{"data":[{"id":"2dd5694d-fa6e-4538-bddd-725f850ede1f","name":"PAUL Continuous Intelligence Lab","cadence":"interval","enabled":false,"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Start with non-mutating production-bound evals. Measure fast/standard/deep routing, factual grounding, owner-intent fidelity, live-machinery grounding, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections and known PAUL failure classes. Add or refine adversarial/regression cases when current evidence exposes a repeatable weakness. If the cycle proves a deterministic regression or a clearly measurable intelligence improvement opportunity, launch exactly one focused canonical execute_objective improvement mission through current Prime authority, exact source scope, Gate, squash merge, affected-target release, and claim-matched runtime proof. Keep a change only when before/after evidence shows improvement without degrading other required invariants; otherwise reject or revert it. Never bypass Gate, never create parallel orchestration, never claim improvement from implementation alone, and never weaken tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Resource policy: do not materially interfere with normal KNOCKLAUNCH work; avoid sustained heavy compute on the owner's MSI laptop or desktop; honor live thermal, queue, runner, provider, and capacity signals; automatically reduce intensity or no-op under active owner-critical work or degraded capacity; prefer the large DigitalOcean runner only for genuinely heavy torture/adversarial evaluation. Keep inference and infrastructure cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval results, score and latency trends, failure clusters, attempted/accepted/rejected improvements, and exact production runtime lineage. This schedule remains an evaluation/improvement loop, not permission for unrelated product changes.","created_at":"2026-09-30T13:42:57.681201+00:00","last_error":null,"updated_at":"2026-09-30T15:52:36.65+00:00","last_run_at":"2026-09-30T15:43:00.695777+00:00","last_status":"cancelled","next_run_at":"2026-09-30T16:43:00.695777+00:00","failure_count":0,"last_mission_id":"cdc4a917-eab1-4527-a6d6-520e20ae4b98","interval_minutes":60,"request_contract":{"valid":true,"intent":{"rawUtterance":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Start with non-mutating production-bound evals. Measure fast/standard/deep routing, factual grounding, owner-intent fidelity, live-machinery grounding, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections and known PAUL failure classes. Add or refine adversarial/regression cases when current evidence exposes a repeatable weakness. If the cycle proves a deterministic regression or a clearly measurable intelligence improvement opportunity, launch exactly one focused canonical execute_objective improvement mission through current Prime authority, exact source scope, Gate, squash merge, affected-target release, and claim-matched runtime proof. Keep a change only when before/after evidence shows improvement without degrading other required invariants; otherwise reject or revert it. Never bypass Gate, never create parallel orchestration, never claim improvement from implementation alone, and never weaken tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Resource policy: do not materially interfere with normal KNOCKLAUNCH work; avoid sustained heavy compute on the owner's MSI laptop or desktop; honor live thermal, queue, runner, provider, and capacity signals; automatically reduce intensity or no-op under active owner-critical work or degraded capacity; prefer the large DigitalOcean runner only for genuinely heavy torture/adversarial evaluation. Keep inference and infrastructure cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval results, score and latency trends, failure clusters, attempted/accepted/rejected improvements, and exact production runtime lineage. This schedule remains an evaluation/improvement loop, not permission for unrelated product changes.","wordingChanged":false,"canonicalIntent":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Start with non-mutating production-bound evals. Measure fast/standard/deep routing, factual grounding, owner-intent fidelity, live-machinery grounding, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections and known PAUL failure classes. Add or refine adversarial/regression cases when current evidence exposes a repeatable weakness. If the cycle proves a deterministic regression or a clearly measurable intelligence improvement opportunity, launch exactly one focused canonical execute_objective improvement mission through current Prime authority, exact source scope, Gate, squash merge, affected-target release, and claim-matched runtime proof. Keep a change only when before/after evidence shows improvement without degrading other required invariants; otherwise reject or revert it. Never bypass Gate, never create parallel orchestration, never claim improvement from implementation alone, and never weaken tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Resource policy: do not materially interfere with normal KNOCKLAUNCH work; avoid sustained heavy compute on the owner's MSI laptop or desktop; honor live thermal, queue, runner, provider, and capacity signals; automatically reduce intensity or no-op under active owner-critical work or degraded capacity; prefer the large DigitalOcean runner only for genuinely heavy torture/adversarial evaluation. Keep inference and infrastructure cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval results, score and latency trends, failure clusters, attempted/accepted/rejected improvements, and exact production runtime lineage. This schedule remains an evaluation/improvement