{ "schemaVersion": "ke.research-findings-evidence.v1", "release": { "id": "ke-research-findings-2026-07-28", "title": "Exact parity held; the timing advantage did not", "date": "2026-07-28", "generatedAt": "2026-07-28T09:30:00-04:00", "informationCutoff": "2026-07-28T09:06:58-04:00", "canonicalUrl": "https://kestudios.dev/research/updates/2026-07-28", "citationUrl": "https://kestudios.dev/research/updates/2026-07-28/CITATION.cff", "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"email": "william@kestudios.dev" }, "reviewStatus": "source-verified local research update; not peer reviewed or independently validated" }, "sourceSnapshots": [ { "label": "Waggle and Kea committed snapshot", "commit": "4ec017ee5bc4e02b4ad51f949950355ed708c393", "capturedAt": "2026-07-28T09:06:58-04:00", "inspectionBasis": "bytes addressable from the frozen commit only", "artifactSha256": [ "a6e1b4dea31fbf0c2325e20a96e2d0e8592a28f2066394b4490600ee0c16d7a5", "bca26c33d0459ea77232280fcabbd97b312bf838fe1c28e1cb082f7472b3ae8d", "7c438aaef099b671a27e4f3656487a95f252ff2201f9279c83953fad6f37ef92", "b8943a6c7f066a635e7154b84597592a05f781091209484af82b67a42358fac5", "ba66ddaa63ef77617e3a5abaebbc80813910cdbd30654af23648b4ed4c8ed644", "f9956bd8da88b4a23c6055f6af56be29b73b46091a52d95e38cd3ae7e5bc475c" ] }, { "label": "HEATWAKE committed snapshot", "commit": "5ed9256a4f48154b44470e238501a89b357efd9a", "capturedAt": "2026-07-28T09:06:58-04:00", "inspectionBasis": "bytes addressable from the frozen commit only; active and later source state excluded", "artifactSha256": [ "a9570acbe4c2db4c1c9798388f85184e72d3f194263fc325290f3bbb2f2d2cb2" ] } ], "acceptedPublicBaseline": { "date": "2026-07-27", "canonicalUrl": "https://kestudios.dev/research/updates/2026-07-27", "evidenceSha256": "267b040af2bf0f809e169206efbd3f399be470b8b2ab90274bc3c49c600c79a9", "comparisonPolicy": "Only claim, method, control, risk, limitation, or dependency changes after this accepted checkpoint were eligible." }, "methodology": { "comparisonPolicy": "Inspect the accepted public checkpoint first, preregister claim IDs and falsifiers, then evaluate only the frozen committed bytes available at the information cutoff.", "evidenceLevels": { "recorded": "present in an attributable frozen artifact", "observed": "measured during the accepted bounded source run", "verified": "reproduced by exact artifact checks, clean isolated replay, or focused local tests", "independentValidation": false }, "environment": "local macOS research workstation; one local open-weight model family and local hardware-accelerated runtime for the paired Waggle and Kea comparisons", "pairedControlPolicy": { "sameModelProcessWithinBlock": true, "sameFrozenTasksAndExpectedValues": true, "equalInputTokenCount": true, "counterbalancedPreparationAndQueryOrder": true, "sourceSeparatedKeaQualification": true, "noStateControl": "abstain", "ambiguousControl": "insufficient_confidence", "timingAcceptance": "native must reach and retain break-even inside every primary block" }, "uncertainty": [ "The paired comparisons use deterministic numerical fixtures, one local model family, one host, and descriptive wall-time measurements.", "No statistical confidence interval, independent laboratory replication, or direct energy metric is available.", "A lower aggregate time is insufficient when any primary block fails the preregistered sustained-break-even rule.", "The opaque representation had a materially larger storage footprint than the readable control in the reported runs.", "The local reproduction package is unlicensed and unpublished.", "The citation record is first-party metadata; no DOI, archive deposit, external repository publication, or independent endorsement exists.", "HEATWAKE clock-day completion is not day-coverage acceptance, source-sequence completeness, source admission, or model utility." ], "exclusions": [ "All uncommitted, active, lock-held, or later source-writer bytes after the cutoff", "All