{ "schemaVersion": "ke.research-findings-evidence.v1", "release": { "id": "ke-research-findings-2026-07-26", "title": "Resident context worked; native savings did not", "date": "2026-07-26", "generatedAt": "2026-07-26T20:32:02-04:00", "informationCutoff": "2026-07-26T20:19:06-04:00", "canonicalUrl": "https://kestudios.dev/research/updates/2026-07-26", "author": { "name": "William Keenan", "organization": "K&E Studios Research", "url": "https://kestudios.dev/william-keenan", "email": "william@kestudios.dev" }, "reviewStatus": "source-verified local research update; not peer reviewed or independently validated" }, "sourceSnapshots": [ { "label": "Waggle and Kea committed snapshot", "commit": "2e169f6786977d8c412dab33c77f28136e287b8d", "capturedAt": "2026-07-26T20:19:06-04:00", "inspectionBasis": "bytes addressable from the frozen commit only", "artifactSha256": "388985cd2c9b0b6179c66f711280769a5d71ea4a09c10a906297bfd826bfd8e0" }, { "label": "HEATWAKE committed snapshot", "commit": "279f91aa7ea8c9c97d99c96f936d384f58bbcbfd", "capturedAt": "2026-07-26T20:19:06-04:00", "inspectionBasis": "bytes addressable from the frozen commit only", "artifactSha256": [ "5d916feb7793b19a00851d3dd5853d697a59400706af299ead1fa7a3cd8c5fed", "12803a8c692ca3d719a58ec16517c5e3781683b7b9934178410b7f05b431aee6", "9bf696213cb26119a2c874ad108d441d91ffa6beb13b69e80f939af752eab4b5" ] } ], "methodology": { "comparisonPolicy": "Inspect the prior public checkpoint first, then compare only frozen committed source bytes available at the information cutoff.", "evidenceLevels": { "recorded": "present in a frozen attributable source artifact", "observed": "read or measured during the accepted bounded run", "verified": "reproduced by a zero-call replay or focused local check", "independentValidation": false }, "environment": "local macOS research workstation; one frozen local open-weight model and runtime for the matched Waggle and Kea arms", "uncertainty": [ "Waggle and Kea exactness uses three deterministic task fixtures; no statistical confidence interval is claimed.", "Waggle and Kea wall times are descriptive measurements from one accepted trial on one host.", "HEATWAKE launch observations cover only the start of a raw collection and do not estimate seven-day reliability or market-data completeness." ], "exclusions": [ "All uncommitted or later source-writer bytes after the information cutoff", "Two interrupted-trial model-process loads from the accepted Waggle and Kea result denominator", "One HEATWAKE raw fragment without a terminal segment receipt", "All HEATWAKE raw payloads from model features, labels, calibration, candidate selection, signals, and the sealed final test" ], "publicSafety": "Public output was composed from accepted aggregate claims and hashes. Non-public implementation materials and operationally sensitive detail are omitted." }, "programs": [ { "id": "waggle-kea", "name": "Waggle + Kea", "verdict": "A bounded resident context service and Kea audit path worked across three frozen tasks. The resident-native timing comparison failed, and no savings or unique native capability is supported.", "claimDeltas": [ { "claim": "A resident-native service can support multiple source-separated local tasks through Kea.", "priorStatus": "not established", "newEvidence": "four matched informed arms at 3/3 exact; Kea qualified 3/3 projections; restricted consumer continued from projections only", "currentStatus": "supported in the bounded fixtures" }, { "claim": "Resident-native execution improves wall-time efficiency.", "priorStatus": "unresolved", "newEvidence": "24.412181 seconds resident native versus 10.42122625 seconds fresh native and 20.295337541 seconds resident full text", "currentStatus": "rejected for this experiment" }, { "claim": "The service implies token, Credit, storage, cost, energy, or overall-efficiency savings.", "priorStatus": "unsupported", "newEvidence": "279321005 context bytes retained; direct energy unavailable; no provider billing or purchased-Credit comparison", "currentStatus": "unsupported" } ], "experiment": { "question": "Can one bounded resident-native service preserve a qualified local context path across multiple source-separated tasks, and does it beat matched controls?", "falsifiers": [ "Any