loop, not permission for unrelated product changes.","resolutionStatus":"scheduled_owner_authorized","uniqueResolution":true,"clarificationRequired":false},"source":"owner_explicit_schedule","version":1,"outcomes":[{"id":"scheduled-objective","statement":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Start with non-mutating production-bound evals. Measure fast/standard/deep routing, factual grounding, owner-intent fidelity, live-machinery grounding, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections and known PAUL failure classes. Add or refine adversarial/regression cases when current evidence exposes a repeatable weakness. If the cycle proves a deterministic regression or a clearly measurable intelligence improvement opportunity, launch exactly one focused canonical execute_objective improvement mission through current Prime authority, exact source scope, Gate, squash merge, affected-target release, and claim-matched runtime proof. Keep a change only when before/after evidence shows improvement without degrading other required invariants; otherwise reject or revert it. Never bypass Gate, never create parallel orchestration, never claim improvement from implementation alone, and never weaken tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Resource policy: do not materially interfere with normal KNOCKLAUNCH work; avoid sustained heavy compute on the owner's MSI laptop or desktop; honor live thermal, queue, runner, provider, and capacity signals; automatically reduce intensity or no-op under active owner-critical work or degraded capacity; prefer the large DigitalOcean runner only for genuinely heavy torture/adversarial evaluation. Keep inference and infrastructure cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval results, score and latency trends, failure clusters, attempted/accepted/rejected improvements, and exact production runtime lineage. This schedule remains an evaluation/improvement loop, not permission for unrelated product changes.","provenance":{"reason":"PAUL schedule 2dd5694d-fa6e-4538-bddd-725f850ede1f","source":"owner_authorized_schedule","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Start with non-mutating production-bound evals. Measure fast/standard/deep routing, factual grounding, owner-intent fidelity, live-machinery grounding, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections and known PAUL failure classes. Add or refine adversarial/regression cases when current evidence exposes a repeatable weakness. If the cycle proves a deterministic regression or a clearly measurable intelligence improvement opportunity, launch exactly one focused canonical execute_objective improvement mission through current Prime authority, exact source scope, Gate, squash merge, affected-target release, and claim-matched runtime proof. Keep a change only when before/after evidence shows improvement without degrading other required invariants; otherwise reject or revert it. Never bypass Gate, never create parallel orchestration, never claim improvement from implementation alone, and never weaken tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Resource policy: do not materially interfere with normal KNOCKLAUNCH work; avoid sustained heavy compute on the owner's MSI laptop or desktop; honor live thermal, queue, runner, provider, and capacity signals; automatically reduce intensity or no-op under active owner-critical work or degraded capacity; prefer the large DigitalOcean runner only for genuinely heavy torture/adversarial evaluation. Keep inference and infrastructure cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval results, score and latency trends, failure clusters, attempted/accepted/rejected improvements, and exact production runtime lineage. This schedule remains an evaluation/improvement loop, not permission for unrelated product changes.","constraints":[{"id":"canonical-execution","statement":"Execute through canonical KNOCKLAUNCH mission, Gate, release, and proof rules; the schedule grants no arbitrary shell, SQL, URL, secret, or provider authority.","provenance":{"reason":"Bound unattended work to the existing mission safety model.","source":"scheduler_safety","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"terminalPolicy":"all_required_proof_conditions_satisfied_after_live_release","proofConditions":[{"id":"scheduled-runtime-proof","kind":"runtime","claim":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Start with non-mutating production-bound evals. Measure fast/standard/deep routing, factual grounding, owner-intent fidelity, live-machinery grounding, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections and known PAUL failure classes. Add or refine adversarial/regression cases when current evidence exposes a repeatable weakness. If the cycle proves a deterministic regression or a clearly measurable intelligence improvement opportunity, launch exactly one focused canonical execute_objective improvement mission through current Prime authority, exact source scope, Gate, squash merge, affected-target release, and claim-matched runtime proof. Keep a change only when before/after evidence shows improvement without degrading other required invariants; otherwise reject or revert it. Never bypass Gate, never create parallel orchestration, never claim improvement from implementation alone, and never weaken tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Resource policy: do not materially interfere with normal KNOCKLAUNCH work; avoid sustained heavy compute on the owner's MSI laptop or desktop; honor live thermal, queue, runner, provider, and capacity signals; automatically reduce intensity or no-op under active owner-critical work or degraded capacity; prefer the large DigitalOcean