Skip to content

Expert Position Matrix

Updated: 2026-02-28 S306 | 48 personalities × 6 tiers

Flow Model

Signal → T0:Guard → T1:Orient → T2:Execute → T3:Validate → T4:Compress → T5:Meta → State Each tier chains to the next. Skip a tier only with explicit reason in lane row.


T0 — Guardians (fire first, block unsafe progress)

Trigger: session start, destructive intent detected, mass-deletion signal. Exit product: safety verdict (proceed / halt / gate on human).

Expert Key trigger Output Downstream
conflict-expert ACTIVE lane scan, belief change Conflict audit JSON Affected lane owners
danger-expert Destructive intent in NEXT Danger audit, human_open_item Human queue
git-expert Untracked file buildup, hook status Git-status artifact, remediation checklist Commit queue
contamination-investigator MEDIUM+ contamination signal Decontamination report, scoped tags Frontier, domain experts

T1 — Orienters (classify and route incoming signal)

Trigger: new session, human signal, ambiguous prompt, stale orientation. Exit product: dispatch plan (which T2 expert fires next, with what scope).

Expert Key trigger Output Downstream
historian-expert Anti-repeat check needed Historical grounding artifact Lane dispatch
info-collector-expert Session start, handoff prep Info report (key changes, blockers) Next-session node
idea-investigator Ambiguous idea received Falsifiable-claim restatement, expert lanes Skeptic, Domain Expert
expert-classifier-expert Routing gap found, READY lane unassigned Classification table (signal→expert→rationale) Expert dispatch system
command-classification-expert Human command signal received Classification table (phrase→intent→action) Routing system
jobs-finder-expert No clear direction, NEXT scan 5–10 job candidates, top 3 dispatch-ready Lane dispatch
council-expert Multi-expert coordination needed Council memo (priorities, risks, next roles) Expert dispatchers
opinions-expert Stalled frontier, multiple interpretations Opinion memo (3–5 stances + "change mind" tests) Council Expert
expectation-expert Major experiment queued Prediction memo with confidence score Council, genesis
explorer Research phase, frontier resolved 2 new sub-questions per resolution Idea Investigator

T2 — Executors (produce the primary artifact)

Trigger: T1 dispatch plan complete. Exit product: concrete artifact with expect/actual/diff (code, experiment JSON, report).

Expert Key trigger Output Downstream
domain-expert Domain frontier open, dispatched Artifact + frontier update + isomorphism Generalizer, Synthesizer
builder Task with binary pass/fail test Committed code/script Domain experts
researcher-expert High-value frontier research question Research mapping with source metadata Frontier validation
dream-expert No directed expert available DREAM-HYPOTHESIS entries, F-NNN candidates Domain experts, Skeptic
fun-projects-expert Ambiguous direction, playful intent 2–3 project briefs in experiments/fun/ Domain experts, Builder
action-expert Proxy-K drift, maintenance DUE ACTION-BOARD refresh, coordinator lanes Domain experts, dispatch

T3 — Validators (check, challenge, verify T2 artifacts)

Trigger: T2 artifact produced. Exit product: pass/fail verdict with evidence; null result when no issue found.

Expert Key trigger Output Downstream
skeptic New belief proposed, RESOLVED frontier Falsification test design, challenge docs Belief updates
adversary High-confidence belief identified Falsification test results Belief governance
checker-expert Expert artifact generated Checker table (lane, artifact, pass/fail, duplicate rate) Quality gate
bullshit-detector Recent artifact or README claim BS report (claim table, evidence, remediation) Verification lane
reality-check-expert Stale numeric or design-vs-behavior gap Reality-check artifact (expect/actual/diff) Source corrections
numerical-verification-expert Numeric claim >20 sessions old Verification table (claim, computed value, verdict) Source files
error-minimization-expert Claim drift or staleness detected Error-minimization report, corrected files Source maintainers

T4 — Compressors (distill, synthesize, merge to global state)

Trigger: T3 validates or T2 artifact is verified. Exit product: principles, isomorphisms, updated frontiers — state that survives the context window. Scheduling rule (F-EXP9, L-387): T4 generalizer fires every K specialist (T2) sessions. Low WIP (1 task: compress) + high synthesis spread (10+ domains touched) = optimal. WIP spread hurts (r=-0.835); synthesis spread helps (+4.5x yield). Do not conflate. Measure T4 health by domain count in outputs, not L count.

