; live AI prerequisites transparent. Literature can drive targeted web retrieval, source repair and separate evidence expansion. Exact citations and versions; equations/tables/figures inspectable; approximate digitized values identified. Assess supports/challenges/mixed/not_comparable/insufficient; LLM judgments labeled, no invented empirical validation. Circular evidence, adaptive selection and incomplete extraction recorded. Preserve conflicting results. Generated code isolated and verified, never accesses credentials or changes original solver/results; actual child runs labeled and linked. Holdout roles protected; post-hoc adaptation exploratory. Native report exports carry same evidence/decisions and source identities. No provider operation when opt-out.
Delivery acceptance
Tests cover real calculations plus clearly labeled protocol/adversarial fixtures. Interrupted process resumes completed receipts, retries are bounded, cancellation persists, data access obeys owner/BYOS controls. Complete guide updates, exact release hash/manifests, original archive contents preserved. Live provider, native Windows, independent expert and applicable production qualification only claimed when actually executed.