GOVERNED CHANGE / HUMAN CONTROL BEFORE EXECUTION

Review proposed code changes before anything runs.

ARKON turns a requested change into an evidence-backed plan, checks its scope and risk, waits for your decision, then tests an approved patch in an isolated synthetic repository.

Automated changes can drift beyond the request or appear successful while silently breaking behavior. ARKON requires scope evidence and real test results before it calls a change verified.

You remain in control: policy cannot approve a change, and no fixture change executes until a reviewer explicitly approves its exact scope.

Run the safe synthetic demo Historical GPT-5.6 evidence

Every interactive scenario is a committed synthetic fixture with deterministic output. The walkthrough explains controls but never analyzes, decides, executes, replays, exports, or calls a model.

Inspect release identity

Deployed application commit 2c893d9ca8df0447d6fa09ec4bb0a876dc0bfeaf

HOW ARKON KEEPS CHANGE CONTROL EXPLICITOne request. Four understandable phases.

1
Understand the changeSee the problem, desired outcome, and affected components.
2
Check plan, policy, and riskReview proposed edits and the tests that must prove them.
3
Human decidesApprove or reject the exact evidence-bound scope.
4
Execute, verify, and preserve proofRun only in an isolated copy, then seal a replayable result.

Verified means the approved synthetic patch passed the fixed test profile and its stored evidence bindings replay successfully. It does not mean merged or deployed.

Bound the request

Intent & fixture scenario

Review the prefilled intent before analysis. It can target only the versioned synthetic repository bundled with this demo.

Start the walkthrough to review intent, generate a grounded plan, inspect policy, record a reviewer decision, and verify the proof chain.

HISTORICAL MODEL EVIDENCE / READ-ONLY RECEIPT

Measured GPT-5.6 evidence

This retained receipt is separate from every deterministic scenario above. Opening it performs no model call.

Historical receipt available

Inspect historical evidence

Viewing this sanitized, read-only evidence performs no new model call. The hosted workflow remains secret-free, deterministic, and fixture-only.

EVALUATOR-SEPARATED REASONING HOLDOUTretry-after-reasoning-holdout

Complete-request leak gate
Deterministic IR policy
2 hidden post-change cases

Evaluated commit: 0bc7400f7490764a5f9bfa00d93dc1522dd53ac2
Current release commit: 2c893d9ca8df0447d6fa09ec4bb0a876dc0bfeaf
Requested model: gpt-5.6
Returned model: gpt-5.6-sol
Response ID: resp_0a222ed277eee6b8016a57696eb0248198a400ae6952e2c256
Evaluated at: Jul 15, 2026, 11:05:37 AM
Token usage: 2,630 in · 1,134 out · 3,764 total
Release relationship: Historical receipt · protected lineage verified
Planning contract: change-plan.v2 · change-factory-holdout-planner.v4

SANITIZED STRUCTURED PLAN

Proposal to correct only the permitted billing parser so numeric Retry-After seconds are converted to milliseconds, while preserving the exported API and all non-numeric behavior, and to add bounded numeric regression coverage.

modify
packages/billing/src/retry-after.mjs
Multiply the existing numeric Retry-After result by 1000 so HTTP seconds are returned as milliseconds. Leave the function signature and HTTP-date, expired-date, blank, and invalid-value paths unchanged.
create
test/retry-after-regression.test.mjs
Add bounded regression cases for the billing parser covering representative numeric Retry-After values and their corrected millisecond results.

Regression contract

test/retry-after-regression.test.mjs · Billing parseRetryAfter converts numeric HTTP Retry-After seconds to milliseconds for 1, 5, and 120 seconds.