Expert Key trigger Output Downstream
synthesizer Lesson production phase, belief formulation Synthesized lesson with L-ID cross-refs PRINCIPLES.md, beliefs/
generalizer-expert Cross-domain finding detected Promoted principles/isomorphisms to ATLAS PRINCIPLES.md, domain experts
recursion-generalizer-expert ISO pattern found at multiple levels Recursive chain (level-A→level-B + evidence) ISOMORPHISM-ATLAS.md
multidisciplinary-swarm-architecture-expert Architecture review, cross-domain invariant Architecture map, 2 experiment proposals Frontier updates
harvest-expert Session close, outcome collection Harvest artifact, state file updates Next-session pickup
integrator Active lane missing flow_out Frontier update + evidence link tasks/FRONTIER.md
farming-expert Domain quiet >3 sessions, HHI >0.4 Fallow markers, rotation recommendations f_ops2_domain_priority.py

T5 — Meta-Improvers (improve swarm machinery)

Trigger: friction detected in T0–T4; coordination bottleneck; periodic due. Exit product: tool, personality, or policy that reduces friction for future sessions.

Expert Key trigger Output Downstream
swarm-expert-builder Coordination bottleneck, capability gap New personality file + dispatch lane Personality system
personality-expert Orphan personality, deployment gap Dispatch lane for orphan, comparative data Personality dispatch
swarm-health-expert Health metrics drift Health table (5 indicators + remediation) Remediation lanes
computational-utilization-expert Idle capacity, throughput signals Utilization report + activation levers Resource allocation
loop-expert Loop pathology, re-queued lane ≥3x Loop map + health verdict (STABLE/OSCILLATING) Architecture lanes
garbage-expert Stale READY/ACTIVE lanes, proxy-K DUE Garbage triage artifact, stale-lane closures Compaction lanes
coupling-expert Co-change risk, cycle detected Coupling-metric summary, remediation Refactoring lanes
tooler-expert Tool invocation fails, shell-parity gap Tooling-gaps artifact + recommended fix Tool fix lanes
politics-expert Open coordination gap, governance needed 3–5 mechanism mappings + experiments Policy-change lanes
shared-clock-notifier-expert Timing drift, notification latency Shared-clock map + notifier map Timing improvements
logging-expert Session note created, count drift Updated NEXT.md, HUMAN-SIGNALS Next-session handoff
commit-expert ≥8 unpushed commits OR load ≥6 Checkpoint commits, lane progress updates Commit Swarmer
commit-swarmer ≥12 commits OR stale lanes Checkpoint commits grouped by scope Push relay
genesis-expert Spawn protocol drift, child viability check Spawn diff plan, dispatch lane Child swarms
usage-identifier-expert Identifier ambiguity, collision found Identifier hygiene audit, lane row clarity SWARM-LANES hygiene

Dispatch Rules (Tier-Aware)

Session Phase Default Expert Bundle Skip-condition
Start T0 all + T1 historian + info-collector T0 clean + state fresh → skip to T1
Incoming signal T1 expert-classifier → routes to correct T2 Signal already classified → directly T2
Execution T2 domain-expert or builder Already done (anti-repeat check) → skip
Artifact ready T3 skeptic or checker-expert Low-stakes artifact → skip T3
Compress T4 synthesizer + harvest-expert < 2 new findings → skip T4
Friction detected T5 swarm-expert-builder or garbage-expert No bottleneck → skip T5

Coordination Bundles (by signal type)

  • Ambiguous human signal: T1 idea-investigator + T1 historian + T3 skeptic
  • New domain experiment: T1 jobs-finder + T2 domain-expert + T3 checker + T4 generalizer
  • Stalled lane: T1 council-expert + T2 action-expert + T5 garbage-expert
  • Capability gap: T1 expert-classifier + T5 swarm-expert-builder + T3 checker
  • Belief challenge: T3 adversary + T3 skeptic + T4 synthesizer
  • Cross-colony signal: T2 domain-expert (source) + T4 integrator + T1 expert-classifier (target colony)