P0, P1, and P2 must be complete. Verify:
[ ] mix benchmark.capability_spec — 80/80 pass
[ ] LongMemEval SHR > 90.4%
[ ] kappa > 0 on > 15% of benchmark queries
[ ] All tests pass, goal progress at 0.85
Implement P3 capabilities: causal edge metadata prep and GraphMemBench Phase 3 (full 200-scenario suite with competitor adapter stubs). Also: finalize the goal, update the spec, update portfolio-review.html with new benchmark numbers.
Retrieve and complete: goal_ae39998e19f363c1519679b15d24744d
Extends :causal edge metadata to support future CausalEngine (deferred until TickTickClock/WebHost ship). No behavioral changes — just ensuring the data substrate can store causal strength, confounders, and intervention history.
Uses existing edge metadata JSON field:
{
"causal_strength": 0.75,
"confounders": ["node_id_1", "node_id_2"],
"intervention_history": [
{"at": "2026-04-10T...", "action": "do(X=0.8)", "result": "Y increased"}
]
}graphonomous/lib/graphonomous/learner.exWhen learn_from_outcome updates a causal edge, also compute and store causal_strength:
causal_strength = case outcome.status do
:success -> min(1.0, old_strength + 0.1)
:partial_success -> old_strength # No change
:failure -> max(0.0, old_strength - 0.15)
:timeout -> max(0.0, old_strength - 0.05)
endSuccessful outcome strengthens causal edge
Failed outcome weakens causal edge
causal_strength persists in edge metadata
confounders field accepted and stored
Category 11: Causal Metadata (8 scenarios)
causal_strength updates on outcome success
causal_strength decreases on failure
confounders stored in edge metadata
Causal subgraph extraction (all causal edges from a node)
Multiple causal parents with different strengths
causal_strength survives consolidation
Zero-strength causal edges flagged for review
Causal edge creation during learn_from_outcome
Category 12: End-to-End Workflows (8 scenarios)
Cold start: empty graph → ingest 50 nodes → retrieve → verify quality
Knowledge evolution: 10 sessions with changing facts → verify revision chain
Skill accumulation: store 5 procedures → retrieve by precondition → verify transfer
Goal lifecycle: create → learn → act → complete with full epistemic coverage
Contradiction resolution: 3-way conflict → consolidation → single winner
Memory lifecycle: ingest → learn → forget → verify no data leak
Attention-driven exploration: survey → identify gaps → learn → verify improvement
Full v0.3 pipeline: all capabilities active, 100-node graph, measure latency
Category 13: Regression Guards (8 scenarios)
Existing retrieval quality preserved (F1 >= 0.415 baseline)
Existing consolidation throughput preserved (>= 27M nodes/sec)
Existing topology accuracy preserved (100% SCC detection)
Existing learning loop integrity (4/4 outcome tests)
Existing goal lifecycle integrity (4/4 lifecycle tests)
kappa=0 fast path latency not degraded by expanded topology window
Embedding quality unaffected by new columns
MCP tool response format backward-compatible
Category 14: Competitor Adapter Stubs (8 scenarios)
Baseline adapter: full-context LLM (stuff everything in prompt)
Mem0 adapter stub: simple vector search + LLM
Zep adapter stub: temporal graph search
Hindsight adapter stub: 4-way parallel retrieval + RRF
Adapter interface contract validation (all adapters implement same interface)
Cross-adapter result format compatibility
Adapter latency measurement framework
Adapter accuracy measurement framework
Category 15: GraphMemBench Reporting (8 scenarios)
JSON output format validation
Learning curve plotting data (accuracy at 5 checkpoints)
Category-level accuracy breakdown
Latency percentile reporting (p50, p95, p99)
Comparison table generation (Graphonomous vs baselines)
Confidence calibration measurement
kappa distribution histogram data
Deliberation accuracy measurement (kappa>0 answers vs kappa=0 fast path)
defmodule GraphMemBench.Adapter do
@callback ingest(session :: map()) :: :ok | {:error, term()}
@callback retrieve(query :: String.t(), opts :: keyword()) :: {:ok, [result()]}
@callback forget(node_id :: String.t()) :: :ok | {:error, term()}
@callback stats() :: map()
endGraphonomous adapter wraps MCP tools. Competitor adapters are stubs that define the interface but return {:error, :not_implemented} — actual integration comes when/if competitors provide APIs.
File: `graphonomous/docs/spec/README.md` Add section for v0.3 capabilities:
Belief revision with provenance chains
Conflict-aware consolidation (Stage 4.5)
Two-phase retrieval (semantic + utility)
Budget-aware forgetting with GDPR hard delete
Scoped uncertainty propagation
Procedural metadata + precondition matching
Multi-agent schema prep
GraphMemBench (120 scenarios, 15 categories)
File: `opensentience.org/docs/spec/OS-E001-EMPIRICAL-EVALUATION.md` Add v0.3 benchmark results:
kappa activation rate (target: >15%)
Two-phase retrieval SHR delta
GraphMemBench pass rate
Forgetting precision/recall
New tool count (22 → 28 MCP tools)
File: `AmpersandBoxDesign/site/portfolio-review.html`
Update KPI card with new SHR
Update benchmark scorecard
Add v0.3 capability summary
Update competitive comparison with new capabilities
Note: kappa activation is now validated, not just theoretical
manage_goal operation=set_progress goal_id=goal_ae39998e19f363c1519679b15d24744d progress=1.0
manage_goal operation=transition_goal goal_id=goal_ae39998e19f363c1519679b15d24744d status=completedlearn_from_outcome action_id=v03_continual_learning status=success
causal_node_ids=[all knowledge nodes linked to goal]
evidence="GraphMemBench 120/120, kappa activation >15%, SHR improved"[ ] mix compile --warnings-as-errors passes
[ ] mix test — ALL tests pass
[ ] mix benchmark.capability_spec — 120/120 scenarios pass (Phase 1 + 2 + 3)
[ ] mix benchmark.longmemeval --split oracle --neural --limit 100 — SHR > 90.4%
[ ] kappa > 0 on > 15% of benchmark queries
[ ] Spec updated with v0.3 capabilities
[ ] OS-E001 updated with new benchmark data
[ ] Portfolio review updated with new numbers
[ ] Goal completed with outcome learning
[ ] Consolidation run to clean session knowledge
[ ] Version bumped to v0.3.0 in mix.exs + npm/package.json