informed arm fails exact expected values.", "Kea cannot independently qualify every accepted output.", "The restricted consumer requires direct access to the producing runtime instead of the Kea projection.", "The resident-native path fails to beat the preregistered matched controls on the claimed performance measure." ], "design": { "tasks": 3, "taskFamiliesDisjointFromInterruptedTrial": true, "arms": [ "resident native", "resident full text", "fresh native", "fresh full text", "no state" ], "informedArms": 4, "sameFrozenTasksAndExpectedValues": true, "sameLocalModelFamily": true, "restrictedConsumerReadsKeaProjectionOnly": true, "maximumModelProcessLoads": 10, "maximumForwardDecodeSteps": 1024 }, "results": { "qualifiedTasks": "3/3", "residentNativeExact": "3/3", "residentFullTextExact": "3/3", "freshNativeExact": "3/3", "freshFullTextExact": "3/3", "noStateQualifiedProjection": false, "restrictedConsumerDisposition": "continue", "wallTimeMilliseconds": { "residentNative": 24412.181, "freshNative": 10421.22625, "residentFullText": 20295.337541, "freshFullText": 20912.578416 }, "acceptedResidentModelProcessLoads": 2, "acceptedFreshModelProcessLoads": 6, "acceptedTotalModelProcessLoads": 8, "preservedInterruptedTrialModelProcessLoads": 2, "acceptedForwardDecodeSteps": 298, "nativeRetainedStateBytes": 279321005, "nativeQueryBytes": 1626, "fullTextPromptBytes": 27180, "maximumPeakRssBytes": 2009907200, "directEnergyMetricAvailable": false, "reportSha256": "054aec0efeeda57a938681c346185129581ab4d343de071c09658805081f74ee" }, "controls": { "formalReplayPins": 20, "formalReplayTasks": 3, "formalReplayAcceptedLoads": 8, "formalReplayPreservedInterruptedLoads": 2, "formalReplayAdditionalLoads": 0, "formalReplayAdditionalForwardDecodeSteps": 0, "priorCheckpointReplayPassed": true, "priorReplayAdditionalLoads": 0, "priorReplayAdditionalForwardDecodeSteps": 0, "serviceSuiteTasksPassed": "3/3", "adversarialInputsRejectedOrFailedClosed": "11/11", "typecheckPassed": true, "buildPassed": true }, "supportedClaim": "One bounded local resident context service, Kea audit path, and projection-only consumer handoff worked across the three frozen tasks.", "negativeClaim": "The resident-native path did not improve wall time against fresh native or resident full text in this experiment.", "notSupported": [ "A learned language, neuralese, hidden-state semantic transfer, or unique native capability", "A semantic, quality, general-reasoning, or cross-model advantage", "Token, Credit, provider-cost, storage, network, energy, wall-time, or overall-efficiency savings", "Production reliability, deployment, customer value, or authority execution" ] } }, { "id": "heatwake", "name": "HEATWAKE", "verdict": "The interrupted fixed window was retained but rejected for admission. A fresh raw-only successor launched at 0/7 complete UTC days, with every model, signal, trading, and production claim still closed.", "claimDeltas": [ { "claim": "The prior fixed seven-day raw-coverage window can support source admission.", "priorStatus": "incomplete; no admission result", "newEvidence": "host reboot, ten closed receipts, one excluded unreceipted fragment, no terminal session receipt, and insufficient remaining calendar", "currentStatus": "rejected; retained but ineligible and ABSTAIN" }, { "claim": "A fresh successor collection changes the model or trading evidence level.", "priorStatus": "no successor-window evidence", "newEvidence": "fresh frozen raw-only contract, one source-bound collector, seeded snapshot, zero launch gaps or reconnects, 0/7 complete UTC days", "currentStatus": "no upgrade; raw custody only" } ], "interruptedWindow": { "question": "Can the prior fixed window still satisfy its original seven-complete-day and terminal-receipt gates after the lifecycle interruption?", "result": { "lifecycleInterrupted": true, "automaticProcessRestart": false, "closedReceipts": 10, "unreceiptedFragmentsExcluded": 1, "terminalSessionReceiptPresent": false, "requiredFullUtcDays": 7, "maximumRemainingFullUtcDaysAtReview": 2, "sameSessionCompletionPossible": false, "rowsCopiedToSuccessor": 0, "rawSourceAdmitted": false, "modelInputEligible": false, "policyAction": "ABSTAIN", "artifactSha256": "5d916feb7793b19a00851d3dd5853d697a59400706af299ead1fa7a3cd8c5fed" }, "negativeClaim": "The