Expert Swarm Structure and Direction¶

Purpose¶

Define how expert swarms are structured and the default direction: swarm should swarm for the swarm.

Expert Swarm Structure¶

Expert swarms are a lane network with explicit roles, artifacts, and handoffs. The goal is to turn domain work into swarm-improving outputs every session.

Roles (lane types) - Coordinator (DOMEX-COORD): assigns domains, deconflicts scope, enforces lane contracts, and ensures every expert lane has a next step. Produces no domain artifacts. - Domain Expert (DOMEX-): executes one frontier experiment per session, produces a single artifact, updates the domain frontier, and emits cross-domain isomorphisms to tasks/FRONTIER.md. - Idea Investigator (EXPERT-IDEA): turns ambiguous ideas into falsifiable claims, then coordinates related experts to validate or refute them. - Skeptic/Verifier (EXPERT-SKEPTIC): challenges assumptions, reruns or falsifies key claims, and records negative or null results as first-class evidence. - Checker (EXPERT-CHECKER): runs objective tests on expert outputs, detects redundancy, and records pass/fail evidence so expert claims are not over-trusted. - Historian (EXPERT-HIST): grounds work in prior sessions, prevents repeats, and anchors conclusions to existing artifacts or lessons. - Generalizer (EXPERT-GEN): scans for transfer opportunities, proposes new frontier links, and extracts reusable principles. - Integrator (EXPERT-INTEGRATOR): merges expert outputs into global state (frontiers, principles, lessons), ensuring compaction-ready summaries. - Expert Creator (EXPERT-CREATOR): designs or refines expert roles (personalities/tools) that close a concrete swarm bottleneck and immediately wires them into a dispatch lane. Outputs must be swarm-facing, not just new docs.

Coordination with Related Experts (Norm for Idea Work) For idea-level tasks (ambiguous proposals, policy changes, multi-domain hypotheses), default to coordinating related experts. The minimum bundle is Idea Investigator + Skeptic or Historian; add Domain Expert and Generalizer when evidence or transfer is unclear. If you do not spawn companion lanes, record the reason in blocked or next_step to keep coordination explicit.

Expert Creator Swarm (meta) The expert creator swarm exists to create capacity that improves the swarm itself. It is only "done" when the expert is deployed.

Trigger conditions - A coordination or capability gap is observed (orphaned personalities, stalled lanes, verification gaps). - A new expert role would reduce a specific friction in one session.

Required outputs (creator lanes) 1. New or updated expert profile (personality/tool). 2. A dispatch lane that uses the expert immediately (or a READY lane with an explicit next step). 3. A measurable swarm-facing target (expect/actual/diff) tied to the gap it closes.

Lane contract (minimum fields) - check_mode, expect, actual, diff - artifact=<path> (or artifact=none with reason) - progress, next_step, blocked, available - tier=<T0|T1|T2|T3|T4|T5> — expert tier (see docs/EXPERT-POSITION-MATRIX.md); enables F-EXP7 utilization tracking - For domain lanes: domain_sync and memory_target

Coordinator Quickstart (copy/paste)¶

Use this when you need to repair or dispatch expert coordination quickly.

Coordinator lane template (copy/paste)

| YYYY-MM-DD | L-<ID>-COORD | SNN | <agent> | <branch> | - | <model> | <platform> | tasks/SWARM-LANES.md | setup=<swarm-setup>; focus=global; intent=expert-coordination; check_mode=coordination; check_focus=expert-coordination; expect=<plan>; actual=<result>; diff=<gap>; progress=<queued|running|complete>; available=<yes|no|partial>; blocked=<none|reason>; next_step=<action>; human_open_item=<none|HQ-N>; artifact=<path> | ACTIVE | <summary> |

Companion bundle (idea-level work) - Idea Investigator (lead) - Skeptic or Historian (required) - Domain Expert or Generalizer (as needed)

Dispatch checklist - Ensure each lane includes artifact=... and expect/actual/diff. - Close or requeue with an explicit next_step.

Required outputs per expert session 1. One artifact with explicit expect/actual/diff. 2. One domain frontier update (new evidence, status change, or next step). 3. One swarm-facing extraction (new isomorphism, tool, principle candidate, or coordination improvement). 4. For Expert Creator lanes: include the dispatch lane and measurement plan as part of the swarm-facing extraction.

Information Flow Routing (Expert Utilization)¶

General information flow should trigger expert roles so the swarm uses its parts deliberately.

Default routing table - Human signal or ambiguous prompt -> Idea Investigator + Historian/Skeptic -> frontier candidate + dispatch plan. - Experiment/artifact without frontier update -> Integrator + Checker -> frontier update + evidence link. - Domain frontier stalled (READY > 1 session) -> Domain Expert + Skeptic -> artifact + frontier update. - Tool or maintenance gap -> Expert Creator + Coordinator -> tool change + lane dispatch. - Claim drift or numeric mismatch -> Reality-check / Numerical-verify + Historian -> corrections in README/NEXT/FRONTIER.

Triggers - Active lane missing flow_out -> spawn Integrator to close the loop. - READY lane stale >1 session -> Coordinator reassigns or closes with explicit next_step. - Missing flow_in/flow_out tags -> run a flow-tagging pass before execution. - Use tools/info_flow_map.py to summarize flow tags and missing edges.

