ContextLattice v3.15.0 - The Graph Earns Its Keep

ContextLattice v3.15.0 - The Graph Earns Its Keep

A graph is not intelligence because it stores edges. It earns that name when it
recovers evidence ordinary ranking misses, fits inside the prompt budget, and
proves the gain without hiding a regression. v3.15 closes that loop.

What changed

• Added contextlattice_memory_graph_repair, a CLI-first, dry-run-default repair
  lane with exact project confirmation and a hard cap on new edges per run.
• Reordered deterministic repair around the strongest identity signals:
  explicit references, session continuity, same-topic evidence, then opt-in
  inferred relationships.
• Made repeated repair batches scan past existing edges and report bounded
  forward progress instead of stalling behind an already-connected prefix.
• Added explicit graph-neighbor holdouts with separate direct seed and graph
  target expectations.
• Split direct-recall health from graph efficacy. Ordinary direct cases no
  longer dilute the graph denominator.
• Required graph targets to resolve to durable memory and hydrate into bounded,
  token-budgeted evidence. A dangling edge cannot pass the gate or enter a
  Context Pack as useful graph context.
• Added skill_retirement.v1 and contextlattice_skill_retire for terminal,
  immutable, non-destructive retirement of inactive Skill Foundry drafts.
• Serialized draft, evaluation, export, and retirement lifecycle mutations so a
  terminal retirement cannot race another public Foundry transition.
• Added interrupted-append recovery: a persisted retirement tombstone remains
  authoritative even if the process stops before the terminal draft snapshot is
  appended.

The boundary

• Repair writes are project-scoped and bounded. Inferred scoring remains opt-in.
• Graph efficacy does not call a model and does not treat an edge pointer as
  evidence.
• Direct recall must remain healthy when graph efficacy is evaluated.
• Draft retirement deletes nothing, installs nothing, and mutates no active
  runtime behavior.
• Retirement changes Foundry draft history only; it never uninstalls or deletes
  a separately installed skill.
• No MCP tool was added, so lifecycle hygiene and graph operations do not inflate
  every agent's tool surface.

Try it

Audit first, then apply one bounded repair batch:

contextlattice_memory_graph_repair --project my-project --pretty
contextlattice_memory_graph_repair \
  --project my-project \
  --write \
  --confirm-project my-project \
  --max-writes 500 \
  --pretty

Build graph-aware holdouts and require direct recall plus positive hydrated
graph lift:

contextlattice_memory_graph_efficacy \
  --refresh-cases \
  --project my-project \
  --graph-max-cases 3 \
  --pretty

Retire a temporary or superseded inactive Foundry draft without deleting its
proof:

contextlattice_skill_retire \
  --draft-id <draft-id> \
  --operator <identity> \
  --reason "temporary proof completed" \
  --pretty

The reproducible baseline, holdout design, economics, and live verification
record are in docs/evals/v3.15-graph-efficacy-foundry-retirement.json.

Proof, not graph theater

• Two scoped repair runs wrote 15,258 high-confidence deterministic edges,
  rejected 29,997 below-gate candidates, and reached zero remaining writes on
  an idempotence pass.
• The durable store restarted with all 39,495 edges loaded and zero policy
  skips.
• Direct recall stayed healthy at recall@5 1.0 and MRR 1.0. Three explicit
  graph holdouts then recovered three hydrated targets absent from direct
  top-K, producing measured graph lift 1.0.
• The Foundry smoke draft retired once, replayed idempotently without another
  write, and remained retired after restart. No evidence was deleted and no
  active runtime was mutated.