runner only for genuinely heavy torture/adversarial evaluation. Keep inference and infrastructure cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval results, score and latency trends, failure clusters, attempted/accepted/rejected improvements, and exact production runtime lineage. This schedule remains an evaluation/improvement loop, not permission for unrelated product changes.","required":true,"targetedActions":[],"allowedReleaseTargets":[],"requiredReleaseTargets":[],"requireCapacityAdmissible":false,"minConsecutiveHealthyHeartbeats":1,"requirePositiveMemoryAttribution":false,"requireMemoryAdmissionPressureFalse":false}],"validationErrors":[]}},{"id":"a3460418-72fa-47a7-b952-c02b1fae85a5","name":"PAUL Continuous Intelligence Lab","cadence":"interval","enabled":false,"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regressions against durable owner corrections. Only when evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity may this cycle mutate the explicitly bounded PAUL source scope supplied with this schedule. Make exactly one focused improvement, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. Reject or revert any change that does not improve the measured weakness or degrades required invariants. Never broaden beyond the supplied PAUL source scope without machine-readable dependency evidence. Never bypass Gate, tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Avoid sustained heavy compute on owner hardware; use available KNOCKLAUNCH capacity and reserve large DigitalOcean compute for genuinely heavy adversarial evaluation. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as a replaceable fallback. Record durable eval results, latency/score trends, accepted/rejected improvements, and exact runtime lineage.","created_at":"2026-09-30T15:53:17.564093+00:00","last_error":null,"updated_at":"2026-09-30T15:55:20.109+00:00","last_run_at":null,"last_status":"cancelled","next_run_at":"2026-09-30T16:53:17.258+00:00","failure_count":0,"last_mission_id":null,"interval_minutes":60,"request_contract":{"valid":true,"intent":{"rawUtterance":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regressions against durable owner corrections. Only when evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity may this cycle mutate the explicitly bounded PAUL source scope supplied with this schedule. Make exactly one focused improvement, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. Reject or revert any change that does not improve the measured weakness or degrades required invariants. Never broaden beyond the supplied PAUL source scope without machine-readable dependency evidence. Never bypass Gate, tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Avoid sustained heavy compute on owner hardware; use available KNOCKLAUNCH capacity and reserve large DigitalOcean compute for genuinely heavy adversarial evaluation. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as a replaceable fallback. Record durable eval results, latency/score trends, accepted/rejected improvements, and exact runtime lineage.","wordingChanged":false,"canonicalIntent":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regressions against durable owner corrections. Only when evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity may this cycle mutate the explicitly bounded PAUL source scope supplied with this schedule. Make exactly one focused improvement, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. Reject or revert any change that does not improve the measured weakness or degrades required invariants. Never broaden beyond the supplied PAUL source scope without machine-readable dependency evidence. Never bypass Gate, tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Avoid sustained heavy compute on owner hardware; use available KNOCKLAUNCH capacity and reserve large DigitalOcean compute for genuinely heavy adversarial evaluation. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as a replaceable fallback. Record durable eval results, latency/score trends, accepted/rejected improvements, and exact runtime lineage.","resolutionStatus":"scheduled_owner_authorized","uniqueResolution":true,"clarificationRequired":false},"source":"owner_explicit_schedule","version":1,"outcomes":[{"id":"scheduled-objective","statement":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regressions against durable owner corrections. Only when evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity may this cycle mutate the explicitly bounded PAUL source scope supplied with this schedule. Make exactly one focused improvement, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. Reject or revert any change that does not improve the measured weakness or degrades required invariants. Never broaden beyond the supplied PAUL source scope without machine-readable dependency evidence. Never bypass Gate, tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Avoid sustained heavy compute on owner hardware; use available KNOCKLAUNCH capacity and reserve large DigitalOcean compute for genuinely heavy adversarial evaluation. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as a replaceable fallback. Record durable eval results, latency/score trends, accepted/rejected improvements, and exact runtime lineage.","provenance":{"reason":"PAUL schedule a3460418-72fa-47a7-b952-c02b1fae85a5","source":"owner_authorized_schedule","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regressions against durable owner corrections. Only when evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity may this cycle mutate the explicitly bounded PAUL source scope supplied with this schedule. Make exactly one focused improvement, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. Reject or revert any change that does not improve the measured weakness or degrades required invariants. Never broaden beyond the supplied PAUL source scope without machine-readable dependency evidence. Never bypass Gate, tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Avoid sustained heavy compute on owner hardware; use available KNOCKLAUNCH capacity and reserve large DigitalOcean compute for genuinely heavy adversarial evaluation. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as a replaceable fallback. Record durable eval results, latency/score trends, accepted/rejected improvements, and exact runtime lineage.","constraints":[{"id":"canonical-execution","statement":"Execute through canonical KNOCKLAUNCH mission, Gate, release, and proof rules; the schedule grants no arbitrary shell, SQL, URL, secret, or provider authority.","provenance":{"reason":"Bound unattended work to the existing mission safety model.","source":"scheduler_safety","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"terminalPolicy":"all_required_proof_conditions_satisfied_after_live_release","proofConditions":[{"id":"scheduled-runtime-proof","kind":"runtime","claim":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regressions against durable owner corrections. Only when evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity may this cycle mutate the explicitly bounded PAUL source scope supplied with this schedule. Make exactly one focused improvement, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. Reject or revert any change that does not improve the measured weakness or degrades required invariants. Never broaden beyond the supplied PAUL source scope without machine-readable dependency evidence. Never bypass Gate, tenant isolation, authorization, completion-proof discipline, or owner-intent fidelity. Avoid sustained heavy compute on owner hardware; use available KNOCKLAUNCH capacity and reserve large DigitalOcean compute for genuinely heavy adversarial evaluation. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as a replaceable fallback. Record durable eval results, latency/score trends, accepted/rejected improvements, and exact runtime lineage.","required":true,"targetedActions":[],"allowedReleaseTargets":[],"requiredReleaseTargets":[],"requireCapacityAdmissible":false,"minConsecutiveHealthyHeartbeats":1,"requirePositiveMemoryAttribution":false,"requireMemoryAdmissionPressureFalse":false}],"validationErrors":[]}},{"id":"1d8d81b9-1a48-4ecf-abee-1abfb861324a","name":"PAUL Continuous Intelligence Lab Proof","cadence":"interval","enabled":false,"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. The only authorized mutation source scope for this cycle is exactly: config/capabilities/platform-sidekick-intelligence.json; config/platform-sidekick-intelligence.json; src/__tests__/contracts/platform-sidekick-intelligence.contract.test.ts; config/capabilities/paul-scheduler.json. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections. If and only if the evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused change within that explicit source scope, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. If no change is justified, record a successful no-op evaluation and terminate cleanly. Never stop merely because source scope is unresolved: the four exact source paths above are authoritative. Never broaden beyond them without machine-readable dependency evidence. Preserve tenant isolation, authorization, owner-intent fidelity, completion-proof discipline, Cloudflare Workers AI as active non-OpenAI primary, and OpenAI only as replaceable fallback. Avoid sustained heavy compute on owner hardware.","created_at":"2026-09-30T15:55:23.352253+00:00","last_error":null,"updated_at":"2026-09-30T15:58:33.529+00:00","last_run_at":"2026-09-30T15:58:02.807835+00:00","last_status":"cancelled","next_run_at":"2026-09-30T15:59:02.807835+00:00","failure_count":0,"last_mission_id":"b85a4ab1-64ca-4541-bb0e-f125b83c2c01","interval_minutes":1,"request_contract":{"valid":true,"intent":{"rawUtterance":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. The only authorized mutation source scope for this cycle is exactly: config/capabilities/platform-sidekick-intelligence.json; config/platform-sidekick-intelligence.json; src/__tests__/contracts/platform-sidekick-intelligence.contract.test.ts; config/capabilities/paul-scheduler.json. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections. If and only if the evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused change within that explicit source scope, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. If no change is justified, record a successful no-op evaluation and terminate cleanly. Never stop merely because source scope is unresolved: the four exact source paths above are authoritative. Never broaden beyond them without machine-readable dependency evidence. Preserve tenant isolation, authorization, owner-intent fidelity, completion-proof discipline, Cloudflare Workers AI as active non-OpenAI primary, and OpenAI only as replaceable fallback. Avoid sustained heavy compute on owner hardware.","wordingChanged":false,"canonicalIntent":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. The only authorized mutation source scope for this cycle is exactly: config/capabilities/platform-sidekick-intelligence.json; config/platform-sidekick-intelligence.json; src/__tests__/contracts/platform-sidekick-intelligence.contract.test.ts; config/capabilities/paul-scheduler.json. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections. If and only if the evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused change within that explicit source scope, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. If no change is justified, record a successful no-op evaluation and terminate cleanly. Never stop merely because source scope is unresolved: the four exact source paths above are authoritative. Never broaden beyond them without machine-readable dependency evidence. Preserve tenant isolation, authorization, owner-intent fidelity, completion-proof discipline, Cloudflare Workers AI as active non-OpenAI primary, and OpenAI only as replaceable fallback. Avoid sustained heavy compute on owner hardware.","resolutionStatus":"scheduled_owner_authorized","uniqueResolution":true,"clarificationRequired":false},"source":"owner_explicit_schedule","version":1,"outcomes":[{"id":"scheduled-objective","statement":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. The only authorized mutation source scope for this cycle is exactly: config/capabilities/platform-sidekick-intelligence.json; config/platform-sidekick-intelligence.json; src/__tests__/contracts/platform-sidekick-intelligence.contract.test.ts; config/capabilities/paul-scheduler.json. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections. If and only if the evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused change within that explicit source scope, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. If no change is justified, record a successful no-op evaluation and terminate cleanly. Never stop merely because source scope is unresolved: the four exact source paths above are authoritative. Never broaden beyond them without machine-readable dependency evidence. Preserve tenant isolation, authorization, owner-intent fidelity, completion-proof discipline, Cloudflare Workers AI as active non-OpenAI primary, and OpenAI only as replaceable fallback. Avoid sustained heavy compute on owner hardware.","provenance":{"reason":"PAUL schedule 1d8d81b9-1a48-4ecf-abee-1abfb861324a","source":"owner_authorized_schedule","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. The only authorized mutation source scope for this cycle is exactly: config/capabilities/platform-sidekick-intelligence.json; config/platform-sidekick-intelligence.json; src/__tests__/contracts/platform-sidekick-intelligence.contract.test.ts; config/capabilities/paul-scheduler.json. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections. If and only if the evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused change within that explicit source scope, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. If no change is justified, record a successful no-op evaluation and terminate cleanly. Never stop merely because source scope is unresolved: the four exact source paths above are authoritative. Never broaden beyond them without machine-readable dependency evidence. Preserve tenant isolation, authorization, owner-intent fidelity, completion-proof discipline, Cloudflare Workers AI as active non-OpenAI primary, and OpenAI only as replaceable fallback. Avoid sustained heavy compute on owner hardware.","constraints":[{"id":"canonical-execution","statement":"Execute through canonical KNOCKLAUNCH mission, Gate, release, and proof rules; the schedule grants no arbitrary shell, SQL, URL, secret, or provider authority.","provenance":{"reason":"Bound unattended work to the existing mission safety model.","source":"scheduler_safety","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"terminalPolicy":"all_required_proof_conditions_satisfied_after_live_release","proofConditions":[{"id":"scheduled-runtime-proof","kind":"runtime","claim":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production intelligence stack. The only authorized mutation source scope for this cycle is exactly: config/capabilities/platform-sidekick-intelligence.json; config/platform-sidekick-intelligence.json; src/__tests__/contracts/platform-sidekick-intelligence.contract.test.ts; config/capabilities/paul-scheduler.json. Begin with non-mutating production-bound evaluations of fast/standard/deep routing, factual and live-machinery grounding, owner-intent fidelity, tenant isolation, ambiguity handling, tool-selection discipline, planner/critic behavior, completion-proof discipline, failure recovery, hallucination resistance, latency, and regression against durable owner corrections. If and only if the evidence proves a deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused change within that explicit source scope, require before/after evidence, canonical Gate, squash merge, affected-target release, and claim-matched production proof. If no change is justified, record a successful no-op evaluation and terminate cleanly. Never stop merely because source scope is unresolved: the four exact source paths above are authoritative. Never broaden beyond them without machine-readable dependency evidence. Preserve tenant isolation, authorization, owner-intent fidelity, completion-proof discipline, Cloudflare Workers AI as active non-OpenAI primary, and OpenAI only as replaceable fallback. Avoid sustained heavy compute on owner hardware.","required":true,"targetedActions":[],"allowedReleaseTargets":[],"requiredReleaseTargets":[],"requireCapacityAdmissible":false,"minConsecutiveHealthyHeartbeats":1,"requirePositiveMemoryAttribution":false,"requireMemoryAdmissionPressureFalse":false}],"validationErrors":[]}},{"id":"c8473f28-bd3a-4844-a788-c4ce3bc385e5","name":"PAUL Source Scope Canary","cadence":"interval","enabled":false,"objective":"Run one bounded PAUL source-scope canary against the current production intelligence stack. This is a canary for durable scheduler handoff: inspect/evaluate PAUL first and make no repository mutation unless a deterministic regression is proven. If mutation would be required, remain bounded to the supplied exact sourceTargets and canonical Gate/release/proof