Research knowledge index — agent and maintainer entry point¶
Navigation, not evidence. This index links canonical research and implementation records. Never promote a development-only result to independent validation, or copy benchmark numbers here as though this page were the experiment ledger.
Tracking: #1712. Canonical cross-session research tracker: #200.
Before resuming development or rerunning an experiment, consult the 2026-11-04 branch and evidence reconciliation. It connects pinned branch/worktree inventories, integration decisions and later result records. This is a dated inventory, not proof that every branch was merged or that an experiment passed; recheck live PR state before acting. Private local DEV inputs remain outside the repository.
2026-11-04 situation after hygiene¶
The reconciliation separates the following states; removing a branch does not resolve its product gap.
| Area | Current state | Source |
|---|---|---|
| Original question → required fields | Existing local DEV failures remain. Complete coverage of extracted needs is not proof that the model extracted every need. | Reconciliation |
| Default API → schema graph | Default planning remains lexical; prepared model/graph field routing is opt-in. | Paper parity |
| Scoped typed field plans | Implemented for declared source contracts, conditions and bindings; this is not general semantic understanding. | Field-first execution |
| Commercial correctness, latency and cost | Not independently qualified by contract tests or small E0/E1 studies. | Acceptance matrix |
| Release | v0.18 remains held. Hygiene and CI success do not clear the release gate. | Held PR #1356 |
Historical variants are preserved with their original outcomes rather than all being merged into the product. The reported OpenClaw vLLM Qwen122B connection is the next model environment to inspect after this organization and situation report; no new model run is part of the cleanup.
Read in this order¶
- Research program and current gates: current priorities, counterexamples, research scope and non-negotiable constraints.
- Research governance: preregistration, development, freeze, calibration, blind-final evaluation and the required session-resume sequence.
- Tool-routing landscape: external OSS comparisons, already-tested approaches, negative outcomes and no-repeat decisions.
- Prior-art roadmap: literature and historic hypotheses.
- Routing status and experiment index: historical experiment interpretation.
- Experiment ledger and prior-art registry: machine-readable provenance; consult these before citing performance claims.
- Research milestone plan: proposed milestone-to-issue mapping; verify live assignments before claiming completion.
Evidence-to-implementation navigation¶
| Stable topic | Question / decision | Evidence source | Implementation or interface | Tracking | Evidence status |
|---|---|---|---|---|---|
| SR-R01 | Typed schema and source authority | Research program | Architecture | #1450 | Design and partial tests; external security gate not independently cleared |
| SR-R02 | Native G5 graph, FieldGroups and fixed K | Research program | Capability graph | #1560 | Development evidence; independent promotion pending |
| SR-R03 | EN/KO/JA retrieval and no-fit | Research program | Routing status | #1339 | Frozen negative evidence retained; subsequent DEV is not blind |
| SR-R04 | OSS tool selection and integration | OSS landscape | Pydantic AI validation, Agents SDK validation | #644, #645 | Limited E0 integration; not broad task success |
| SR-R05 | Open-set rejection and prior art | Prior-art roadmap | Experiment index | #388 | Mixed and rejected historical approaches |
| SR-R06 | Full-catalog versus bounded retrieval | Research program | Experiment ledger | #432 | Frozen independent negative result; do not overwrite |
| SR-R07 | Research lifecycle and reproducibility | Governance | Experiment ledger | #200 | Process documented; individual artifact completeness not audited here |
Interpretation: A link to source code proves a navigation path, not that the linked implementation satisfies a research gate. PR merge, CI-green, development benchmarks, independent confirmation and release approval are distinct states. Check live issues and PRs before updating this table.
Before starting a new experiment¶
- Search this index, the OSS landscape, prior-art registry and experiment ledger.
- Search GitHub issues and open PRs for the same hypothesis and exact evaluation slice.
- Record the existing decision, negative evidence and remaining uncertainty.
- Create or resume a canonical issue; preregister allowed and forbidden data, frozen configuration, baseline, success and abstention gates.
- Link issue → protocol → run/artifact → ledger → decision → implementation PR.
- Update this navigation table only after the original evidence and decision records have been committed. Do not silently change historic statuses.
Agent retrieval contract¶
- Start with this small index; open only the relevant canonical sources.
- Prefer source-owned schemas and the ledger over summaries or chat recollection.
- Never treat semantic similarity, score > 0, a retrieved candidate, or a successful CI run as execution authority or proof of downstream task completion.
- If an evidence/implementation link cannot be verified, mark it unverified.
- Keep English/Korean navigation aligned; avoid duplicating metric tables in translations.
- For a new session, follow governance's resume order and verify live issue/PR/workflow state.