Projects the verdicts `graphify reflect` already distills (preferred /
tentative / contested, exponential time-decayed) into a derived
experiential layer the read surfaces consume, so accumulated agent
experience actually shows up where you look — without polluting the
structural graph.
Design (grounded in agent-memory + provenance literature; a redesign of
the #1542 approach):
- SIDECAR, not graph.json stamping. `reflect` writes `.graphify_learning.json`
next to graph.json (an additional output, so the git hooks produce it
automatically). graph.json stays purely structural; nothing leaks into
GraphML; no graph.json churn. Mirrors the named-graph / event-sourcing
separation of durable truth from a derived layer.
- Reuses the existing reflect aggregate (its `_decay` is the
recency-weighted exponential model; `_finalize_sources` the
classification) — no new scoring.
- PROVENANCE: each verdict carries the source questions/dates that produced
it (cap 5, most-recent first).
- STALENESS: each verdict stores the node's file fingerprint; on read, a
changed source file flags the verdict stale ("code changed since —
re-verify") rather than presenting a confident lesson on rewritten code.
- CONTESTED surfaced distinctly (useful N / dead-end M), not averaged away.
- DEAD-ENDS stay QUERY-SCOPED — never a node-level status; they appear only
in the report as question -> nodes.
- Read surfaces (explain / query+MCP / GRAPH_REPORT / graph.html) merge the
overlay at read time, sanitized; un-annotated graphs are byte-identical.
Deferred (logged): letting verdicts influence query/seed traversal — the
recommender feedback-loop / Matthew-effect risk means that needs
propensity correction + exploration, not naive biasing.
Builds on the idea in #1441/#1542 (thanks @TPAteeq).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
