44 lines
4.9 KiB
Markdown
44 lines
4.9 KiB
Markdown
|
# §G
|
||
|
|
|
||
|
|
Broaden GEWU ingestion so every project considers technical, conceptual, cultural/historical, and financial/market material. Persist plausible but uncertain relevance as a review flag alongside normal ingestion, optionally assess relevance and lens with JEV, and let explicit project-scoped user feedback guide later ingestion without weakening evidence validation or silently changing project policy.
|
||
|
|
|
||
|
|
# §C
|
||
|
|
|
||
|
|
- The four research lenses are defaults to consider, not quotas; a source need not contain material for every lens.
|
||
|
|
- A plausible connection with uncertain relevance follows the normal proposal flow and also receives a durable user-review flag; the flag neither blocks nor auto-publishes the proposal.
|
||
|
|
- Exclude only clearly out-of-scope material, unsupported/unavailable source content, or work deferred for an explicit budget; retain section-level coverage and reasons.
|
||
|
|
- Evidence validity, source attribution, scope authorization, and proposal acceptance rules do not change.
|
||
|
|
- User-review flags are stored separately from proposal validation's existing `needs-review` status and protected owner decisions. They notify/filter for human feedback and do not control proposal acceptance.
|
||
|
|
- Feedback is scoped to one project. It may inform later triage only when the user explicitly chooses to save it as project guidance; it does not train the model or alter acceptance policy.
|
||
|
|
- Existing projects receive the four default lenses when read, while retaining their current questions and project-specific context.
|
||
|
|
- Optional Jev relevance/lens scoring is advisory; missing configuration or provider failure never blocks the normal ingest/proposal path.
|
||
|
|
|
||
|
|
# §I
|
||
|
|
|
||
|
|
- Project context exposes default lenses: `technical`, `conceptual`, `cultural-historical`, `financial-market`, plus optional project-specific lenses.
|
||
|
|
- Agent-to-service interface: create a bounded review flag containing project/run/source revision, extraction and exact section locator, bounded evidence excerpt, lens, concise candidate summary and relevance rationale. The agent still follows normal proposal evaluation/submission. Optionally attach backend JEV relevance/lens probabilities; its failure does not block ingestion. It must not accept arbitrary local paths or credentials.
|
||
|
|
- Human review interface: an authenticated GEWU web view lists items within authorized project scope; a human decides `include`, `exclude`, or `defer`; optionally saves a concise reusable guidance rule for that project. Decisions record reviewer, time, rationale, and item identity.
|
||
|
|
- Agent project context returns only approved project guidance derived from human review. Review-item content and guidance remain distinct from canonical knowledge proposals.
|
||
|
|
- Candidate names are provisional: MCP `ingest_review_create`; authenticated review read/decision routes; `get_project_context` returns `research_lenses` and approved `ingest_guidance`.
|
||
|
|
|
||
|
|
# §V
|
||
|
|
|
||
|
|
V1: Every ingest run considers all four default lenses unless the human project context explicitly narrows scope.
|
||
|
|
V2: Plausible but low-confidence relevance follows normal ingestion and creates a review flag with source-bound section locator and rationale; it is not silently dropped or held from proposal submission.
|
||
|
|
V3: A review item cannot become accepted knowledge without a separately validated proposal and the existing acceptance policy.
|
||
|
|
V4: User review decisions are attributable, idempotent, project-authorized, and immutable; corrections append a new decision.
|
||
|
|
V5: Only an explicit human choice promotes a review decision into reusable project guidance; guidance never overrides evidence, authorization, or acceptance checks.
|
||
|
|
V6: Every run reports selected, included, review-queued, excluded, and budget-deferred coverage with reasons.
|
||
|
|
V7: Review items and feedback from one project are never returned in another project's context without explicit authorized sharing.
|
||
|
|
V8: An optional Jev provider never becomes a source of evidence, an acceptance gate, or a requirement for storing review flags.
|
||
|
|
|
||
|
|
# §T
|
||
|
|
|
||
|
|
id|status|task|cites
|
||
|
|
T1|x|Define research-lens and ingest-review records, authorization, retention, and migration behavior|I,V3,V4,V7
|
||
|
|
T2|x|Add four default lenses to project creation/read/update and expose approved review guidance in `get_project_context`|I,V1,V5
|
||
|
|
T3|x|Revise `kb-ingest` and its pipeline to retain plausible cross-lens material for review and report complete dispositions|V1,V2,V6
|
||
|
|
T4|x|Add service-side review-flag creation, scoped listing, and authenticated human decisions, separate from proposal lifecycle and owner review|I,V3,V4,V7
|
||
|
|
T5|x|Add an explicit option to promote a human review decision into project-specific reusable guidance|I,V5
|
||
|
|
T7|x|Call optional OpenRouter Jev for relevance/lens probabilities and store advisory results without blocking ingestion|V2,V8
|
||
|
|
T6|.|Verify default lenses, review queue isolation, idempotent/attributable decisions, feedback application, no auto-publication, evidence integrity, and resume behavior|V1,V2,V3,V4,V5,V6,V7
|