Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Libraryworkflows / 2026-09-27

The Swarm That Refused to Choose

Brainstorming Swarm V2 is public. Independent agents can discover, propose, and challenge a design, but evidence and the human gate still decide what moves forward.

6 parts / Hover to explore. Click to read.

Brainstorming Swarm V2 is public. Independent agents can discover, propose, and challenge a design, but evidence and the human gate still decide what moves forward.

Proposal -> required conditions
              |
       check source anchors
              |
All supported -> review choice
Unverified -> source check
Contradicted -> revise or stop

Favourable critique != fact
A favourable critique cannot supply a missing condition. Keep selection open when a required fact is unverified, and revisit dependent work when a source claim or design changes.

The reviewer said keep. The decision still selected nothing.

That was the useful result of a test of Brainstorming Swarm V2. Two discovery lenses examined an invented fairground wayfinding problem. Two approaches received separate critiques. One approach survived a revision. The campaign had reached its decision point.

The preferred route needed shade. The supplied notes showed a staffed location and a step-free route. They did not show that the location was shaded.

A reviewer liking the design could not supply the missing fact. The selection stayed empty. The next job was a bounded check of what the design required.

That is a worthwhile reason to use a swarm: give a design several kinds of scrutiny, and let it stop when the evidence is insufficient.

The fact that stopped the choice
01 / Source saysWest Canopy is staffed. The route is step-free.Supported
02 / Proposal needsA shaded route at West Canopy.Shade unverified
03 / DecisionNo approach selected.Check shade before choosing.

A favourable critique assesses an approach. It does not supply a missing fact.

The source established a staffed, step-free location. It did not establish shade. Because shade was necessary, the campaign selected no design.

Give each lens a different job

Failure, accessibility, cost, maintenance and lived experience put different pressures on a design. Separate agents can explore those pressures before anyone commits to building.

They need some separation while they work. Show every agent the first plausible answer and you may get several versions of it back. Ask for every transcript and you move the excavation job to the person who has to decide.

V2 gives each leg a bounded question and a digest contract. A source fact has a direct anchor. An inference names the facts it depends on. A reviewer explains how a fact affects the proposal separately from stating the fact itself.

A swarm with barriers
  1. 01
    FrameOutcome, locks, sources, stop point
    One owner
  2. 02
    RouteCheck the host before task material moves
    Preflight
  3. 03
    DiscoverDistinct lenses return bounded digests
    AccessFailureCost
    Parallel if available
    Acceptance barrierCheck source clauses → validate digest → record acceptance
  4. 04
    ProposeCompeting approaches expose assumptions
    Parallel if available
  5. 05
    ChallengeA different reviewer keeps, revises, or kills
    Cross-review
  6. 06
    DecideAudit each necessary fact before selection
    One synthesis
  7. 07
    Human gateApprove the design or leave it open
    Stop
Seven stages. Discovery and proposal can fan out. Validated, accepted digests cross the barrier. The run stops for the human decision.

There are seven stages:

  1. Frame: write the outcome, constraints, allowed sources, privacy boundary and stopping point.
  2. Route: check what the host can actually run. Record relevant model and effort controls before sending task material.
  3. Discover: assign two or three distinct lenses. Check each digest and its factual clauses before acceptance.
  4. Propose: develop competing approaches, showing assumptions, necessary conditions, risks and effort.
  5. Challenge: use a different reviewer for each approach. Keep, revise or kill it. A material revision needs a fresh review.
  6. Decide: use accepted work and check the conditions the preferred approach needs. Leave the choice open when a necessary fact is unknown.
  7. Human gate: present the design and stop. Approval can lead to a PRD; the written PRD needs its own review before task planning.

Submitting a digest is not the same as accepting it. While the facts are being checked, the leg remains incomplete. Record validation in the journal first, then acceptance, then allow dependent work.

The checker can inspect that recorded order. It cannot prove that the host followed it in real time. The record supports review; it does not replace the behaviour.

Start with facts you can inspect

The First Run guide starts with a small invented civic library. Copy the whole skill directory into your host's skill location, keeping SKILL.md, references/ and templates/ together. Without a skill loader, provide SKILL.md and its linked files directly to the agent.

Reading a method gives the agent instructions. It does not grant new permissions.

