Action Model / Resource rooms / cross-pack synthesis
rev 2 · no build
BR-04 / detailed resource room

First principles: provide the most client value with the fewest tokens.

This room expands the director synthesis from the Builder Research space. It reconstructs the problem below “recreate Lovable,” maps what the other lanes cover, and turns the missing pieces into a bounded evidence backlog.

8 irreducible problemscoverage matrixP0 / P1 / P2 backlogzero implementation receipt
01 / First-principles reconstruction

What problem are we really solving?

Not “make another prompt-to-app website.”

Atomic problem

How can agents and internet code provide the most useful client outcome as quickly and cheaply as possible, under a semi-custom brief?

That breaks into specification, design direction, retrieval, safe composition, execution, verification, approval, delivery, and learning. Each step has an evidence and authority boundary.

The loop the research points toward

  1. Guide the client into a structured brief without a rigid questionnaire.
  2. Design the direction before expensive implementation.
  3. Retrieve only when the task decomposition says evidence can help.
  4. Assemble from admitted scaffold/component contracts.
  5. Validate build, browser, visual, data, and license gates.
  6. Approve / revise as structured state, then ship.
02 / Eight irreducible problems

What remains unowned or under-proven

These are the reasons the pack stays a research program.

01

Requirements elicitation

Users omit implicit requirements; guided questioning can recover them, but no mature OSS benchmark or contract exists.

02

Design → token IR

Image-to-design-token extraction is underdeveloped. Pixels underdetermine type, spacing, radius, and layout; confidence/quantisation are required.

03

Retrieval timing

Always-on retrieval wastes tokens. Search should occur after task decomposition and use license-clean, AST-aware, bounded evidence packs.

04

Safe composition

Turning a repo into an installable block requires manifest, ownership, adaptation, dependency, and rollback semantics.

05

Backend foundations

Cheap models are most dangerous around auth, schema, migrations, billing, tenancy, and data isolation; scaffolds must own the contract.

06

Deployment / tenancy

Sandbox-green is not production-green. Previews, domains, isolation, credentials, and rollback are a separate subsystem.

07

Evaluation

There is no mature public benchmark for a “good app” across design, build, operation, and evidence. The validator harness is part of the product.

08

Guide agent / voice

Voice should narrate deltas while the screen owns the artifact and approval state; silence is not approval.

03 / Coverage matrix

What is solved, what is a substrate, what is open

Capability existence is not composition.

LayerExisting prior artAction Model research stateNext proof
UI generationLovable, v0, Bolt, Replit, screenshot-to-code, shadcn registries.Polished per-lane components exist; composition within a governed contract is open.One archetype, registry, design round, build and visual diff.
Data / CRUDTeable, Baserow, Payload, Frappe, Supabase starters.Many substrates; migration, tenancy, authority and client ledger boundaries remain.Read-only block first, synthetic fixtures, schema receipt.
Agent operationActionist docs, browser agents, CopilotKit, workflow engines.Long-tail/API-less outcome loop with approval/read-back/recovery is not proven end-to-end.One disposable app flow with side-effect denial and recovery.
Voice / guideWebRTC, LiveKit, Pipecat, realtime APIs.Transport exists; guided elicitation and structured approval protocol are open.Chat-first guide with optional voice narration and stateful approval.
EvidenceSPDX, ScanCode, SBOM, CI, browser/screenshot tooling.Ingredients exist; a single admission envelope joining them is the novel operating layer.One candidate with complete receipts, owner, and rollback.
04 / Bounded backlog

The work that should happen before scale

Prioritised, falsifiable, and explicitly no-build.

P0

Prove the contract

  • MODEL-EVAL: golden prompts + mechanical validators.
  • LICENSE-EXEC: execute the source/license/SBOM gate.
  • ACTIONIST-CONTRACT: authority, approval, recovery, evidence.
  • DEPLOYMENT: isolated preview → production distinction.
P1

Prove one bounded vertical

  • Choose finance, CRM, or operations archetype.
  • One data boundary and one read-only block.
  • One design direction and registry contract.
  • One synthetic fixture and screenshot baseline.
P2

Expand only after receipts

  • Per-industry corpus sweeps.
  • Write-side adapters and long-tail GUI actions.
  • Voice guide and marketplace distribution.
  • Factory scale, tenancy, and multi-client release.
Program posture: zero scan/build/eval/API receipts were claimed in the synthesis report. The correct next move is a narrow falsifiable pilot, not more undirected breadth.
05 / Evidence shelf

Open the director report and design principles

Local report copy is kept under the pack.

143 linesResearch synthesis rev 2

Sources, first-principles problem, coverage matrix, decision ledger, gaps, backlog, standing posture.

Open full synthesis ↗
designBuilder design principles

Guide agent, image-gen front end, interstitial retrieval, architecture, and phase-1 shape.

Open BUILDER-DESIGN.md ↗
principle 1Interstitial retrieval research

Retrieval timing, granularity, hybrid index, license gate, and evidence budget.

Open retrieval report ↗
principle 3Voice guide research

Chat-first elicitation, voice narration, screen-owned artifacts, and approval state.

Open guide report ↗
Local source readerfull report text on demand
Choose a source above. The local reader loads the original artifact.