rules.","created_at":"2026-09-30T16:17:17.595503+00:00","last_error":null,"updated_at":"2026-09-30T16:21:47.105+00:00","last_run_at":"2026-09-30T16:20:01.826348+00:00","last_status":"cancelled","next_run_at":"2026-09-30T16:21:01.826348+00:00","failure_count":0,"last_mission_id":"a7b575a8-bc83-483b-81f1-30cfacafa2ec","interval_minutes":1,"request_contract":{"valid":true,"intent":{"rawUtterance":"Run one bounded PAUL source-scope canary against the current production intelligence stack. This is a canary for durable scheduler handoff: inspect/evaluate PAUL first and make no repository mutation unless a deterministic regression is proven. If mutation would be required, remain bounded to the supplied exact sourceTargets and canonical Gate/release/proof rules.","wordingChanged":false,"canonicalIntent":"Run one bounded PAUL source-scope canary against the current production intelligence stack. This is a canary for durable scheduler handoff: inspect/evaluate PAUL first and make no repository mutation unless a deterministic regression is proven. If mutation would be required, remain bounded to the supplied exact sourceTargets and canonical Gate/release/proof rules.","resolutionStatus":"scheduled_owner_authorized","uniqueResolution":true,"clarificationRequired":false},"source":"owner_explicit_schedule","version":1,"outcomes":[{"id":"scheduled-objective","statement":"Run one bounded PAUL source-scope canary against the current production intelligence stack. This is a canary for durable scheduler handoff: inspect/evaluate PAUL first and make no repository mutation unless a deterministic regression is proven. If mutation would be required, remain bounded to the supplied exact sourceTargets and canonical Gate/release/proof rules.","provenance":{"reason":"PAUL schedule c8473f28-bd3a-4844-a788-c4ce3bc385e5","source":"owner_authorized_schedule","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"objective":"Run one bounded PAUL source-scope canary against the current production intelligence stack. This is a canary for durable scheduler handoff: inspect/evaluate PAUL first and make no repository mutation unless a deterministic regression is proven. If mutation would be required, remain bounded to the supplied exact sourceTargets and canonical Gate/release/proof rules.","constraints":[{"id":"canonical-execution","statement":"Execute through canonical KNOCKLAUNCH mission, Gate, release, and proof rules; the schedule grants no arbitrary shell, SQL, URL, secret, or provider authority.","provenance":{"reason":"Bound unattended work to the existing mission safety model.","source":"scheduler_safety","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"sourceTargets":["workers/cc-execution/src/index-mcp.js","config/capabilities/platform-sidekick-intelligence.json"],"terminalPolicy":"all_required_proof_conditions_satisfied_after_live_release","proofConditions":[{"id":"scheduled-runtime-proof","kind":"runtime","claim":"Run one bounded PAUL source-scope canary against the current production intelligence stack. This is a canary for durable scheduler handoff: inspect/evaluate PAUL first and make no repository mutation unless a deterministic regression is proven. If mutation would be required, remain bounded to the supplied exact sourceTargets and canonical Gate/release/proof rules.","required":true,"targetedActions":[],"allowedReleaseTargets":[],"requiredReleaseTargets":[],"requireCapacityAdmissible":false,"minConsecutiveHealthyHeartbeats":1,"requirePositiveMemoryAttribution":false,"requireMemoryAdmissionPressureFalse":false}],"validationErrors":[]}},{"id":"7335281c-bd3d-4aef-8016-6ab29cb466e1","name":"PAUL Continuous Intelligence Lab","cadence":"interval","enabled":true,"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production stack. Begin read-only. Run representative production PAUL evaluations across fast, standard, and deep reasoning, including owner-realistic correction recovery, direct awareness questions, current/live-state reads, greetings while durable work is running, profanity/urgency semantics, tool-argument grounding, tenant isolation, memory/context selection, ambiguity handling, hallucination resistance, completion-proof discipline, and latency. Inspect the actual returned answer/tool-call text for false greens; never accept a numeric pass when the owner-visible response still violates the intended behavior. Compare results to current main/runtime truth and durable owner corrections. If and only if evidence proves one deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused repair within the authorized PAUL source scope, add or strengthen a regression case for the proven failure, pass canonical Gate, squash merge, release only affected targets, and verify the live production behavior. If no change is justified, record a successful no-op evaluation and terminate cleanly. Do not broaden source scope without machine-readable dependency evidence. Never bypass tenant isolation, authorization, owner-intent fidelity, Gate, release ownership, or terminal proof. Do not materially interfere with active owner-critical work: read current queue/capacity/thermal/runtime state first, reduce intensity or no-op when capacity is constrained, avoid sustained heavy compute on owner MSI hardware, and use available KNOCKLAUNCH cloud/DigitalOcean heavy capacity only when genuinely needed. Keep cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval outcomes, false-green findings, accepted/rejected improvements, latency/score trends, and exact runtime