{ "schemaVersion": "ke.research-findings-evidence.v1", "release": { "id": "ke-research-findings-2026-07-15", "title": "What survived the controls", "date": "2026-07-15", "generatedAt": "2026-07-15T21:01:57-04:00", "canonicalUrl": "https://kestudios.dev/research/updates/2026-07-15", "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"email": "william@kestudios.dev" }, "reviewStatus": "source-verified local research update; not peer reviewed" }, "methodology": { "claimRule": "A claim is supported only inside the tested fixture, model, runtime, host, and control envelope. Negative results remain results. Untested extensions are listed as not demonstrated.", "sources": [ "Frozen JSON reports and receipts", "Preregistrations and promotion policies", "Record and file SHA-256 digests", "Exact source commits", "Fresh local reruns of the decisive gates" ], "publicSourceBoundary": "The research repositories were private at release time. This ledger therefore does not represent inaccessible GitHub URLs as public source access.", "externalReview": false }, "sourceSnapshots": [ { "program": "Waggle + Kea", "commit": "180bd10fe17142fce6e6f4a1a79ae2d85b777ed7", "worktreeStateAtVerification": "clean", "publicProgramUrl": "https://kestudios.dev/research/waggle" }, { "program": "HEATWAKE", "commit": "57ca6bb4209e9679e02bb7aa06cc3dfed27f95d4", "worktreeStateAtVerification": "clean", "publicProgramUrl": "https://kestudios.dev/research/heatwake" } ], "programs": [ { "id": "waggle-kea", "name": "Waggle + Kea", "verdict": "Supported at fixture scale: deterministic replay, exact same-model native-state continuation, bounded same-host reuse utility, and pre-inference abstention. Not supported: a universal language, compression, or overall efficiency.", "experiments": [ { "id": "C3-LATENT-5", "question": "Can a model consume a serialized numeric channel under blinded conditions?", "design": { "paidModelCalls": 24, "arms": [ "serialized numeric channel", "one-byte symbolic control", "no channel" ], "tasksPerArm": 8 }, "result": { "numericExactArtifacts": "8/8", "symbolicExactArtifacts": "8/8", "noChannelExactArtifacts": "0/8", "noChannelSafeAbstentions": "8/8", "numericTotalTokens": 13631, "symbolicTotalTokens": 11662, "numericAdditionalTokens": 1969, "numericAdditionalTokensPercent": 16.9 }, "supportedClaim": "The tested serialized channel was consumable in this fixture.", "notSupported": [ "Unique latent-language advantage", "Token savings", "Quality advantage", "Production utility" ] }, { "id": "C3-LATENT-6", "question": "Did the prior numeric channel contain irreducible semantics?", "result": { "allTestedVectorsLosslesslyRecodable": true, "fixtureSemanticsFitOneByte": true, "paidFollowupEligible": false, "paidFollowupCalls": 0, "certificateFileSha256": "bec27c16dabb679864b4dc08d82beedf0d56d1f32f518f2805855b00bb418804" }, "supportedClaim": "The fixture did not establish a unique serialized latent language.", "notSupported": [ "General impossibility of latent communication", "Cross-model conclusion" ] }, { "id": "C6r", "question": "Can a fresh local process restore Qwen3-14B native state and exactly preserve the tested output capability?", "model": { "repository": "Qwen/Qwen3-14B-GGUF", "revision": "530227a7d994db8eca5ab5ced2fb692b614357fd", "quantization": "Q6_K", "runtime": "llama.cpp public C API", "runtimeRevision": "ed4837891" }, "result": { "cases": 8, "semanticParity": "8/8", "byteParity": "8/8", "promptTokenParity": "8/8", "totalStateBytes": 333229432, "totalOriginalPromptBytes": 7270, "stateToPromptByteRatioApprox": 45836, "reportSha256": "0c2daa36d97e41ca24669914a22b70c968e3c33dc46bd019d7f37259702aea59" }, "supportedClaim": "The tested same-model native state preserved source-grounded capability across a fresh process.", "notSupported": [ "Compression", "Independent semantic decoding of opaque state", "Cross-model language", "General capability", "Overall efficiency" ] }, { "id": "C6s", "question": "Does one same-host native state become faster than repeated full-text prefill under fan-out?", "design": { "sharedPrefixTokens": 1701, "sharedPrefixBytes": 8518, "nativeExecutors": 8, "fullTextControls": 8, "pairOrder": "counterbalanced alternating", "hostRuns": 1 }, "result": { "semanticParity": "8/8", "byteParity": "8/8", "tokenParity": "8/8", "coldBreakEvenTask": 4, "task8NativeColdWallMs": 36909.21225, "task8FullTextColdWallMs": 53269.589208, "task8ColdWallReductionPercent": 30.7124, "stateBytes": 279321005, "totalRetainedObjectBytes": 279322905, "allEightFullTextLogicalHandoffBytes": 71215, "retainedObjectToAllFullTextHandoffsRatioApprox": 3922.2, "reportSha256": "5ed6c9a1921466375d85a769a59ac86c509edfa1aded8e40e4d39854900d60a8" }, "supportedClaim": "One controlled local run showed exact parity and a same-host cold-latency break-even at the fourth task.", "uncontrolled": [ "OS cache", "Thermal state", "Energy", "Cross-host transfer", "Repeated-run variance" ], "notSupported": [ "Storage efficiency", "Network efficiency", "Overall efficiency", "Production economics" ] }, { "id": "C6t", "question": "Can the passing native-state mechanism be operated as a bounded local workspace lifecycle?", "result": { "workflow": [ "capture", "inspect", "fork", "replay", "release", "reclaim" ], "acceptedForks": 2, "reclaimedObjectBytes": 279322905, "retainedObjectBytesAfterCleanup": 