/method/closing-the-loop

Closing the Loop

What Closing the Loop means

Closing the Loop is the name for the synthetic, end-to-end path that connects an external agent session to a review candidate that an operator can traverse, without ever letting the agent session write an admitted record. The path runs through five named systems — the ARCS Anchor scenario driver, the ARCS Amnesiac bridge, the GARP local scenario sink, the GARP Workbench Scenario Traversal Cockpit, and the GARPedia public projection — and stops at the public projection boundary. The grammar is the operating claim of GARP, applied to AI-assisted input: GARP separates AI assistance from institutional record admission.

The loop is "closed" in the sense that an outside agent action has an observable destination on an operator surface and, eventually, a public projection — not in the sense that the agent has produced a record. Generated text is not a record. Agent outcomes are review candidates, not admitted records. Receipts and review candidates do not verify model output. The seam between an AI-assisted outcome and an admitted record is the whole point of the loop, and the loop is built so that seam stays legible at every hop.

This page describes the loop's shape, what it proves, and — equally important — what it does not prove. It does not authorize any new public surface, does not lift any renderer authority hold, and does not introduce a new admitted record.

The synthetic loop, in order

The loop is described as a strict sequence so each hop's owner is named. Every hop is synthetic on this surface: a fictional scenario id (currently SYN-FSI-MNPI-0001) drives the path, and no real agent session, real tool call, or real operator material is rendered.

external agent sessionAn external agent runs a scenario driven by the ARCS Anchor synthetic scenario driver. The agent is outside the institutional record admission boundary; its session produces tool calls, not records.
tool-call dispositionEach tool call has a disposition recorded against the scenario. The disposition is a structured outcome of the call, not a claim about the underlying subject matter.
connector receipt consistencyThe connector emits a receipt for the tool-call disposition, and consistency between the disposition and the receipt is checked against the scenario. A receipt is an artifact of the call, not a verification of the model that produced the call.
AgentOutcomeObjectThe ARCS Amnesiac surface assembles the disposition and the receipt into an AgentOutcomeObject — a structured record of what the agent did, scoped to that scenario. An AgentOutcomeObject describes an agent outcome; it is not an admitted record about the world.
CandidateClaim bridgeThe AgentOutcomeObject is bridged into a CandidateClaim. The bridge is the explicit boundary between "this is what the agent produced" and "this is a candidate that might enter review". Crossing the bridge does not admit the claim; it routes it.
scenario_id correlationThe GARP local surface correlates the CandidateClaim back to the scenario by scenario_id and writes it into a synthetic scenario sink. The scenario_id is the link that lets an operator follow the trail back to the originating session.
Workbench scenario traversalThe GARP Workbench Scenario Traversal Cockpit lets the operator owner walk the scenario: which session, which tool calls, which AgentOutcomeObjects, which CandidateClaims. The cockpit is a review surface, not an admission surface; it does not edit a record in place.
GARPedia / public projection boundaryGARPedia is the public projection of records that the upstream custody and render owners have already chosen to project. The loop terminates here as a read surface. There is no path from a CandidateClaim into a rendered GARPedia page on this surface, and there is no agent write into GARPedia at all.

What the loop proves

That an external agent session can reach an operator review surface, end-to-end, without writing through the renderer or the admitted pool.
That the seam between an AI-assisted outcome and an admitted record is observable at every hop: tool-call disposition, connector receipt, AgentOutcomeObject, CandidateClaim, scenario sink, and Workbench cockpit.
That scenario_id is a working correlation key for traversal across the synthetic path, including in the Workbench Scenario Traversal Cockpit and the GARPedia projection cockpit contract checks the garp-ops Closing the Loop verifier exercises.
That the garp-workbench check:demo step now verifies scenario traversal and the GARPedia projection cockpit contracts, so regressions in the demo path surface mechanically rather than by inspection.

What the loop does not prove

Generated text is not a record. The presence of a CandidateClaim in the scenario sink does not make the underlying agent output a fact about the world.
Agent outcomes are review candidates, not admitted records. An AgentOutcomeObject describes what an agent did; reaching the candidate/admission seam does not promote it.
Receipts and review candidates do not verify model output. A connector receipt proves a call happened; it does not adjudicate the model that produced the call.
scenario_id is a correlation key, not an admission key. Sharing a scenario_id with an admitted record would still not make the candidate admitted.
No model-truth claim is made by passing through the loop. The loop is a route, not an oracle.
No compliance or legal determination is made by passing through the loop. The synthetic FSI scenario is fictional and asserts nothing about a real subject.

Why submitted is not admitted

The most consequential distinction the loop makes is between "submitted" and "admitted". submitted means routed to the candidate/admission seam, not admitted. A CandidateClaim that lands in the scenario sink and is traversable in the Workbench has been submitted; it has not been admitted, and it does not act on the admitted pool, the renderer authority list, or any public page model.

Admission remains a separate, upstream act owned by the custody owner, with renderer authorization owned by the render owner and continuation owned by the operator owner. The loop delivers material to the seam where admission can be considered; it never crosses the seam from the agent's side. There is no fallback path in this loop that admits a candidate, and there is no shortcut that lets a candidate render on GARPedia without an explicit upstream admission and renderer-authority decision.

How this relates to GARPedia / reverse wiki

The reverse wiki grammar at /method/reverse-wiki states that authority lives in the admission gate and the renderer authority list, and that the page is downstream of both. Closing the Loop is the upstream counterpart: it describes how AI-assisted material reaches the seam where admission can be considered, without short-circuiting either gate.

Public projection is not public admission mutation. GARPedia renders the projection of records that the custody and render owners have already chosen to project; it does not consume CandidateClaims, does not subscribe to the scenario sink, and does not expose Workbench operator state. A reader who walks from a Workbench scenario into GARPedia is walking from a review surface to a public read surface, not from a candidate into a rendered page.

Public-safety boundary

The loop is public-safe by construction: this page describes a synthetic scenario, names the public-surface stops, and does not expose the private operator path. The following are explicitly excluded from this page and from the public projection.

Private prompts, private source text, tool arguments, transcripts, raw model output, and private operator paths are excluded. None of these appear on this page or on any GARPedia public projection; they live in the operator and custody surfaces upstream.
No admitted fixture is introduced by this page. The admitted pool manifest is unchanged, the renderer authority list is unchanged, and the CP-SIG renderer-authority hold stands in full.
No public write, admission control, or moderation surface is added or implied. The Workbench remains the operator surface; GARPedia remains a public read surface.
No FSI determination, compliance determination, or legal determination is made by the loop or by this page. The synthetic FSI scenario is illustrative only.
GARPedia is not the source of truth. It renders the projection; the source of truth lives in the upstream admitted pool, governed by custody and renderer authority decisions described elsewhere in the method index.