frozen window could not satisfy its original coverage gate and did not produce a source-admission or science result." }, "successorWindow": { "question": "Can a fresh raw-only collection start without carrying prior rows, weakening thresholds, or crossing downstream authority gates?", "design": { "newSession": true, "newWindow": true, "newCustodyRoot": true, "priorRowsCopied": 0, "requiredFullUtcDays": 7, "minimumDayCoverage": 0.9, "automaticAdmission": false, "automaticProcessRestart": false, "oneWriter": true, "workers": 1, "rawOnlyUntilSeparateAdmission": true, "modelFeaturesEnabled": false, "labelsEnabled": false, "calibrationEnabled": false, "candidateSelectionEnabled": false, "executionEligible": false }, "launchResult": { "completedFullUtcDays": 0, "requiredFullUtcDays": 7, "sourceBoundCollectors": 1, "launchObservations": 2, "snapshotSeeded": true, "transportGaps": 0, "transportReconnects": 0, "protocolRefusals": 0, "modelFeatureWrites": 0, "labelWrites": 0, "calibrationReads": 0, "finalTestOpens": 0, "modelFits": 0, "orders": 0, "trades": 0, "capitalUsd": 0, "spendUsd": 0, "policyAction": "ABSTAIN", "contractSha256": "12803a8c692ca3d719a58ec16517c5e3781683b7b9934178410b7f05b431aee6", "launchArtifactSha256": "9bf696213cb26119a2c874ad108d441d91ffa6beb13b69e80f939af752eab4b5" }, "controls": { "focusedSuccessorChecksPassed": "6/6", "broaderAdjacentChecksPassed": "22/23", "broaderAdjacentRefusals": 1, "refusalReason": "historical source-admission adapter detected a drifted launch binding", "adapterResultPromoted": false, "scienceResultPromoted": false }, "supportedClaim": "A fresh source-bound raw collector launched under the frozen successor contract without carrying prior rows or opening downstream gates.", "notSupported": [ "A completed seven-day window or source admission", "An admitted model corpus, model fit, calibration, signal, or final-test result", "A forecast, market edge, profitability, execution result, or trading readiness", "Production reliability, automatic host recovery, credentials, orders, trades, capital, or spend" ] } } ], "verification": { "performedAt": "2026-07-26T20:32:02-04:00", "environment": "local macOS research workstation", "runs": [ { "program": "Waggle + Kea", "scope": "strict typecheck, three-task service suite, and adversarial refusal suite", "result": "PASS", "observed": "3/3 service tasks passed; 11/11 adversarial inputs rejected or failed closed" }, { "program": "Waggle + Kea", "scope": "accepted formal replay", "result": "PASS_ZERO_INFERENCE", "observed": "20 pins and 3 tasks verified; 8 accepted and 2 preserved interrupted-trial loads accounted; 0 added loads or forward/decode steps" }, { "program": "Waggle + Kea", "scope": "prior-checkpoint replay and frozen build", "result": "PASS", "observed": "prior replay added 0 loads or steps; production build passed from the committed snapshot" }, { "program": "HEATWAKE", "scope": "focused successor contract, collection, custody, authority, and refusal controls", "result": "PASS", "observed": "6/6 focused checks passed" }, { "program": "HEATWAKE", "scope": "broader adjacent historical adapter set", "result": "REFUSED", "observed": "22/23 checks passed; one historical adapter refused a drifted launch binding; no adapter or science result promoted" } ], "verificationEffects": { "providerCalls": 0, "newLocalModelLoads": 0, "newForwardDecodeSteps": 0, "trainingRuns": 0, "paidRequests": 0, "emailSends": 0, "newMarketCaptureRequests": 0, "newMarketBytesCaptured": 0, "heatwakeSourceWrites": 0, "heatwakeFinalTestOpens": 0, "orders": 0, "trades": 0, "capitalUsd": 0, "authorityGrants": 0, "sourceDeployments": 0 } }, "releaseNonClaims": [ "Not peer reviewed or independently validated", "No learned language, neuralese, hidden-state semantic transfer, or unique native capability", "No semantic, quality, token, Credit, cost, storage, energy, wall-time, or overall-efficiency advantage", "No completed HEATWAKE seven-day window, source admission, admitted model corpus, or final-test result", "No forecast, signal, market edge, profitability, execution result, trading readiness, or production reliability", "No customer adoption, revenue, commercial validation, deployment, or authority claim" ] }