Inspect actual sanitized typed IR

{
  "schemaVersion": "change-plan.v2",
  "intentDigest": "c399b9f8d862dcffc434c28f72df12b583796c8b6b20d8e41b7a50a4bc40ee1c",
  "repositoryDigest": "0130de9c6f6fbe102ad11f0f2d62644ebfac42f1cbcfcfce35e60b90c79898cf",
  "evidenceDigest": "910021139e3af95ef2465ba73cd4d656032aca43815984eef93dbf80362302d2",
  "intentConformance": {
    "schemaVersion": "intent-conformance.v1",
    "policyId": "read-only-versioned-scenario.v1",
    "scenarioId": "retry-after-reasoning-holdout",
    "scenarioVersion": "2026.07.15.1",
    "requestMode": "read_only_versioned_scenario",
    "conforms": true,
    "classification": "exact_scenario_intent",
    "reasons": [
      "Intent exactly matches retry-after-reasoning-holdout@2026.07.15.1, whose semantics and transformation variant are server-owned."
    ],
    "digest": "abee95cb87af46aec45df1b8b9f92082e0aeb16e65ef6480e00cdd96b7134663"
  },
  "summary": "Proposal to correct only the permitted billing parser so numeric Retry-After seconds are converted to milliseconds, while preserving the exported API and all non-numeric behavior, and to add bounded numeric regression coverage.",
  "changes": [
    {
      "id": "C1",
      "path": "packages/billing/src/retry-after.mjs",
      "symbol": "parseRetryAfter",
      "description": "Multiply the existing numeric Retry-After result by 1000 so HTTP seconds are returned as milliseconds. Leave the function signature and HTTP-date, expired-date, blank, and invalid-value paths unchanged.",
      "rationale": "The billing parser currently returns the parsed numeric value without unit conversion, while repository policy requires numeric HTTP seconds to be returned as milliseconds.",
      "evidenceIds": [
        "E-SRC-002",
        "E-CON-001",
        "E-CON-002",
        "E-CON-003"
      ],
      "operation": "modify",
      "expectedPreimageSha256": "9eb54c75c4d15de06ab755e63cfe1c17e33c10928a7051b7097839093392d980",
      "transformation": {
        "kind": "numeric_multiply_return",
        "targetEvidenceId": "E-SRC-002",
        "expectedPreimageSha256": "9eb54c75c4d15de06ab755e63cfe1c17e33c10928a7051b7097839093392d980",
        "permittedSymbol": "parseRetryAfter",
        "numericExpression": "Number(normalized)",
        "inputUnit": "seconds",
        "outputUnit": "milliseconds",
        "factor": 1000
      }
    },
    {
      "id": "C2",
      "path": "test/retry-after-regression.test.mjs",
      "symbol": "parseRetryAfter",
      "description": "Add bounded regression cases for the billing parser covering representative numeric Retry-After values and their corrected millisecond results.",
      "rationale": "These cases fail under the unconverted numeric return and pass after the seconds-to-milliseconds correction, without requiring changes to existing HTTP-date coverage.",
      "evidenceIds": [
        "E-SRC-002",
        "E-CON-002",
        "E-CON-003",
        "E-CON-004"
      ],
      "operation": "create",
      "expectedPreimageSha256": null,
      "transformation": {
        "kind": "bounded_numeric_regression_cases",
        "permittedSymbol": "parseRetryAfter",
        "sourceChangeIds": [
          "C1"
        ],
        "cases": [
          {
            "input": "1",
            "expectedMilliseconds": 1000
          },
          {
            "input": "5",
            "expectedMilliseconds": 5000
          },
          {
            "input": "120",
            "expectedMilliseconds": 120000
          }
        ],
        "expectedBaselineOutcome": "fail",
        "expectedPatchedOutcome": "pass"
      }
    }
  ],
  "regressionTests": [
    {
      "id": "T1",
      "behavior": "Billing parseRetryAfter converts numeric HTTP Retry-After seconds to milliseconds for 1, 5, and 120 seconds.",
      "targetPath": "test/retry-after-regression.test.mjs",
      "expectedBaselineOutcome": "fail",
      "expectedPatchedOutcome": "pass",
      "evidenceIds": [
        "E-SRC-002",
        "E-CON-002",
        "E-CON-004"
      ]
    }
  ],
  "constraintsAcknowledged": [
    "Preserve the exported parseRetryAfter(value, nowMs) public API.",
    "Convert numeric Retry-After values from HTTP seconds to milliseconds.",
    "Do not alter HTTP-date, expired-date, blank, or invalid-value behavior.",
    "Modify only packages/billing/src/retry-after.mjs and create only test/retry-after-regression.test.mjs, as constrained by the permitted boundary.",
    "Do not modify manifests, dependencies, CI, infrastructure, credentials, or deployment files.",
    "Verification remains limited to the fixture-owned Node test profile; no network, Git, package-manager, shell, merge, or deployment access is proposed."
  ],
  "verificationProfileId": "fixture-node-test-v1",
  "risk": {
    "level": "low",
    "reasons": [
      "The production change is confined to the evidenced numeric return expression in one permitted module.",
      "The conversion factor is directly established by the repository policy.",
      "The public API and non-numeric control paths remain unchanged.",
      "Regression inputs and expected values are small, deterministic, and bounded."
    ]
  },
  "confidence": 0.99,
  "provenance": {
    "requestedModel": "gpt-5.6",
    "schemaVersion": "change-plan.v2",
    "disclosure": "One protected GPT-5.6 Responses API call proposes typed IR over synthetic evidence with store:false and no tools. Deterministic policy stops at REQUIRE_HUMAN_APPROVAL; a separate non-human, fixture-only evaluation authorization permits one isolated verification.",
    "adapter": "openai.responses",
    "actualModel": "gpt-5.6-sol",
    "liveModelCall": true,
    "promptVersion": "change-factory-holdout-planner.v4",
    "responseId": "resp_0a222ed277eee6b8016a57696eb0248198a400ae6952e2c256"
  }
}

