Design Docs¶
Breakdown documents for the Release Dependency Map workstreams. The ROADMAP owns what and in which order; these docs own how and why at one level more detail. Keep them in lockstep: when a workstream's scope changes, update both.
Conventions:
- No dates. Bands are gated on exit criteria; sequencing is dependency order.
- Validation cost is part of the design. Every doc states what can be verified free (deterministic tests, mocks) vs what needs a paid run, and budgets the paid part.
- Explicitly-not sections are binding. They exist so scope doesn't get re-litigated; changing one is a deliberate decision, not drift.
- No real company names in examples - placeholders only (see
docs/CONTRIBUTING.md).
| Doc | Workstream | Version band |
|---|---|---|
| company-analyst-product-contract.md | Canonical product, artifact, economics, evaluation, feature-admission, and GitHub/PyPI release contract for the company-and-website workflow | cross-cutting |
| research-tradecraft.md | Collection-first → hypothesis-first: framing, Day-1 hypothesis tree, plan checkpoint; content within a curated structure (deepens #4) | 1.x → 2.0 |
| eval-plan.md | Pre-registered, cheapest-first plan for the pending paid evals (calibration baseline, framed-vs-unframed, content-depth): exact commands, instruments, acceptance criteria, cost | 1.x |
| premium-quality-eval.md | Blinded five-company Standard/Premium evaluation for epistemic quality, argument coherence, and current writing-topology differentiation | 1.x quality validation |
| agentic-balance.md | Rule vs judgment: when a primr component stays a deterministic workflow vs becomes a Level-2 model decision; the keep-list and budget couplings under the tradecraft work | cross-cutting |
| run-epistemic-ledger.md | Run-scoped source receipts, evidence anchors, findings, strategic inferences, a small semantic firewall, and the shadow-to-promotion path into claim-aware writing and memory | 1.43 → 2.x |
| open-knowledge-format.md | Adopt OKF v0.2 (markdown + YAML frontmatter + linked tree) as the findings-interchange shape, with permissive consumers, a stricter Primr producer contract, and confidence distinct from verification; the polished report stays unchanged | cross-cutting (2.0 / 3.0) |
| 1x-completion.md | Finishing the excellent single-shot brief; immediate slice is evidence-grounded validation, with label calibration as the first measurable slice | 1.x |
| engineering-excellence.md | Anti-slop enforcement layer: dev-facing AGENTS.md contract, architectural fitness functions, file-size ratchet, CLI verb convention, toolchain currency | cross-cutting |
| runtime-language-boundaries.md | Evidence-gated runtime policy: Python product layer, supervised workers, optional Rust acceleration, and explicit Go/Mojo trigger conditions | cross-cutting |
| 23-orchestrator-refactor-map.md | Working map for the #23 orchestrator refactor (stages, tangles, batch order) | 1.x |
| 2.0-backend-freedom.md | Capability routing + local/hybrid inference; routed utility-stage metadata shipped for fast.scrape_summary, fast.source_relevance, and fast.hiring_signals; next slice is provider-health and eval-backed host/local promotion |
2.0 |
| provider-expansion.md | OpenAI/Anthropic recipes, governed OpenRouter preview, Bedrock/Foundry gateways, and $0 local profile (provider catalog refreshed September 2026) | 1.x Phase A; 2.0 Phases B/C |
| 2.0-research-memory.md | Company-run continuity and delta mode; indexed memory is a 2.0 candidate only after representative report lift | 1.43, conditional 2.0 / 2.x |
| 2.0-agent-control-plane.md | Per-tool authz, approval tokens, tool-call, resource-read, and A2A skill-call audit log, job-scoped artifact metadata, QA summary, usage/cost summary, source appendix summary, scrape trace summary, verification summary, calibration summary resources, MCP report-read scope separation, A2A skill-scope parity, A2A compact-resource parity, A2A report-read parity, and A2A research approval/budget parity; next slice is remaining output negotiation | 2.0 |
| competitive-positioning.md | Plain fit vs chat deep research, enterprise CI, and research APIs; strengths, gaps, and humble talk track (not marketing) | cross-cutting |
| progressive-artifacts.md | Time-to-first-useful: working brief layers during long runs without replacing the full Strategic Overview | 1.x usability |
| grok-default-routing.md | Defaults track current provider flagships (2026-10-28 owner decision); earlier eval-gated Grok policy | 1.x model routing |
| 3.0-research-frontier.md | VLM extraction, knowledge compounding, artifact handoff | 3.0 |