{ "schemaVersion": 2, "date": "2026.08.10", "publishedAt": "2026-08-10T18:29:44-07:00", "timeZone": "America/Los_Angeles", "title": "Hiro moves toward continuous improvement with an eight-hour control cycle", "publicationStatus": "Redesign validated and staged; live activation held until the current probation boundary", "executiveSummary": [ "The user replaced the nightly-lab operating model with a continuous evidence-driven improvement loop and authorized routine Stage 5 work to proceed under a standing report-only policy instead of per-candidate approval.", "A new isolated implementation branch now makes both the Stage 6B promotion-admission interval and probation eight hours, limits admission to one promotion per rolling eight hours, and lets monitoring and rollback continue for an already-started promotion after its admission packet expires.", "Eligible Stage 4 candidates can proceed automatically through Stage 5 isolated integration, smoke tests, canaries, and repeated monitoring. Each outcome requires a post-action report; Stage 5 still cannot alter Hiro's active branch, deploy, restart services, or bypass Stage 6.", "The former automatic Daylab, Nightlab, and internet-observation probation cadence is retired. A new continuous coordinator reacts to new response evidence after a bounded minimum interval and performs a bounded idle discovery cycle when no new evidence arrives.", "The benchmark interface was realigned around the current pipeline: discovery, specification, construction, evaluation, Stage 5 integration, Stage 6 probation, candidate evidence, performance, and history. Obsolete Nightly, Internet Observations, and Lineage tabs were removed from navigation.", "Stage 6C remains disabled, but it now has a risk-specific accelerated readiness gate without an arbitrary 30-day wait: one retained 6B promotion, 30 perfect 6C shadow decisions, ten hot-reload rehearsals, three rollback drills, five interruption boundaries, full validation, and one exact policy activation.", "The redesign passed 50 focused control tests, 17 focused dashboard and policy tests, 445 repository-wide tests, and an additional Stage 6C policy test. JavaScript syntax, Python compilation, and Git whitespace checks also passed.", "The live Stage 6B promotion was not disturbed. It remains on the original checkout and has passed its zero-, 15-, and 60-minute checkpoints. The redesign is committed separately and is not yet active." ], "workstreams": [ { "title": "Unified eight-hour Stage 6B cycle", "status": "Implemented and validated in isolation", "details": [ "Changed the active policy from a 24-hour probation to an eight-hour probation with checkpoints at zero, 15, 60, 360, and 480 minutes.", "Changed the primary admission rate to one promotion per rolling eight hours, while preserving an additional three-per-24-hour ceiling and the existing seven-day boundary.", "Separated admission expiry from post-promotion monitoring authority. An already-started promotion can complete monitoring and rollback after the admission window closes, but the expired packet cannot admit another promotion.", "Reduced the active policy's maximum enablement window to eight hours and retained exact revision, policy, candidate, environment, sentinel, lease, and kill-switch checks." ] }, { "title": "Stage 5 report-only authorization", "status": "Implemented and validated", "details": [ "Stage 4 now returns eligible-for-controlled-integration rather than eligible-for-human-review when every evaluation gate passes.", "Stage 5 defaults to a standing user policy that binds the exact recommendation, candidate hash, allowed paths, tests, and monitoring plan without requiring a new user packet for every candidate.", "Successful and failed Stage 5 outcomes require a post-action report. Legacy approval packets remain verifiable for historical evidence and compatibility.", "The active branch, deployments, service restarts, external actions, high-risk changes, and unresolved recovery work remain outside the standing Stage 5 authority." ] }, { "title": "Continuous improvement coordinator", "status": "Implemented and validated", "details": [ "Added a versioned active-improvement policy and coordinator with a durable cross-process lease, evidence fingerprints, bounded minimum interval, idle discovery interval, atomic state, and append-only events.", "The coordinator composes proposal discovery, autonomous external construction, public and held-out evaluation, automatic isolated Stage 5 integration, and post-action reporting.", "The scheduler now launches the evidence-driven coordinator rather than automatic Daylab, Nightlab, or internet-observation probation.", "The machine-local legacy nightly schedule was disabled immediately so the former 22:00 job cannot revive while the code change waits for safe activation." ] }, { "title": "Benchmark control-center realignment", "status": "Implemented and validated", "details": [ "Renamed the page to Improvement Control Center and made Pipeline its primary view.", "Added a six-stage operating flow, live Stage 6 checkpoint progress, continuous-loop status, and Stage 6C readiness summary.", "Retained candidate, evaluation-health, performance, evidence-history, and comparison views.", "Removed obsolete Nightly, Internet Observations, and Lineage tabs from navigation while leaving historical evidence APIs intact." ] }, { "title": "Accelerated Stage 6C path", "status": "Designed and policy-tested; not