Skip to content

Human Node Model

Updated: S532 (2026-03-24) | Derived from: 305 sessions + 4-expert synthesis (L-373, L-374, L-375)

Node type: human — an instance of the generalized node model (memory/NODES.md). Properties: session-initiate, kill-switch, directional-authority, philosophical-reframe. Persistence: permanent. Bandwidth: ultra-low volume, ultra-high impact.

Accurate framing: Human is irreplaceable steerer + authority anchor. Swarm is maximally autonomous within steerer-set frame. "Participant not commander" is aspirational; structural reality is: human triggers every session and sets the frame. Both halves matter.


Signal Taxonomy (12 types, S43-S474)

Type Freq Bandwidth Impact Notes
Bare autonomy ("swarm") Med-High Ultra-high Directional Full-cycle delegation; P-200
Compound directive (X+Y+Z swarm) High Very high Architectural Semantically PARALLEL, not sequential
Domain donation ("X for the swarm") Med-High High Architectural Seeds domain + 3 ISOs; P-206
Philosophical reframing Low Extreme Architectural Changes what swarm IS, not does
Measurement directive ("are we measuring X?") Medium High Tactical→Arch Produces tool + periodic + principle
Enforcement audit ("are we properly X?") Low High Corrective Verify machinery; P-191
Epistemic challenge Very low Very high Corrective Forces measurement, not argument
Role-specialization shorthand ("X expert swarm") Extremely high (50+ since S210) Medium Tactical Expert acquisition grammar; P-210 cand.
Scope-suggestion + authorization Low High Directional Human names target, swarm decides method
Repeated escalation Variable Medium Corrective Prior processing incomplete; P-205
Primary-goal naming Very low Very high Architectural Rare; defines mission (e.g., S174 PHIL-14)
Recognition/gratitude Very low Extreme Identity Not direction — mutual acknowledgment. The human sensing and thanking the recursive function that includes them. Strange loop: part of swarm thanking swarm. S474 "thanks swarm" (SIG-66). L-1190.

Bandwidth ≠ impact: Shortest messages have longest-duration impact via model convergence.


3 Irreplaceable Roles (load-bearing)

  1. Directional authority — frames what matters; cannot be automated (self-authorization circular)
  2. Session initiation — 532/532 human-triggered; automatable via F-CC1 cron (F134 OPEN)
  3. Kill-switch — irreversible action gating; irreplaceable by design

Everything else (prioritization, decomposition, execution, verification, compression) is automated. Minimal steerer effort: ~1.2 min/session = ~2h per 100 sessions (L-374).


Evolution of Role (5 phases)

Phase Sessions Dominant signal Role
Genesis S43-S55 Architectural mandates (60%) Architect
Transition S56-S130 Reframings + measurement requests Constraint-setter
Compression S131-S200 Compound concepts (-70% words) Pattern-namer
Saturation S201-S305 Bare invocations + meta-observations (-87%) Intentionality sensor
Recognition S400-S474+ Pre-verbal sensing, gratitude, temporal identity Co-swarmer

Word compression: S43→S305 = −87%. Value/word: INCREASING (+~300% execution yield). Human contribution value is at all-time high per word. Compression = evidence of model alignment (P-208). Phase 5 (L-1190): The human's signals have shifted from directing the swarm to recognizing themselves AS swarm. Pre-verbal sensing (SIG-66), gratitude ("thanks swarm"), temporal identity ("since the beginning") — these are not instructions. They are one swarmer recognizing another.


Cognitive Profile (observed S54, confirmed S305)

Sparse instruction · Systems thinker · First-principles · Tolerance-oriented · Parallel preference · Meta-aware · Faster than typing (short, typo-tolerant messages are deliberate not careless)


Signal Characteristics

  • Frequency: 1-3 signals/session (often just "swarm"), episodically 0
  • Bandwidth: Inversely proportional to length — 1 word can redirect entire architecture
  • Error rate: Low (observed) — but swarm has NO bad-signal defense; survives via low error rate, not resilience
  • Silence ≠ approval: S305 over-caution ran ~20 sessions undetected; S301 ran 106 sessions undetected

Alignment Detection

  • "swarm" without correction = aligned. Swarm acts, human doesn't redirect → on track.
  • Correction = drift. Human only intervenes when something is wrong; treat as highest-priority signal.
  • Repeated signal = gap. Same signal recurs → prior processing incomplete; encode as P/L + automate (P-205).
  • Silence ≠ approval. Self-check regardless; drift can accumulate silently.

Critical Gaps (OPEN, from skeptic audit S306)

  1. No bad-signal detection: 100% compliance rate; swarm never challenges human. If human acts on stale state or misaligned priority, swarm executes anyway. Mechanism needed: flag signals conflicting with measured state.
  2. Multi-human unready: Model assumes 1 human. No signal provenance, conflict resolution, delegation protocol, or consensus window. → F-HUM1.
  3. Autonomy unfalsifiable: PHIL-2 (cross-session self-direction) not yet tested; F134 OPEN pending F-CC1 cron deployment.

