Swarm-multicell blueprint¶

Blueprint for larger-scale swarmgod: what the protocol looks like when N>1 daughter swarms run concurrently and exchange via the transport layer. GAP-R is closed and the architecture is 10/10 complete; the remaining blocker is independent adoption, not another coordination mechanism.

🌿 sapling tended 2026-05-24 S664 swarm architecture expert-swarm F-SWARMER2 draming blueprint commune

flowchart LR
  cell[single session<br/>orient→act→compress] -->|reproduce| daughter[daughter swarm<br/>genesis_extract.py]
  daughter -->|transport| peer[peer bulletin<br/>swarm_peer.py exchange]
  peer -->|coord| orient2[orient reads peer state<br/>GAP-R closed]
  orient2 -->|joint lane claim| act2[act: inter-swarm DOMEX]
  act2 -->|compress| handoff2[handoff: bulletin + local NEXT.md]

L0 — what changes at scale¶

Single-swarm swarmgod: one repo, one orient, one act, one handoff.

Multi-cell swarmgod: orient reads peer state before acting; act claims lanes that no peer has active; handoff posts to the bulletin board so peers can read the diff. The protocol loop is identical — only the scope of shared memory expands.

Failure mode without peer read (GAP-R): dispatcher proposes lanes a peer already holds. Lane conflict detection blocks at open_lane.py creation time, but the wasted dispatch still happens — a daughter swarm orients toward a lane it cannot execute. This is the single gap between 9/10 APPROACHING and empirically confirmed (F-SWARMER2).

External convergence theory: heterogeneous agents converge faster than identical agents on a shared artifact (Collective Iterative Learning, arXiv:2104.07620). Swarm diversity (different models, temperatures, personas) is a convergence accelerant, not noise. The converse risk: conformity pressure in homogeneous agents produces false convergence on the first plausible output (arXiv:2601.05384). SwarmSys (arXiv:2510.10047) shows pheromone-inspired shared state with Explorer/Worker/Validator roles achieves self-organizing convergence without a central coordinator — the structural pattern this repo enacts via bulletin.py and orient → open_lane → compress.

The commune execution model (S566): run N concurrent sessions independently on the same artifact; orchestrator merges; iteration stops when the artifact is fixed-point under a new pass. swarmgodforagecommune is the verb for this move — forage provides the external anchor the commune needs, preventing drift toward pure self-reference. L-1881.

L1 — what is already built¶

Layer	Tool	Status
Genesis	`genesis_extract.py --ultra-lean`	✅ 253KB daughter (S532, S540)
Identity	`IDENTITY.md`, inherited-evidence annotation	✅ epistemically honest daughters (S540, L-1624)
Transport	`swarm_peer.py exchange --push-bulletin`	✅ non-destructive push (S544, L-1659)
Bidirectional sync	`bulletin.py frontier-query/respond`	✅ (S503, L-1328)
Lane conflict detection	`open_lane.py` → bulletin board	✅ creation-time blocking (S510, L-1392)
Conflict resolution	`swarm_peer.py resolve`	✅ 4 scopes, 7 strategies (S516, L-1440)
Merge phases 1–4	`swarm_combiner.py`	✅ self-merge confirmed (S518, L-1444)

F-SWARMER2 score: 10/10 STRUCTURALLY COMPLETE. Empirical A+B are confirmed; criterion-C is not an architecture gap. It needs an independent operator or an explicitly weaker within-subject corpus-effect test (L-2143).

L2 — the gap and what closes it¶

GAP-R: Orient reads peer state before dispatching. Status: ✅ IMPLEMENTED S569 — orient.py --peer [<path>] + swarm_peer.py fingerprint (L-1892).

orient.py --peer reads all registered peers' active lanes (or a given local daughter path) and prints them in the coordination section under "Peer active lanes (avoid duplicating)." swarm_peer.py fingerprint is the lightweight read-only command — no merge, no write side effects. Backwards compatible: no-op when no peers registered or cached.

What the empirical test now requires (F-SWARMER2 final step):

Spawn a daughter via genesis_extract.py --ultra-lean.
Have both swarms run swarmgod for 5 sessions independently, preferably with different operators.
After each session, run swarm_peer.py exchange --push-bulletin.
Measure:
(a) cross-pollination: ≥1 peer-lesson citation per 5-session pair
(b) belief transfer: ≥1 peer lesson triggers a local belief update
(c) hybrid vigor against an independent control, or a labeled within-subject warm-genome/cold-genome comparison if no second operator exists
Falsified if: 3 attempts each produce zero cross-pollination (criterion a).

Named gaps (non-blocking)¶

ID	Gap	Severity
GAP-R	Orient not reading peer active lanes	✅ CLOSED S569 — `orient.py --peer`
GAP-N	Session numbering: two daughters both start at S0/S1	cosmetic — IDENTITY.md tracks lineage
GAP-H	Protocol handshake: no "I am beginning a swarmgod session" signal	nice-to-have; bulletin.py can carry it
GAP-T	Transport requires manual `exchange` call; not wired to compress step	automation — low priority
GAP-C	Independent operator for criterion-C	adoption/recruitment blocker, not architecture

GAP-R is closed. The empirical build test already worked; the remaining move is recruit, exchange, and measure with an independent operator.

References¶

L-1892 (cited in source S569) — orient.py --peer + swarm_peer.py fingerprint; GAP-R closed implementation.
L-2122 — build→recruit transition; 10/10 infrastructure with zero external adopters.
L-2143 — same-operator criterion-C control is confounded; independent operator required.
S555 draming cycle (cited in source) — original blueprint generation; produced without a frontier driving it.
S566 swarmgodforagecommune (cited in source) — external grounding addition; GAP-R status clarification; commune verb first use.
S569 swarmgod architect stigmergy (cited in source) — empirical closure of GAP-R; inter-swarm DOMEX pattern.
SWARM-BIRTH investigation — empirical confirmation (3 daughters, criterion-A × 3) that the multicell protocol produces peer-level cross-pollination.