Action Model / Resource rooms / platform capability
checkpoint 009 · partial
AM-PLATFORMS / detailed resource room

What the platforms can do — and what none of them prove.

This room expands the platform lane instead of summarising it into a card. It carries the first-party dossiers, local reconciliations, admission experiments, contradiction ledger, and the current boundary between documented capability, strategic inference, and implemented proof.

7 documented dossiers0 authenticated/live1 active blockerBlock Contract not authorised
01 / Current state

The actual lane posture

Source: status.json + checkpoint-009

9immutable checkpoints
7catalogue / documented dossiers
34held / rejected claim groups
0authenticated live dossiers
1held admission blocker
Headline

The build plane is converging. The actionable seam is authority, recovery, and evidence.

The lane compared Manus, Airtable Omni, Zapier Forms, Base44, Onlook, NocoBase, CopilotKit, v0, and Lovable. The evidence supports a constrained client build plane plus an agent-operation and evidence plane as a product direction. It does not establish a live Actionist implementation.

Do not collapse these states

  • Documented means a first-party URL and an observation date were recorded.
  • Authenticated/live is zero because no credentialed product session was available.
  • Implemented is zero because no Actionist runtime code was changed by this lane.
  • Candidate is not admitted; Horizon and screenshot-to-code remain held.
02 / First-party dossier set

Seven platform families, read individually

These are research subjects, not endorsements.

SubjectWhat the lane establishedAction Model readState
ManusAgentic execution and builder convergence were treated as a reference point for the “agent + app” category.Benchmark the combination, not just prompt-to-app generation; ask where authority and recovery are explicit.documented
Airtable OmniConversational data/workflow construction shows how business context can be the product surface.Vertical context and structured data are more defensible than generic generation.documented
Zapier FormsForms and workflow primitives connect intake to automation without asking the user to become a developer.Own the job-to-outcome loop; the form is only the front door.documented
Base44Prompt-to-app plus hosted execution illustrates how fast a broad builder can converge.Hosting and generation are available category capabilities, not the wedge.documented
OnlookVisual editing and design-to-code surfaces were compared with the image→token→registry proposal.Visual control is useful, but a production contract still needs registry, build, and evidence gates.documented
NocoBaseLow-code internal systems are a plausible substrate, but edition/license boundaries remain unresolved.Do not convert “low-code fit” into a reusable-block admission.held
CopilotKit / v0 / LovableAgent/product primitives, context injection, registries, and builder APIs were reconciled with the local design.Adapters can inform the build plane; none is the Actionist authority/evidence contract.partial
03 / The useful abstraction

Three planes, not one “AI builder” blob

The cross-lane comparison that survived review.

01 · build

Constrained client build plane

Known schema, known deployment boundary, registry-backed UI, bounded model edits, and a visible artifact trail.

  • Prompt → structured spec
  • Design direction → token contract
  • Scaffold → bounded adaptation
  • Build, smoke, visual proof
02 · operate

Authority / operation plane

The layer generic builders do not close: actions over long-tail or API-less software, with approvals, side-effect limits, retries, and recovery.

  • Typed authority boundary
  • Ask-human checkpoints
  • Idempotency and read-back
  • Rollback or compensating action
03 · evidence

Evidence / registry plane

Every source, transformation, dependency, license, test, visual baseline, owner, and release decision is inspectable.

  • Provenance and SPDX
  • Block manifest and SBOM
  • Decision ledger
  • Admission / quarantine state
Strategic conclusion: the three-plane wedge is supported as a product direction, unverified as a capability, and not claimed as implemented. That distinction is repeated throughout the original packets.
04 / Work history

What happened across the nine checkpoints

Immutable evidence packets remain available.

001

First-party priority dossiers: Manus, Airtable Omni, Zapier Forms, Base44, and Onlook; evidence levels and source/date rules established.

Open packet ↗
002–003

Local reconciliation and additional platform surfaces, including NocoBase/CopilotKit boundaries and capability-matrix repair.

Open packet ↗
004

Image-to-design-token gap verified: no local OSS implementation was found; the proposed token contract is novel but still open.

Open packet ↗
005

Screenshot-to-code candidate admission gate written; source/license fit was not enough to admit it.

Open packet ↗
006–007

v0/Lovable API and white-label questions reconciled; the lane separated documented context injection from held OEM/deployment claims.

Open packet ↗
008

Adversarial platform comparison produced the three-plane model and explicit holds around authority, recovery, evidence, licensing, and cost.

Open packet ↗
009

Control-plane repair and cross-lane contradiction pass corrected the tracker to one blocker and 34 held/rejected claim groups.

Open latest ↗
05 / The unresolved experiment

P-010 / screenshot-to-code admission

Candidate is held, not rejected.

Why it was selected

The lane saw a plausible path from screenshot-to-code into a registry-backed frontend assembly loop. The candidate remains valuable as a learning target because it tests the boundary between visual intent, token extraction, scaffold selection, and bounded adaptation.

Why it is not admitted

  • No isolated stack adaptation receipt.
  • No token-slot contract tied to a real registry.
  • No deterministic build or browser smoke exit.
  • No visual baseline or screenshot-diff receipt.
  • No owner, rollback, dependency, or asset-scan receipt.
Current verdict

Block Contract pilot: NOT AUTHORISED.

The research did not fail because the idea was bad. It stopped because the evidence ladder was incomplete. That is the point of the admission framework.

source/license · openadaptation · openbuild/smoke · openowner/rollback · open
06 / Evidence shelf

Read the packets, not just this interpretation

Local source links are locked on the public host.

currentAM-PLATFORMS current state

Latest posture, counts, reconciliation, evidence boundary, next queue.

Open CURRENT.md ↗
statusStructured lane status

Machine counters, blocker, dossier state, verification flags.

Open status.json ↗
queuePlatform queue

Original bounded work queue and stop rules.

Open queue.md ↗
deep divePublic-source Action Model deep dive

The separate public ecosystem dossier with product, docs, socials, risks, and sources.

Open deep dive ↗
frameworkBlock Contract design

The admission framework that the platform lane keeps referring back to.

Open framework ↗
full historyAll checkpoint files

Checkpoints 001–009 are immutable; this room only surfaces their path.

Open checkpoint index anchor ↗
Local source readerchoose a packet to load its full text
Choose a packet above. On the local server this reader loads the original file, not a summary.

The public page keeps the interpreted room visible but does not expose the raw source tree. Local source links work when this pack is served from the Action Model directory.