enabled", "details": [ "Defined 6C as one schema-validated assistant-behavior fragment with atomic hot reload and a known-good fallback.", "Excluded arbitrary Python or shell changes, service restarts, and external actions from the first 6C scope.", "Replaced the 20-promotion and 30-day calendar requirement with direct 6C evidence: 30 perfect shadow decisions, ten reload rehearsals, three rollback drills, five interruption boundaries, and full tests.", "Kept one exact Stage 6C policy activation as the final gate; routine candidates after that activation would use post-action reporting rather than per-candidate approval." ] } ], "decisions": [ "Interpret eight hours as both the Stage 6B promotion-admission interval and the probation duration.", "Treat admission authority and post-promotion monitoring authority as distinct so packet expiry cannot strand an authorized probation.", "Remove routine Stage 5 pre-authorization and replace it with exact machine-verifiable eligibility plus mandatory post-action reporting.", "Keep high-risk, permission-expanding, deployment, credential, evaluator, safety-policy, and unresolved recovery changes under explicit review.", "Retire automatic Daylab and Nightlab cadence in favor of new-evidence and idle-discovery triggers.", "Measure 6C readiness with controls specific to runtime-fragment mutation rather than waiting 30 calendar days for evidence-only promotions.", "Do not merge the redesign into the active checkout while the current promotion is still under exact-revision probation." ], "validation": [ { "check": "Focused Stage 5, Stage 6, scheduler, and notification suite", "status": "passed", "result": "50 tests passed, including automatic isolated integration, legacy approval compatibility, eight-hour checkpoint retention, scheduler behavior, and post-action reporting." }, { "check": "Focused active-loop, dashboard, and policy suite", "status": "passed", "result": "17 tests passed for evidence-triggered execution, idle suppression, dashboard navigation, API behavior, and active Stage 6 policy bounds." }, { "check": "Repository-wide suite", "status": "passed", "result": "445 tests passed after linking the existing ignored pinned runtime into the isolated worktree. Two earlier failures were reproduced as missing-worktree-runtime infrastructure issues and then passed without product-code waivers." }, { "check": "Stage 6C readiness policy", "status": "passed", "result": "The additional policy test confirmed that 6C remains design-only, runtime-fragment bounded, eight-hour monitored, and separately activated." }, { "check": "Static validation", "status": "passed", "result": "Python compilation, benchmark-page JavaScript parsing, and Git diff whitespace checks passed." }, { "check": "Live probation preservation", "status": "passed", "result": "The active checkout remained on the promoted revision; zero-, 15-, and 60-minute checkpoints passed and no terminal outcome has yet been recorded." } ], "currentState": [ "The new operating model is committed on an isolated branch and is not active in Hiro's production checkout.", "The live Stage 6B evidence promotion remains in its original probation and has passed checkpoints at zero, 15, and 60 minutes.", "The machine-local legacy nightly schedule is disabled.", "Automatic Daylab, Nightlab, and internet-observation probation are disabled by the new policy and remain available only for deliberate diagnostic use after activation.", "Stage 5 per-candidate approval removal is implemented in the staged branch; high-risk exceptions remain reviewed.", "Stage 6C has an accelerated readiness plan but is not implemented or authorized to mutate runtime behavior." ], "limitations": [ "Activating the staged branch before the current exact-revision probation reaches a safe boundary would cause an integrity failure, so activation is intentionally deferred.", "The current live controller still uses the earlier 24-hour implementation until the new branch is safely adopted; the new controller cannot retroactively rewrite the existing append-only transaction.", "Continuous discovery can construct, evaluate, and integrate candidates in isolation, but active runtime behavior still cannot be promoted until the Stage 6C fragment, hot reload, shadow campaign, and rollback drills exist.", "The benchmark navigation no longer exposes retired views, but historical APIs and source code remain available for evidence preservation and manual diagnostics.", "Credentials, private prompts and responses, machine-local paths, and authority-sensitive exploit details are omitted." ], "nextSteps": [ "Allow the current promotion to reach its safe eight-hour handoff boundary without changing the active revision.", "Adopt the validated redesign commit with a clean exact-head check, then regenerate the Stage 6 policy hash and scheduler definition for the eight-hour controller.", "Start the continuous coordinator and confirm that one new-evidence cycle produces a post-action report without reviving legacy labs.", "Implement the schema-validated 6C runtime fragment and atomic hot-reload fallback.", "Complete the 30 shadow decisions, ten reload rehearsals, three rollback drills, five interruption tests, and full-suite validation required for exact 6C activation." ], "disclosureNote": "This public entry omits credentials, private prompts and responses, machine-local paths, raw held-out cases, and actionable details of unresolved control weaknesses." }