Deepr v2.37.0 ships bounded, evidence-first expert investigations for persistent expert councils, with separate factual and perspective learning lanes and stricter no-surprise-bills controls.
Highlights
- Persistent experts can investigate a question with distinct evidence lenses, challenge and revise positions across rounds, and produce provenance-indexed artifacts without mutating expert state during the run.
- Facts and perspectives are kept separate. Hypotheses, theories, stances, concepts, null hypotheses, and original ideas can remain useful without being mislabeled as verified facts or human-reviewed work.
- Explicit bulk learning preflights every selected expert and lane, hash-verifies producer artifacts, locks all targets, and applies atomically. A budget is a ceiling, not consent.
- Remote MCP, REST, WebSocket, A2A, queue, and worker paths now preserve scoped identity and ownership while failing closed on unproven authority, metered consent, or accounting.
- Claude Code is the only currently executable plan-quota adapter, and only after a fresh provider observation proves paid extra usage is disabled. Other plan CLIs remain visible but execution-blocked until Deepr can prove safe confinement and billing posture.
- URL retrieval, outbound MCP, local approvals, graph commits, report absorption, deployment secrets, reservations, and the append-only cost ledger received broad security and lifecycle hardening.
- The council and capacity guides now give exact commands, cost semantics, learning behavior, and honest works-now versus planned boundaries.
Validation
- GitHub CI passed lint, strict type islands, frontend tests and packaging, dependency and secret audits, and the full unit suite on Python 3.12, 3.13, and 3.14.
- 9,203 unit tests passed locally with 85% branch coverage, above the blocking 80% threshold.
- CodeQL passed for Python, JavaScript/TypeScript, and Actions.
- Live validation used local Ollama and safely proven subscription quota only. Paid and metered-at-margin validation spend was exactly $0.00.
- The corrected three-expert pilot staged six candidate writes but applied none. Semantic superiority and held-out evidence quality remain explicit evaluation gates, not release claims.
See docs/CHANGELOG.md for the complete record.