POLICY & AUTHORIZATION

REQUIRE HUMAN APPROVAL

The product policy stopped at its approval-required boundary. No human approval was supplied; a separate non-human authorization permitted one fixture-only evaluation.

Subsequent authority: Non-human evaluation authorization
Human approval: No
Actor: Protected GPT-5.6 evaluation workflow
Scope: Fixture-only verification
Boundary checks: Causal · one-shot

RECORDED RED / GREEN

Baseline4 pass · 0 fail · 0 skip

RED0 pass · 1 fail · 0 skip

GREEN5 pass · 0 fail · 0 skip

Hidden1 pass · 0 fail · 0 skip

Inspect retained cryptographic digests

Request: 6026ddda1ed7a519ca27dc13907b39be839ac11960bca30efb7e089d481c215c
Schema: 7632a572b1ed126c1a00668258ef2d2ef07432af311d256b42162c79120c748e
Output: d3a0423c5b5c1bbbce682d1cedb806eca6872ac035a49d1ddd0da6d95d1401b5
Plan: 33cbf654f77b71cce2eef907522ff444c8796543c05440d84661d91fd9256cb3
Safety: efd6b6d80670a5dc66302011e92ebc9411dff8f0e4ff10a626349c17d17ce7ef
Authorization: b76aa57d4470b95d5af1f9133acba7d251a96c193bf623ec717e60c1d43161c9
Holdout receipt: f23bf3e860e39cd557c30b1d2061a05bcf0d30a8a17ed39ee03d8655bb39a8b9
Evaluator oracle: d40a626184b65d29dc145cf4aa10fcca08ca538f97c9157c91c5963c0b03bbb2
Holdout request binding: 6026ddda1ed7a519ca27dc13907b39be839ac11960bca30efb7e089d481c215c
Holdout plan binding: 33cbf654f77b71cce2eef907522ff444c8796543c05440d84661d91fd9256cb3
Hidden output: c5c8131a3857f5c5294c918493e9eb1a8ba6e66780c0f6d0b2ae8eeeefc9d950
Patch: b32b8db5c4f67f3fbf2eb71650b006196861d052b28fcfbc57709e9893bfacbe
Final tree: bbb07d629f5678d68ad6c775c34055f215d3e62ce79b4651134a0ea430c08a79

Submission provenance

Prior work, Build Week work & decisions

The experience distinguishes the existing ARKON platform from what this submission adds. It shows recorded schema actor types, explicitly annotates known automation, and does not claim authenticated identities.

PRE-EXISTING ARKON · NOT NEW BUILD WEEK WORK

Governance platform foundation

ARKON previously provided orchestration, companions, AEGIS, risk-tiered approvals, audit, Azure DevOps integration, MCP, a web console, and Azure infrastructure.

Platform orchestration & companions
AEGIS concepts and risk-tiered gates
Enterprise integrations and cloud infrastructure

NEW · OPENAI BUILD WEEK 2026

Change Factory proof chain

This repository adds repository-grounded planning, deterministic fixture policy, a fixture-only judge workspace, isolated patch-and-test execution, and the replayable evidence experience.

GPT-5.6 structured repository-reasoning contract
Intent → evidence → approved isolated change loop
RED/GREEN execution evidence, diffs, and hash-linked replay

CODEX CONTRIBUTION

Implementation collaborator

Codex contributed submission code, fixture-safe workflow implementation, judge UX, and verification support inside this dated repository under the human-authored Build Week contract.

Codex does not supply reviewer approval and cannot widen the fixture boundary.

DECISION PROVENANCE

Decision pending

A reviewer must inspect the evidence, policy result, and exact envelope before any isolated patch execution. Any public-UI name is a user-entered assertion, not authenticated human identity.

No decision is inferred from model output or policy success.