---
source_hash: 9328b46a513bcd289777ce484be5fa8fac57bd3041d11ebb365e389398f72682
post: /posts/workflows/the-swarm-that-refused-to-choose
---

# Context Pack: The Swarm That Refused to Choose

Reference for the September 2026 Brainstorming Swarm V2 account. Public repository: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2. The method is MIT licensed; host support is a separate evidence question.

## Verbatim excerpts

> A reviewer liking the design could not supply the missing fact.

> Submitting a digest is not the same as accepting it.

> Sometimes that decision is to find out one more thing before choosing.

## The test and its correction

An invented fairground wayfinding campaign had two discovery lenses, two approaches and separate critiques. A desired route needed shade. The supplied facts established a staffed location and step-free route, not shade. A keep verdict did not make the condition true, so final selection remained empty.

An earlier internal synthetic decision inferred shade from a staffed location and was vetoed. The method then required decision dependencies: necessary conditions marked supported, unverified or contradicted against direct source anchors.

The public corrected run revoked an accepted discovery digest before drafting when it had inferred shade from a place name. A revised digest marked shade unverified. A material design change invalidated the old critique and required a fresh one. Shade and full pilot cost remained unknown; bounded checks were the next step. An independent artefact reviewer passed the corrected reasoning/trail, not native-host support. Earlier revisions and verdicts remain visible.

## Seven stages

Frame outcome/constraints/sources/privacy/stop point; route after checking actual host capability and relevant model/effort controls; discover with two or three distinct lenses; propose with assumptions/conditions/risks/effort; challenge with a different reviewer and fresh review after material changes; decide using only accepted work and supported dependencies; stop at the human gate. Design approval can lead to a PRD; review the written PRD before task planning. Implementation/publication remain separate decisions.

Keep source facts separate from inferences and the reviewer's explanation of their bearing. Each fact has a direct anchor; inferences name supporting facts. Bounded digests preserve a readable main conversation and approaches should not converge by seeing each other's answers prematurely.

## Files and validation order

Keep the whole skill directory: SKILL.md, references and templates. Without a skill loader, provide those files directly. Instructions do not confer permissions.

Campaign files: SCOPE.md, LEDGER.md, campaign.json, initially empty events.ndjson, digests/. Scope names boundaries; the role map records authors; the ledger points to current/earlier revisions; the journal records validation then acceptance.

Use python3 scripts/check_campaign.py path/to/campaign --leg g1 for the actual submitted leg ID and inspect its sources. Keep the leg incomplete while validating. Journal validation, then accept and journal acceptance. Run the --events check before dependent work, and the full campaign check at completion. File existence or valid structure alone is not acceptance.

The checker inspects files, links, required fields and recorded event order. It cannot prove truthful citations, independent authors, actual runtime barrier order or concurrent computation. solo_analysis is explicitly one agent's analysis, not independent peer review.

## Public release boundaries

The method/checker/examples/host guides were built in dependent waves and reviewed at their seams. Failure cases included missing judges, unsupported facts inside well-formed JSON and stale verdicts after revision. The public repository was assembled from reviewed files without private development history. Literal scans and separate infrastructure/identity reviews complemented one another; private-context reports stay private.

The support matrix dated 27 September 2026 lists Claude, Codex, Grok and Muse rows as candidate. The fairground audit did not prove exact child model identity, concurrent computation, cancellation or two-host-family support. Record actual serial/solo behaviour and stop when a required route cannot be verified.

## Sources and retained figures

- First Run: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2#first-run
- Method: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2/blob/main/SKILL.md
- Exact trial: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2/blob/main/evidence/codex-method-trial-2026-09-27.md
- Trial/review: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2/tree/main/evidence
- Support: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2/blob/main/SUPPORT.md
- Scrub: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2/blob/main/evidence/scrub-summary.md
- Licence: https://github.com/PUSHINGSQUARES/brainstorming-swarm-v2/blob/main/LICENSE

SwarmEvidenceGate, SwarmStageMap and SwarmRevisionTrail illustrate the source story, validation barriers and retained revisions. They do not run or certify a campaign.
