Skip to main content

Intelligence categories

Eden ships 22 intelligence surfaces, grouped into four phases. This page is the catalogue.

Phase 1 — Observation (all shipped)​

The platform-level observation surface — every LLM call, every tool call, every user event becomes a queryable signal.

Cat 1 — Trace Replay & Workflow​

Reconstruct any trace from cold storage, replay it against the current prompt registry, diff against the original. Includes checkpoint rewind, live tail SSE, and per-step custom properties.

Tables: trace_replays, trace_replay_steps, trace_replay_diffs.

Cat 2 — RCA & Incidents​

Rule-based pre-filter, deterministic cluster keys (clu_<sha256[:16]>), Jaccard secondary similarity. Auto-fires incident.created webhooks on high-severity clusters.

Tables: incident_clusters, incident_alerts, incident_summaries.

Cat 4 — Retrieval Intelligence​

Utilization scorer, gap detector, chunk attribution, reorderer. Reads retrieval_logs and telemetry_events.response_body.

Tables: retrieval_logs, retrieval_gaps, chunk_attributions.

Cat 5 — Memory Intelligence​

Attributor, timeline builder, impact scorer, reweaver. Token-overlap influence + retrieval-score weighting.

Tables: memory_attributions, memory_timeline, memory_impacts.

Cat 6 — Agent Intelligence​

Agent performance scorer, relationship graph, handoff analyzer, benchmark runner, fleet service. Fires on_drift_alert(drift_type="fleet_deprecation:...") on identity re-registration.

Tables: agent_performance_daily, agent_relationships, handoff_log.

Cat 7 — APM Depth​

Sampling policy (5 keep rules), tail sampling buffer, K8s enrichment, Pyroscope continuous profiling.

Tables: apm_sampling_decisions, apm_k8s_enrichment_log, trace_step_profile_links.

Cat 8 — Cost Intelligence​

Efficiency scorer, ROI analyzer, anomaly detector (z-score), OLS forecast. Closed-loop hooks fire on cost.threshold_breach and cost.anomaly.

Tables: cost_summaries_daily, cost_anomalies, cost_forecasts.

Phase 2 — Quality & Evaluation​

Judges, scoring, and regression detection.

Cat 3 — Evaluation Quality​

13 services: deterministic judges, LLM judges, rubric management, regression detection, golden datasets, prompt management, eval runs, batch replay, and a shared helper. The largest cat.

Persists to all 14 Cat 3 tables + guardrail_policies (Cat 15 cross-cat write). Fires on_judge_regression and on_judge_threshold_breach.

Cat 9 — User Intelligence​

Hierarchical Beta-Binomial satisfaction correction (per arXiv 2605.12177), 6 v2 behavioural trackers (Regeneration, Edit, Return, Abandonment, CopyPaste, Retry), pure-Python Levenshtein. Fires on_drift_alert(drift_type="low_user_satisfaction").

Tables: user_sessions, user_behavioural_signals, user_satisfaction_corrected.

Cat 10 — Governance Intelligence​

4 framework evidence packs (SOC 2, HIPAA, GDPR, ISO 27001), Datadog-SDS-parity PII scanner, RTBF anonymisation, BFS lineage graph, dashboard rollup.

Tables: compliance_evidence_packs, data_lineage, pii_policies.

Cat 11 — Outcome / ROI Intelligence​

Per-department ROI rollup. 3 anomaly types (low_roi, cost_without_outcomes, roi_regression). Fires on_drift_alert(drift_type="low_roi_department").

Tables: outcomes, outcome_rollups.

Phase 3 — Experimentation & Prediction​

A/B testing and forecasting.

Cat 12 — Experimentation​

Traffic splitter (deterministic SHA-256 weighted bucket), auto evaluator, experiment recommender (Bayesian beta-binomial), winner detector (hard-judge majority + veto). Replay production in staging.

Tables: experiments, experiment_splits, experiment_assignments, experiment_results.

Cat 13 — Predictive​

Quality predictor, cost predictor, capacity predictor, risk scorer, drift detector. PSI drift on judge-score distribution, 64-bucket hash-histogram embedding drift.

Tables: predictions, drift_alerts.

Cat 18 — Query Performance / Columnar​

Per-org partitions (eden.<org>.<YYYY_MM>), hourly/daily/weekly rollups, 4h-SLA compaction, read-only SQL allowlist, Postgres-fallback.

Tables: managed by ClickHouse Sink + compaction workers.

Cat 19 — Trace Query DSL / EQL​

SQL-only EQL parser (no sqlglot), span search, field selection. Per-step edit DSL step_<n>.<key>.

Tables: read-only against telemetry_events.

Cat 20 — Time-Travel Debug​

Non-destructive replay, cache-miss warnings, per-step edit DSL, checkpoint rewind.

Tables: read-only against trace_replays.

Phase 4 — Autonomy​

Self-healing and code modification.

Cat 14 — Autonomous Intelligence​

Investigation agent, report generator, recommendation engine, self-healer, NL query engine, auto dashboard builder, NL report generator. Single-shot LLM calls, graceful degradation on LLM failure.

Tables: investigations, recommendations, self_heal_proposals.

Cat 15 — Guardrails​

Pre-flight check, runtime check, post-flight check. Fires on_pii_match, on_judge_threshold_breach, on_drift_alert(drift_type="compliance_gap:<framework>").

Tables: guardrail_policies, guardrail_decisions.

Cat 21 — Custom Metric Studio​

User-defined metrics with custom LoRA heads. Multi-tenant isolation, label-based slicing.

Tables: custom_metrics, custom_metric_lora_jobs.

Cat 22 — APM Depth + Tracing Depth (v1.4)​

5-keep-rule sampling, tail sampling buffer (15k slots), K8s enrichment (Downward API + SA JWT), Pyroscope continuous profiling.

Tables: same as Cat 7, but the v1.4 surface adds the Pyroscope-side trace_step_profile_links.

Cat 23 — Autonomous Code Modification (STUB surface)​

Patch proposals, sandbox validation, two-stage opt-in. STUB only — real end-to-end ship is gated on the 4 hard gates and the SLM Initiative. See the Code change service docs for the opt-in flags.

Tables: 19 tables in code_change schema.

Closed-loop hooks​

The intelligence services communicate via a single hook bus (promotion_hooks.py). Hooks fire on:

  • on_judge_threshold_breach (Cat 3)
  • on_judge_regression (Cat 3, Cat 12)
  • on_drift_alert (Cat 6, Cat 9, Cat 11, Cat 13)
  • on_cost_threshold_breach (Cat 8)
  • on_cost_anomaly (Cat 8)
  • on_compliance_gap (Cat 10)
  • on_security_event (Cat 10)
  • on_pii_match (Cat 10, Cat 15)
  • on_winner (Cat 12)
  • on_trace_completed (Cat 1)
  • on_incident_created (Cat 2)

Each hook can write to multiple destinations (DB tables, webhooks, Slack channels) — see Reference → Webhooks.

What's NOT in v1.7​

  • Cat 17 — Streaming Intelligence (planned v1.7.x)
  • Cat 16 — Code Generation Eval (planned v1.8)
  • The full Cat 23 real ship (planned v1.6 Phase 6, 8 weeks, gated on the SLM Initiative)

Read the per-category write-ups in docs/docs/observability/ for the implementation details (internal docs, excluded from the public site).