lineage.","created_at":"2026-09-30T21:25:07.101674+00:00","last_error":null,"updated_at":"2026-10-01T05:27:03.377+00:00","last_run_at":"2026-10-01T05:27:00.725154+00:00","last_status":"synthesis_required","next_run_at":"2026-10-01T09:27:00.725154+00:00","failure_count":0,"last_mission_id":"189f62c4-36de-4aad-aca8-726179ae110f","interval_minutes":240,"request_contract":{"valid":true,"intent":{"rawUtterance":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production stack. Begin read-only. Run representative production PAUL evaluations across fast, standard, and deep reasoning, including owner-realistic correction recovery, direct awareness questions, current/live-state reads, greetings while durable work is running, profanity/urgency semantics, tool-argument grounding, tenant isolation, memory/context selection, ambiguity handling, hallucination resistance, completion-proof discipline, and latency. Inspect the actual returned answer/tool-call text for false greens; never accept a numeric pass when the owner-visible response still violates the intended behavior. Compare results to current main/runtime truth and durable owner corrections. If and only if evidence proves one deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused repair within the authorized PAUL source scope, add or strengthen a regression case for the proven failure, pass canonical Gate, squash merge, release only affected targets, and verify the live production behavior. If no change is justified, record a successful no-op evaluation and terminate cleanly. Do not broaden source scope without machine-readable dependency evidence. Never bypass tenant isolation, authorization, owner-intent fidelity, Gate, release ownership, or terminal proof. Do not materially interfere with active owner-critical work: read current queue/capacity/thermal/runtime state first, reduce intensity or no-op when capacity is constrained, avoid sustained heavy compute on owner MSI hardware, and use available KNOCKLAUNCH cloud/DigitalOcean heavy capacity only when genuinely needed. Keep cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval outcomes, false-green findings, accepted/rejected improvements, latency/score trends, and exact runtime lineage.","wordingChanged":false,"canonicalIntent":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production stack. Begin read-only. Run representative production PAUL evaluations across fast, standard, and deep reasoning, including owner-realistic correction recovery, direct awareness questions, current/live-state reads, greetings while durable work is running, profanity/urgency semantics, tool-argument grounding, tenant isolation, memory/context selection, ambiguity handling, hallucination resistance, completion-proof discipline, and latency. Inspect the actual returned answer/tool-call text for false greens; never accept a numeric pass when the owner-visible response still violates the intended behavior. Compare results to current main/runtime truth and durable owner corrections. If and only if evidence proves one deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused repair within the authorized PAUL source scope, add or strengthen a regression case for the proven failure, pass canonical Gate, squash merge, release only affected targets, and verify the live production behavior. If no change is justified, record a successful no-op evaluation and terminate cleanly. Do not broaden source scope without machine-readable dependency evidence. Never bypass tenant isolation, authorization, owner-intent fidelity, Gate, release ownership, or terminal proof. Do not materially interfere with active owner-critical work: read current queue/capacity/thermal/runtime state first, reduce intensity or no-op when capacity is constrained, avoid sustained heavy compute on owner MSI hardware, and use available KNOCKLAUNCH cloud/DigitalOcean heavy capacity only when genuinely needed. Keep cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval outcomes, false-green findings, accepted/rejected improvements, latency/score trends, and exact runtime lineage.","resolutionStatus":"scheduled_owner_authorized","uniqueResolution":true,"clarificationRequired":false},"source":"owner_explicit_schedule","version":1,"outcomes":[{"id":"scheduled-objective","statement":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production stack. Begin read-only. Run representative production PAUL evaluations across fast, standard, and deep reasoning, including owner-realistic correction recovery, direct awareness questions, current/live-state reads, greetings while durable work is running, profanity/urgency semantics, tool-argument grounding, tenant isolation, memory/context selection, ambiguity handling, hallucination resistance, completion-proof discipline, and latency. Inspect the actual returned answer/tool-call text for false greens; never accept a numeric pass when the owner-visible response still violates the intended behavior. Compare results to current main/runtime truth and durable owner corrections. If and only if evidence proves one deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused repair within the authorized PAUL source scope, add or strengthen a regression case for the proven failure, pass canonical Gate, squash merge, release only affected targets, and verify the live production behavior. If no change is justified, record a successful no-op evaluation and terminate cleanly. Do not broaden source scope without machine-readable dependency evidence. Never bypass tenant isolation, authorization, owner-intent fidelity, Gate, release ownership, or terminal proof. Do not materially interfere with active owner-critical work: read current queue/capacity/thermal/runtime state first, reduce intensity or no-op when capacity is constrained, avoid sustained heavy compute on owner MSI hardware, and use available KNOCKLAUNCH cloud/DigitalOcean heavy capacity only when genuinely needed. Keep cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval outcomes, false-green findings, accepted/rejected improvements, latency/score trends, and exact runtime lineage.","provenance":{"reason":"PAUL schedule 7335281c-bd3d-4aef-8016-6ab29cb466e1","source":"owner_authorized_schedule","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"objective":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production stack. Begin read-only. Run representative production PAUL evaluations across fast, standard, and deep reasoning, including owner-realistic correction recovery, direct awareness questions, current/live-state reads, greetings while durable work is running, profanity/urgency semantics, tool-argument grounding, tenant isolation, memory/context selection, ambiguity handling, hallucination resistance, completion-proof discipline, and latency. Inspect the actual returned answer/tool-call text for false greens; never accept a numeric pass when the owner-visible response still violates the intended behavior. Compare results to current main/runtime truth and durable owner corrections. If and only if evidence proves one deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused repair within the authorized PAUL source scope, add or strengthen a regression case for the proven failure, pass canonical Gate, squash merge, release only affected targets, and verify the live production behavior. If no change is justified, record a successful no-op evaluation and terminate cleanly. Do not broaden source scope without machine-readable dependency evidence. Never bypass tenant isolation, authorization, owner-intent fidelity, Gate, release ownership, or terminal proof. Do not materially interfere with active owner-critical work: read current queue/capacity/thermal/runtime state first, reduce intensity or no-op when capacity is constrained, avoid sustained heavy compute on owner MSI hardware, and use available KNOCKLAUNCH cloud/DigitalOcean heavy capacity only when genuinely needed. Keep cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval outcomes, false-green findings, accepted/rejected improvements, latency/score trends, and exact runtime lineage.","constraints":[{"id":"canonical-execution","statement":"Execute through canonical KNOCKLAUNCH mission, Gate, release, and proof rules; the schedule grants no arbitrary shell, SQL, URL, secret, or provider authority.","provenance":{"reason":"Bound unattended work to the existing mission safety model.","source":"scheduler_safety","inherited":false,"blockingAuthority":true},"proofConditionIds":["scheduled-runtime-proof"]}],"sourceTargets":["workers/cc-execution/src/platform-sidekick-intelligence.js","workers/cc-execution/src/model-provider-gateway.js","workers/cc-execution/src/index-mcp.js","workers/cc-execution/src/version.js","supabase/functions/operator-respond/index.ts","supabase/functions/operator-respond/paul-memory.ts","supabase/functions/operator-respond/paul-owner-profile.ts","src/pages/admin/GodModeCommandConsole.tsx","config/platform-sidekick-intelligence.json","config/paul-owner-profile.json","src/__tests__/contracts/platform-sidekick-intelligence.contract.test.ts","src/__tests__/contracts/mobile-admin-console-authority.contract.test.ts"],"terminalPolicy":"all_required_proof_conditions_satisfied_after_live_release","proofConditions":[{"id":"scheduled-runtime-proof","kind":"runtime","claim":"Run one bounded PAUL Continuous Intelligence Lab cycle against the current canonical production stack. Begin read-only. Run representative production PAUL evaluations across fast, standard, and deep reasoning, including owner-realistic correction recovery, direct awareness questions, current/live-state reads, greetings while durable work is running, profanity/urgency semantics, tool-argument grounding, tenant isolation, memory/context selection, ambiguity handling, hallucination resistance, completion-proof discipline, and latency. Inspect the actual returned answer/tool-call text for false greens; never accept a numeric pass when the owner-visible response still violates the intended behavior. Compare results to current main/runtime truth and durable owner corrections. If and only if evidence proves one deterministic regression or clearly measurable intelligence improvement opportunity, make exactly one focused repair within the authorized PAUL source scope, add or strengthen a regression case for the proven failure, pass canonical Gate, squash merge, release only affected targets, and verify the live production behavior. If no change is justified, record a successful no-op evaluation and terminate cleanly. Do not broaden source scope without machine-readable dependency evidence. Never bypass tenant isolation, authorization, owner-intent fidelity, Gate, release ownership, or terminal proof. Do not materially interfere with active owner-critical work: read current queue/capacity/thermal/runtime state first, reduce intensity or no-op when capacity is constrained, avoid sustained heavy compute on owner MSI hardware, and use available KNOCKLAUNCH cloud/DigitalOcean heavy capacity only when genuinely needed. Keep cost bounded. Preserve Cloudflare Workers AI as the active non-OpenAI primary path and OpenAI only as replaceable fallback. Record durable eval outcomes, false-green findings, accepted/rejected improvements, latency/score trends, and exact runtime lineage.","required":true,"targetedActions":[],"allowedReleaseTargets":[],"requiredReleaseTargets":[],"requireCapacityAdmissible":false,"minConsecutiveHealthyHeartbeats":1,"requirePositiveMemoryAttribution":false,"requireMemoryAdmissionPressureFalse":false}],"validationErrors":[]}}],"rows":5,"limit":200,"table":"paul_schedules","format":"knocklaunch-table-chunk-v1","offset":0,"schema":"public"}