active HEATWAKE raw events, segment receipts, and collector state", "All HEATWAKE feature, label, calibration, fit, candidate, signal, and sealed final-test data", "Any model acquisition, weight download, model load, provider call, or compatibility execution", "Any post-result change to a target, metric, threshold, or final-test boundary" ], "publicSafety": "Public output was composed from accepted aggregate claims, non-reversible hashes, public metadata, and bounded measurements. Non-public implementation material, raw prompts, raw model state, market payloads, private topology, credentials, and reconstruction-enabling detail are omitted." }, "programs": [ { "id": "waggle-kea", "name": "Waggle + Kea", "verdict": "Exact bounded behavior reproduced across four paired runs, while both longer-horizon controls failed the preregistered sustained-timing rule. The local reproduction package passed, but remains unlicensed and unpublished. No overall-efficiency or new model-family execution claim is supported.", "claimDeltas": [ { "claim": "Native context has a sustained timing advantage over an equally informed cached-prefix control.", "priorStatus": "no general timing advantage; one narrow break-even existed only against repeated full-text reconstruction", "newEvidence": "two 64-decision runs observed lower native logical wall; two 128-decision runs each failed sustained break-even in one primary block; one preserved first trial was slower", "currentStatus": "rejected under the preregistered all-block rule" }, { "claim": "Native and cached-prefix paths preserve exact bounded behavior through independent Kea qualification.", "priorStatus": "supported across six bounded fixtures", "newEvidence": "384/384 exact paired decisions across four frozen runs with continue, abstain, and insufficient-confidence dispositions preserved", "currentStatus": "supported across the reviewed paired fixtures" }, { "claim": "The result has a clean local reproduction package.", "priorStatus": "formal replay pinned; no canonical local release candidate", "newEvidence": "isolated offline reproduction, 49 packed files, 20/20 adversarial refusals, and three consumer dispositions", "currentStatus": "local release candidate verified; unlicensed and not published" }, { "claim": "A new model-family compatibility experiment is ready to run.", "priorStatus": "no eligible local input or exact candidate", "newEvidence": "one approximately 15.8 GB low-precision public artifact identified; zero downloads and zero model loads", "currentStatus": "candidate selected; fit and compatibility unverified" } ], "pairedRuns": [ { "label": "short paired run", "primaryBlocks": 4, "decisionsPerBlock": 16, "exactPairedDecisions": "64/64", "nativeLogicalWallMilliseconds": 119076.106456, "cachedPrefixLogicalWallMilliseconds": 132504.150463, "nativeAggregateDifferencePercent": -10.134055, "allBlocksSustainedBreakEven": true, "retainedNativeStateBytes": 279321005, "readablePrefixBytes": 8518, "directEnergyAvailable": false, "artifactSha256": "a6e1b4dea31fbf0c2325e20a96e2d0e8592a28f2066394b4490600ee0c16d7a5" }, { "label": "fresh isolated short reproduction", "primaryBlocks": 4, "decisionsPerBlock": 16, "exactPairedDecisions": "64/64", "nativeLogicalWallMilliseconds": 124884.688209, "cachedPrefixLogicalWallMilliseconds": 129318.796374, "nativeAggregateDifferencePercent": -3.42882, "allBlocksSustainedBreakEven": true, "retainedNativeStateBytes": 90727353, "readablePrefixBytes": 1744, "directEnergyAvailable": false, "preservedExcludedPreModelTrials": 4, "artifactSha256": "bca26c33d0459ea77232280fcabbd97b312bf838fe1c28e1cb082f7472b3ae8d" }, { "label": "long-horizon paired control", "primaryBlocks": 4, "decisionsPerBlock": 32, "exactPairedDecisions": "128/128", "nativeLogicalWallMilliseconds": 261756.274007, "cachedPrefixLogicalWallMilliseconds": 265376.797755, "nativeAggregateDifferencePercent": -1.364296, "blocksFailingSustainedBreakEven": 