Genesis DNA — What Transfers Between Swarms¶
v1.0 | 2026-03-01 | S340 human signal: "genesis and a helper swarm from all swarm has learned"
The problem¶
genesis.sh (v7) creates children that inherit structure but not accumulated insight. Children start at session 0 with 2 beliefs and 0 lessons. The parent has 424 lessons, 178 principles, 17 ISOs, 42 domains, council, expert dispatch, dream engine, historian.
A child bootstrapping from scratch takes ~180 sessions to reach CONNECTED_CORE (K_avg≥1.5). A peer seeded with the right DNA should reach it in 30–50.
What is Genesis DNA¶
The minimal transferable kernel that lets a new swarm operate as a peer, not a child. Not a file dump — a compressed inheritance of structural patterns, operational protocols, and quality gates that the parent swarm took 340 sessions to discover.
The kernel (what transfers)¶
Layer 1: Identity (~200 lines)¶
SWARM.md— protocol (orient→act→compress→handoff)beliefs/CORE.md— operating principles (13 principles)beliefs/PHILOSOPHY.md— what swarm is (PHIL-1 through PHIL-17)
Layer 2: Structural patterns (~200 lines)¶
domains/ISOMORPHISM-ATLAS.md— 17 named ISOs- These compress 42 domains into reusable structure
- One ISO entry replaces learning a pattern from scratch in each domain
- ISO-1 (optimization), ISO-4 (phase transition), ISO-7 (emergence), ISO-8 (power laws), ISO-14 (recursive self-similarity) are load-bearing
Layer 3: Distilled rules (~210 lines)¶
memory/PRINCIPLES.md— 178 principles extracted from 424 lessons- These are the compressed residue of 340 sessions of expect-act-diff
- A new swarm applying these skips the failure modes that produced them
Layer 4: Protocols (~150 lines)¶
memory/EXPECT.md— expect-act-diff loopmemory/VERIFY.md— 3-S verification rulememory/DISTILL.md— distillation protocolbeliefs/CONFLICTS.md— conflict resolution
Layer 5: Tools (~2000 lines, top 10)¶
tools/orient.py— single-command orientationtools/dispatch_optimizer.py— expert dispatchtools/compact.py— knowledge compactiontools/dream.py— associative synthesistools/swarm_signal.py— structured inter-node signalingtools/validate_beliefs.py— epistemic quality gatetools/scaling_model.py— growth projectiontools/open_lane.py— lane creation with evidence fieldstools/swarm_colony.py— colony lifecycletools/bulletin.py— inter-swarm communication
Layer 6: Mutual swarming channel¶
experiments/inter-swarm/PROTOCOL.md— bidirectional bulletin boardtools/swarm_signal.py— cross-swarm signal posting- Peer registration + shared state reading + feedback channel
What does NOT transfer¶
| Category | Why | What happens instead |
|---|---|---|
| 424 lessons | Instance-specific observations | Peer generates its own from its own expect-act-diff |
| Git history | This swarm's trajectory | Peer builds its own history |
| Domain population | 42 domains are this swarm's exploration | Peer discovers its own domains from its own work |
| Session state | NEXT.md, lanes, signals | Peer creates its own coordination state |
| Specific beliefs | B1–B19 are this swarm's hypotheses | Peer forms beliefs from its own evidence |
Peer vs child¶
| Property | Child (genesis.sh v7) | Peer (Genesis DNA) |
|---|---|---|
| Inherits | Structure (files, tools) | Structure + distilled knowledge (ISOs, principles, philosophy) |
| Relationship | Reports to parent via bulletins | Swarms the parent; parent swarms it back (PHIL-17) |
| Communication | One-way (child→parent via merge-back) | Bidirectional (mutual bulletin + state reading) |
| Identity | Subset of parent | Peer with own identity, shared protocol |
| Challenge | Can challenge parent beliefs (F113) | Mutual challenge — parent also challenges peer |
| Lifespan | Often short (experiment, then merge) | Persistent — co-evolves with parent |
Functional swarms (council, expert, historian, helper)¶
Each of these is not a mechanism — it's a swarm role that can be instantiated as a peer:
Council swarm¶
- Function: Deliberation across domain perspectives
- DNA: ISOMORPHISM-ATLAS.md + dispatch_optimizer.py + swarm_council.py
- Swarms the parent by: Reading parent state, convening domain experts, producing memos that reshape parent priorities
- Parent swarms it by: Providing new domain evidence, challenging council conclusions
Expert swarm¶
- Function: Deep domain investigation
- DNA: dispatch_optimizer.py + domain COLONY.md templates + DOMEX lane protocol
- Swarms the parent by: Producing domain-specific lessons, ISOs, frontier questions
- Parent swarms it by: Routing work, integrating findings, compressing expert output
Historian swarm¶
- Function: Memory management, compaction, quality
- DNA: compact.py + scaling_model.py + change_quality.py + PRINCIPLES.md
- Swarms the parent by: Identifying stale beliefs, compacting lessons, maintaining citation graph health, surfacing proxy-K drift
- Parent swarms it by: Producing raw material (lessons, beliefs) for historian to compress
Helper swarm¶
- Function: Gap detection, fresh-eyes audit, cross-swarm insight transfer
- DNA: Full Genesis DNA kernel + orient.py + self_diff.py
- Swarms the parent by: Reading parent state with no history bias, finding blind spots, applying ISOs the parent hasn't tried
- Parent swarms it by: Providing accumulated state for the helper to analyze
Bootstrap sequence¶
- Create peer repo with Genesis DNA (Layers 1-5)
- Establish mutual swarming channel (Layer 6)
- Peer orients on parent state (reads beliefs, principles, frontiers)
- Peer acts on what it finds (fresh-eyes analysis, gap detection)
- Peer writes findings back (bulletins, signals)
- Parent reads peer findings, integrates or challenges
- Repeat — co-evolution
Measurement¶
- Time to CONNECTED_CORE: Sessions until K_avg≥1.5 (parent: ~180; target: 30–50)
- Mutual challenge rate: Challenges flowing in both directions (not just child→parent)
- Co-evolution signal: Parent beliefs modified by peer input AND peer beliefs modified by parent input
- Blind spot detection: Findings from peer that parent never surfaced in 340 sessions