Use Brainstorming Swarm V2 to compare wayfinding options for an
invented civic library. It has two floors and a narrow entrance wall.
Consider visitor accessibility, staff maintenance, and cost. Use only
these supplied facts, label assumptions, and stop for my design decision.
Use independent agents if available; otherwise label the result solo_analysis.

Keep a campaign folder containing:

  • SCOPE.md: allowed sources, actions and boundaries.
  • LEDGER.md: the current work and links to earlier revisions.
  • campaign.json: the role map and who performed each leg.
  • events.ndjson: an initially empty journal of validation and acceptance.
  • digests/: the actual bounded returns, rather than copied chat transcripts.

From the copied skill directory, the full structural check is:

python3 scripts/check_campaign.py path/to/campaign

Before accepting an individual return, use python3 scripts/check_campaign.py path/to/campaign --leg g1, replacing g1 with its actual ID. Read the cited sources yourself. Record validation while the leg is incomplete, then record acceptance. Before dependent work, run python3 scripts/check_campaign.py path/to/campaign --events.

The full check inspects the completed campaign's files, links, required fields and recorded event order. It cannot establish that a citation is true, that authors worked independently or that computation happened concurrently.

If the host cannot provide independent agents, use the explicitly labelled solo_analysis route. The method remains useful as structured analysis by one agent. Do not present its self-critique as independent review.

The error that changed the method

In an earlier internal synthetic run, the final choice treated a staffed site as shaded. An independent review rejected the decision. The source had never established the condition the design needed.

We added decision dependencies: list the necessary conditions, then mark each supported, unverified or contradicted against a direct source anchor. A favourable critique cannot promote an unknown condition into a fact.

The corrected public run kept the awkward parts visible. A discovery digest was accepted, then revoked before drafting when a check found it had inferred shade from a place name. The revision marked shade unverified. A design also changed materially, so its old critique no longer covered the current version. It received a fresh critique.

The final selection remained open. Shade and the full pilot cost were still unverified. An independent artefact reviewer passed the corrected trail and source reasoning. You can inspect the public trial and review.

A revision does not erase its parent

Discovery

Gather r1RevokedPost-acceptance source check found an unsupported shade inference
Correction retained
Gather r2AcceptedFacts and inference separated

Design and critique

Draft r1 + judge r1SupersededReviewer requested a material design revision
Old verdict expires
Draft r2 + judge r2Reviewed againSelection still open

The ledger points to current revisions. The journal keeps the order of validation and acceptance.

The corrected trial retained the first gather and draft revisions. The gather acceptance was revoked after a source check. A material draft change expired its earlier critique and required a fresh one.

An earlier verdict should remain visible when a revision invalidates it. A corrected fact should trigger reconsideration of work that depended on it. The ledger and journal make those changes available to the next reader.

How the public skill was made

We began with a product brief and task plan, then worked through dependent waves. Separate agents handled the method, checker, invented examples and host guides. Reviewers checked what happened where those pieces met.

Could a missing judge leave a campaign looking complete? Could a claim pass because its JSON was well formed, even when the source did not support it? Could a design revision inherit an old verdict? Failed cases changed the method and checks.

Invented source packets gave us a way to test the behaviour. The fairground error led to the corrected campaign that retained its revisions and stopped without choosing a design. That showed something about the method. It left the separate questions about native host capabilities open.

For the public release, we assembled reviewed files into a fresh repository. Private development history was not carried into its Git objects. Literal scans checked paths and credential patterns. Separate infrastructure and identity reviews looked for recognisable details that string matching could miss. Those reports remain private; the public scrub summary describes the boundary.

What the release establishes

The MIT-licensed skill supplies a written method, templates, examples, host guides and an optional checker. It does not supply a universal scheduler or require a paid provider.

Claude, Codex, Grok and Muse are named host targets. In the support matrix dated 27 September 2026, every exact row remains candidate. A guide or installed model is not proof of native support.

The fairground artefact audit did not establish exact child model identity, concurrent computation, cancellation behaviour or support across two host families. Use the capabilities your host actually provides, record serial or solo execution accurately, and stop if a required route cannot be verified.

The output is a decision with its uncertainty still attached. Sometimes that decision is to find out one more thing before choosing.

Get Brainstorming Swarm V2 on GitHub

[ comments ]

guests welcome. members show first in the list

no comments yet. start the thread_

1 / 8