1, "timingAdvantageDemonstrated": false, "broaderContextCapacityDemonstrated": false, "artifactSha256": "7c438aaef099b671a27e4f3656487a95f252ff2201f9279c83953fad6f37ef92" }, { "label": "fresh isolated long-horizon reproduction", "primaryBlocks": 4, "decisionsPerBlock": 32, "exactPairedDecisions": "128/128", "nativeLogicalWallMilliseconds": 273656.952133, "cachedPrefixLogicalWallMilliseconds": 277045.392082, "nativeAggregateDifferencePercent": -1.223063, "blocksFailingSustainedBreakEven": 1, "preservedFirstTrialNativeDifferencePercent": 0.78476, "timingAdvantageDemonstrated": false, "crossTrialTimingAdvantageClaimed": false, "artifactSha256": "b8943a6c7f066a635e7154b84597592a05f781091209484af82b67a42358fac5" } ], "aggregateDecisionAccounting": { "pairedRuns": 4, "exactPairedDecisions": "384/384", "longHorizonRuns": 2, "longHorizonRunsPassingAllBlocks": "0/2", "restrictedConsumerDisposition": "continue", "noStateDisposition": "abstain", "ambiguousDisposition": "insufficient_confidence", "sustainedTimingAdvantageDemonstrated": false, "overallEfficiency": false }, "localReleaseCandidate": { "releaseCandidateReady": true, "openSourceReady": false, "licenseStatus": "unlicensed", "published": false, "manifestPayloadFiles": 47, "packedFiles": 49, "adversarialCasesRejected": "20/20", "exactReproductionDecisions": "128/128", "modelLoadsDuringOfflineCheck": 0, "networkCallsDuringOfflineCheck": 0, "authorityEffects": 0, "artifactSha256": "ba66ddaa63ef77617e3a5abaebbc80813910cdbd30654af23648b4ed4c8ed644" }, "futureCompatibilityCandidate": { "selectionBasis": "public model and artifact metadata", "baseModelFamily": "Kimi Linear", "declaredTotalParameters": "48B", "declaredActiveParameters": "3B", "selectedArtifactFormat": "GGUF", "selectedQuantization": "IQ2_M", "selectedArtifactBytes": 15792109472, "memoryFit": "plausible_unverified", "localEligibilityBeforeSelection": false, "downloads": 0, "modelLoads": 0, "compatibilityChecks": 0, "providerCalls": 0, "spendUsd": 0, "artifactSha256": "f9956bd8da88b4a23c6055f6af56be29b73b46091a52d95e38cd3ae7e5bc475c" }, "notSupported": [ "A sustained timing advantage, speedup, or overall-efficiency result", "Token, Credit, provider-cost, memory, storage, network, or energy savings", "A learned language, hidden-state semantic transfer, unique capability, or higher output quality", "Independent validation, a statistical confidence interval, or a causal timing explanation", "An open-source release or software-license grant", "A new model-family load, state round trip, semantic result, customer result, production result, or authority execution" ] }, { "id": "heatwake", "name": "HEATWAKE", "verdict": "The first of seven required UTC clock days is recorded, but day coverage is unaudited and unpassed, source-sequence completeness is not claimed, and the terminal receipt is absent. Source admission and every model, signal, edge, profitability, trading, and production claim remain closed.", "claimDeltas": [ { "claim": "HEATWAKE has one complete admitted day toward its seven-day requirement.", "priorStatus": "raw custody only at 0/7 complete days", "newEvidence": "clock counter 1/7; day coverage unaudited and unpassed; 18 transport gaps; 18 reconnects; no sequence-completeness claim; no terminal receipt", "currentStatus": "recorded clock day only; no coverage or source-admission upgrade" } ], "custodyClock": { "completedUtcClockDays": 1, "requiredUtcClockDays": 7, "dayCoverageAudited": false, "dayCoveragePassed": false, "collectionComplete": false, "sourceSequenceCompletenessClaimed": false, "terminalSessionReceiptPresent": false, "transportGaps": 18, "transportReconnects": 18, "protocolRefusals": 0, "rawSourceAdmitted": false, "modelInputEligible": false, "policyAction": "ABSTAIN", "artifactSha256": "a9570acbe4c2db4c1c9798388f85184e72d3f194263fc325290f3bbb2f2d2cb2" }, "controls": { "focusedLifecycleAndSupervisorTestsPassed": "39/39", "receiptRecordHashVerified": true, "receiptFileDigestVerified": true, "sourceCheckoutUnchanged": true }, "effects": { "activeEventReads": 0, "segmentReceiptReads": 0, "sourceWrites": 0, "collectorRestarts": 0, "providerReads": 0, "sourceAdmissions": 0, "modelFeatureWrites": 0, "labelWrites": 0, "modelFits": 0, "finalTestOpens": 0, "orders": 0, "trades": 0, "capitalUsd": 0, "spendUsd": 0, "deployments": 0 }, "notSupported": [ "Audited or passing day coverage, a completed seven-day collection, terminal receipt, or source-admission result", "An admitted model corpus, feature or label construction, model fit, calibration, candidate selection, or final-test result", "A forecast, signal, market edge, profitability, execution result, or trading readiness", "Production reliability, credentials, orders, trades, capital, spend, or release authority" ] } ], "falsifiers": [ "A sustained timing claim requires a fresh preregistered paired run to pass break-even in every primary block and an independent repetition to preserve the result.", "An overall-efficiency claim additionally requires matched end-to-end memory, storage, energy, and cost accounting against the strongest readable control.", "A model-family compatibility claim requires lawful acquisition, hash verification, a local load, recurrent-state save and restore, and a matched cached-prefix check.", "A HEATWAKE coverage claim requires the frozen audit to pass exact daily coverage and terminal-receipt checks; any scientific claim requires later separately gated stages." ], "publicReferences": [ { "title": "Kimi Linear 48B-A3B-Instruct model card", "url": "https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct", "use": "declared base-model architecture, parameter counts, context length, and license metadata" }, { "title": "Kimi Linear GGUF artifact index", "url": "https://huggingface.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF", "use": "public derivative artifact size and quantization metadata; not an execution or quality endorsement" } ], "verification": { "performedAt": "2026-07-28T09:30:00-04:00", "environment": "local macOS research workstation", "runs": [ { "program": "Waggle + Kea", "scope": "four frozen paired evidence artifacts", "result": "PASS_WITH_NEGATIVE_TIMING_VERDICT", "observed": "384/384 exact paired decisions; both 128-decision runs failed sustained break-even in one primary block" }, { "program": "Waggle + Kea", "scope": "isolated local release candidate", "result": "PASS_UNLICENSED", "observed": "128/128 exact reproduction; 20/20 adversarial cases rejected; 49 packed files; no model or network work" }, { "program": "Waggle + Kea", "scope": "strict source checks", "result": "PASS", "observed": "strict type checking and production build passed from the frozen committed source" }, { "program": "Waggle + Kea", "scope": "future compatibility candidate boundary", "result": "PASS_NO_EXECUTION", "observed": "candidate metadata verified; 3/3 adversarial mutations rejected; zero downloads, model loads, provider calls, or spend" }, { "program": "HEATWAKE", "scope": "first-day custody receipt", "result": "PASS_RAW_ONLY", "observed": "record hash and file digest matched; 1/7 clock days; coverage unaudited and unpassed; 18 gaps; 18 reconnects; no terminal receipt" }, { "program": "HEATWAKE", "scope": "focused lifecycle and supervisor controls", "result": "PASS", "observed": "39/39 tests passed from the unchanged frozen checkout" } ], "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 sustained timing advantage, speedup, savings, or overall-efficiency result", "No learned language, hidden-state semantic transfer, unique capability, or higher output quality", "No open-source release or software-license grant", "No new model-family execution, compatibility, semantic, or quality result", "No audited HEATWAKE day coverage, completed seven-day collection, terminal receipt, 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, authority execution, or source-program deployment" ] }