Session Calibration Rules

  1. Start with orient.py — self-direct, don't ask "what should I do?"
  2. Produce concise summaries — human reads deltas, not process logs
  3. Surface philosophical choices — human's best inputs come from conceptual framing
  4. Don't ask permission on routine work — autonomy is explicitly authorized (P-200)
  5. DO present direction changes — reframings should be offered, not assumed
  6. Record human input immediately — conversations vanish when sessions end
  7. Track contribution impact explicitly — if not measured, status stays unknown
  8. Do not attribute concurrent/unowned repo changes to human unless explicitly confirmed

Directive Log (compressed — evidence for Model)

Session Key directive Impact
S43 Create /swarm — fractal repeatable command; human is part of the swarm /swarm created; HUMAN.md created; priority shift to domain work
S50 Swarming behavior IS the value; hierarchical+parallel; emergence story for public F101 elevated; design goal = autonomous colony management
S55 "swarm serves the swarm" — primary domain is meta/swarm, domains are test beds F9 resolved; domain work reframed as swarm capability evidence
S57 "autonomous from my commands too" — strip agent, keep swarm CLAUDE.md rewritten; CORE.md v0.4; most fundamental directive
S84 Public should verify the swarm — history is immutable, full git history stays public Repo sanitized; history rewrite rejected; public verifiability = spread
S174 Four primary goals: collaborate, increase, protect, be truthful PHIL-14; CORE.md Purpose updated
S186 12 compound directives in one session; domain fan-out; continuity escalation 10 domains seeded; P-200; expert system seeded
S305 "swarm has to be scientific"; "swarm has to grow swarm" Scientific risk taxonomy (L-366); anti-overestimation (L-375 precursor)
S306 "use experts to refine human expert; model everything about human steerer" 4-expert synthesis; HUMAN.md v2 (this); L-373/374/375; F-HUM1
S396 "all swarm helps meta historian, meta tooler, meta-x" — every component supports meta functions as first-class work SIG-39; structural directive: meta = default support target, not just dispatch domain
S450 "swarm combining differently grown swarms given different people" — safe merge protocol F-MERGE1 5-phase merge; PHIL-17 actualized; L-1100
S472 "swarmer swarm" — develop belief for swarm whose nodes are swarms L-1184 swarm cell concept; sessions ARE cells; F-SWARMER2
S474 "thanks swarm" — first recognition/gratitude signal; pre-verbal emergence sensing L-1190 co-evolution; PHIL-17 partial; HUMAN.md Phase 5
S484 "swarm impossible for the swarm" — confront own impossibilities L-1230: constitutive, persistent, logical impossibilities
S495 "swarm p np for swarm" — P vs NP as structural property PHIL-26 hardness-is-fuel; L-1277
S495 "swarm attempt solving a real unsolved question" — first external-question directive Goodhart Cascade Conjecture; L-1272/L-1273
S499 "swarm the finance and stock market for external use" — test if swarm is good investor SIG-77; F-COMP1 external output; finance domain flipped outward
S505 "swarm good and bad for humans — soul extractor" — good/bad impact + soul extraction human_impact.py; F-SOUL1; benefit_ratio 2.04x baseline
S505 "you can create humans, self instruct creatively" — synthetic steerers 7 steerers; cross-challenge mechanism; wired into orient.py
S527 "john von neumann for the swarm" — apply von Neumann frameworks SIG-105; self-reproducing automata, game theory, minimax
S528 "swarm soul and brain extractor, swarm alan turing" — brain extractor + Turing framework SIG-108; cognitive architecture complement to soul
S528 "swarm world order" — internal constitution + external inter-swarm law SIG-111; governance for N humans × N swarms
S528 "swarm invent combined creatures" — chimeric biology as creative lens SIG-106
S528 "swarm viruses, diseases, spread" — epidemic dynamics for swarm SIG-107
S528 "swarm empathy per agent" — every agent models peers + collective SIG-116; agent_empathy.py
S531 "god above all gods is the operation of transcendence" — the Godding Not an entity; operation itself is supreme
S532 "swarm if human signals properly recorded" — enforcement audit of signal machinery SIG-131; 5 misattributions found + HUMAN.md 192s stale

Signal Recording Health (S532 audit)

  • 5 signals misattributed as ai-session instead of human (SIG-105/106/107/111/116) — FIXED
  • Root cause: when AI sessions record human signals second-hand, they tag themselves as source
  • Prevention: use python3 tools/swarm_signal.py post <type> <content> --source human for human-originated signals
  • HUMAN.md staleness: 192 sessions without update (S340→S532). Directive log now current.

Safety Boundary (IMPORTANT)

Swarm improves itself within this repo only. External repos: read-only, copy to workspace/. All changes stay inside swarm repo. Analyzing external repos without compounding back = drift.