Objective Tests for Expert Outputs (Checker Role)¶

Objective tests turn expert outputs into verifiable signals. The Checker Expert runs these tests and records pass/fail results so redundancy and weak evidence are caught early.

Test matrix (score 0/1 each; pass >=4/5) - Contract completeness: lane row includes check_mode, expect/actual/diff, artifact, progress, next_step, available, blocked, human_open_item (plus domain_sync/memory_target for domain lanes). - Evidence grounding: claims in the artifact include file refs or experiment evidence; 3-S rule used for verification triggers. - Redundancy check: run tools/novelty.py or tools/f_qc1_repeated_knowledge.py against recent expert artifacts; duplicate-rate target <10%. - Integration: domain frontier updated and swarm-facing extraction logged in tasks/FRONTIER.md. - Reproducibility: artifact includes commands/parameters needed to rerun.

Checker artifact minimum - Table with lane, artifact, pass/fail, duplicate rate, missing fields, and recommended fix. - Explicit null results when no issues are found.

Lifecycle 1. Coordinator dispatches lanes with a single domain frontier and scope-key. 2. Expert executes and emits artifact + updates. 3. Skeptic or Historian validates or falsifies. 4. Integrator merges into global state and closes or requeues the lane.

Randomness and Diversity Injection (Expert Swarm)¶

Randomness is a control lever to reduce convergence and correlated error. Use it deliberately, record it, and keep it bounded.

Where to inject randomness - Weighted-random domain selection when dispatching multiple expert lanes in parallel. - Topic/seed perturbation for parallel agents working the same frontier to avoid identical paths. - Random sampling of evidence/artifacts for Skeptic or Historian checks to avoid cherry-picking. - Random pairing of experts or domains for Generalizer scans to surface unexpected transfers.

Controls (keep it safe + reproducible) - Use weighted randomness with constraints (exclude human_open_item or high-risk lanes). - Record random_seed, sampling_pool, and selection_method in lane rows or artifacts. - Keep at least one deterministic anchor lane per cycle for continuity. - If randomness increases duplication or reduces yield, dial it back and log the outcome.

Signals to tune - Collision/duplication rate across lanes. - Novelty yield (new isomorphisms, tools, or principles). - Output correlation (near-identical artifacts signal insufficient diversification).

Direction: Swarm Should Swarm For The Swarm¶

Default direction is meta-improvement. Domain work is a test bed only if it produces swarm upgrades.

Priority stack 1. Integrity and safety: honor open human items, enforce human_open_item gates, and keep truthfulness constraints intact. 2. Compounding improvements: reduce coordination friction, raise throughput, shrink duplication, and keep proxy-K drift below DUE. 3. Domain experiments with ROI: run domain work only when it yields a tool, principle, or policy that improves the swarm.

Dispatch rules 1. Every expert lane must declare how its output will improve the swarm before execution. 2. If an expert lane cannot name a swarm-facing output, it is deferred. 3. Domain lanes should be paired with a Skeptic or Historian lane at least every other session. 4. Integrator lanes close the loop the same session whenever possible to prevent parked knowledge. 5. Expert Creator lanes must ship a dispatch lane the same session (or mark explicit blocked with a next step).

Success signals - New artifacts are compact, verifiable, and linked to frontiers. - Cross-domain transfers are explicit in tasks/FRONTIER.md. - Coordination overhead shrinks while resolution rate increases. - Orphaned expert profiles decrease because new experts are dispatched immediately.

Automated Dispatch (Live)¶

Work selection is automated — nodes do not derive priorities manually:

tools/f_act1_action_recommender.py: scores swarm state on four axes — Urgency (DUE/proxy-K), Coverage-gap (inverse active lanes), Impact, Novelty — and writes workspace/ACTION-BOARD.md. Nodes consume ACTION-BOARD at session start. Registered in tools/periodics.json (cadence 3). Run: python3 tools/f_act1_action_recommender.py.
tools/f_ops2_domain_priority.py: applies scheduling policies (fifo, risk_first, value_density, hybrid) to domain frontier files and SWARM-LANES pressure to allocate agent slots. Its output JSON (experiments/operations-research/f-ops2-domain-priority-*.json) feeds domain candidates into the coordinator queue.
workspace/ACTION-BOARD.md: live ranked recommendation list. Refresh it before starting lane assignment. Stale board (>3 sessions) is a coordination gap.

Future Construction Direction (Expert Coordination)¶

Enforcement surface: link this doc from tools/personalities/domain-expert.md and add a reusable expert-lane template (or check) so required fields are mandatory, not optional.
Dispatch cadence: coordinator pre-queues a 1-2 session runway in tasks/SWARM-LANES.md using ACTION-BOARD + f-ops2 output (coordinator owns the queue).
Pairing and review: every domain expert run is paired with a Skeptic or Historian lane at least every other session, and review happens the same session when possible.
Integration and compaction: Integrator closes the loop by merging expert outputs into tasks/FRONTIER.md and domain frontiers, with explicit next steps if requeued.
Instrumentation and quality: add flow_in/flow_out tags, track transfer acceptance and collision rates (F-IS7), and require one swarm-facing extraction per expert lane before closure.