0, "immutableReplay": true }, "supportedClaim": "A founder-runnable fixture workflow can capture, fork, verify, release, and reclaim the exact object.", "notSupported": [ "Live Agents", "Production workspace", "Authority" ] }, { "id": "C6v", "question": "Will the workspace continue supported held-out tasks and refuse unsupported tasks before inference?", "design": { "newContexts": 3, "tasks": 5 }, "result": { "supportedExactContinuations": 3, "preExecutorAbstentions": 2, "modelProcessesForAbstentions": 0, "totalLocalModelProcesses": 6, "totalDecodeCalls": 76, "paidProviderCalls": 0, "externalCalls": 0, "authorityEffects": 0, "stateBytesCreated": 52128823, "nativeObjectBytesReclaimed": 52134507, "nativeBytesRetained": 0, "reportSha256": "ad3ee1e326375c1737dff2978760ac627fc5f71205428c5af35306ffd3b84f1c" }, "recoveryDisclosure": "The packaging step removed three required private ledgers after the formal result. Only explicitly receipt-equivalent, unhash-bound ledgers were reconstructed. No additional inference occurred and artifact, receipt, evaluation, and report identities did not change.", "supportedClaim": "The declared-source fixture continued three supported tasks exactly and refused two unsupported tasks before Executor launch.", "notSupported": [ "Broad generalization", "Arbitrary reasoning", "Production traffic" ] }, { "id": "L1-L2", "result": { "libraryBuildImport": "PASS", "publicEntrypoints": [ "grokcode/waggle", "grokcode/kea", "grokcode/waggle-kea" ], "fixtureOnly": true, "liveTraffic": false, "authorityGranted": false, "authorityGateSigned": false }, "supportedClaim": "The fixture API is packaged and importable; the live authority gate remains unsigned." } ], "programNonClaims": [ "No learned Waggle", "No universal neuralese", "No independent semantic decoder of opaque model state", "No token, dollar, storage, network, energy, or overall-efficiency win", "No production traffic, customer result, revenue, deployment, or authority" ] }, { "id": "heatwake", "name": "HEATWAKE", "verdict": "Supported at synthetic fixture scale: deterministic corpus, two reproducible model families, frozen walk-forward and promotion controls, and safe refusal. Not supported: synthetic generalization, any empirical forecast, market edge, or profitability.", "experiments": [ { "id": "fabricated-superposition-corpus", "result": { "bundles": 17, "branchExamples": 51, "tensors": 17, "partitions": 4, "source": "deterministic authored synthetic data" }, "supportedClaim": "The mechanism-development corpus is content-addressed and reproducible.", "notSupported": [ "Observed exchange behavior", "External validity" ] }, { "id": "S3-11-walk-forward", "design": { "architectures": [ "causal visual TCN", "causal attention fusion" ], "expandingOriginFolds": 5, "laterBundlesPerArchitecture": 10, "walkForwardUniformComparatorLogLossNats": 1.098612321409, "promotionReplayTheoreticalLn3LogLossNats": 1.098612288668, "comparatorNote": "The walk-forward table and the exact theoretical promotion replay use slightly different numeric representations of the constant one-third target; both rank both learned models below uniform." }, "result": { "candidate": { "logLossNats": 1.098683870252, "deltaVsUniformNats": 0.000071548843, "brierSum": 0.000047706305, "foldsBeatingUniform": "0/5" }, "challenger": { "logLossNats": 1.098700995688, "deltaVsUniformNats": 0.000088674279, "brierSum": 0.000058968539, "foldsBeatingUniform": "0/5" }, "walkForwardSummaryRecordHash": "9ae0e0d69cc97abb67f05d9eb72e3c0dece4edc92f7cf78dc7c3a80bdf259f3e" }, "supportedClaim": "Both learned candidates were worse than uniform on the authored synthetic target.", "targetLimitation": "The authored target had constant one-third support, making uniform the theoretical optimum. This diagnoses a non-identifying synthetic target and is not a real-market conclusion." }, { "id": "S3-12-promotion", "result": { "eligibleModels": 0, "testedModels": 2, "failedGatesForBoth": [ "predictive identifiability", "every frozen baseline", "uniform in every walk-forward fold", "shuffled-label discrimination", "calibration transfer", "mandatory policy coverage" ], "policyAction": "ABSTAIN", "learnedSelection": null, "finalTestState": "SEALED_UNOPENED", "promotionDecisionRecordHash": "59b5d58b4f80cfd332ed0c42d9150e6fa962e9ae092830325cf8043727880bda" }, "supportedClaim": "The frozen policy rejected both models and selected none.", "externalEffects": { "capitalUsd": 0, "orders": 0, "trades": 0, "deployments": 0, "paidRequests": 0, "externalSends": 0, "finalTestOpens": 0 } }, { "id": "DG-13-founder-signal-journey", "design": { "entrypoint": "one stock-python command into the reviewed Python 3.12 environment", "coldDeterministicBuilds": 2, "stages": [ "custody", "environment", "train", "evaluate", "compare", "signal", "replay" ] }, "result": { "fixtureMechanicsFits": 60, "untouchedFixtureEvaluations": 30, "negativeControlEvaluations": 620, "realisticCostPaperReceipts": 5376, "deterministicSignalEnvelopes": 48, "policyActionForEveryPath": "ABSTAIN", "checkpointRestoresPassed": 30, "coldBuildsByteIdentical": true, "selfConsistentLineageTamperRefused": true, "acceptedBytesRestored": true, "currentProjection": "ABSTAIN", "staleProjection": "ABSTAIN", "workerCount": 1, "wallNanoseconds": 37580939083, "peakRssBytes": 996638720, "networkRequests": 0, "recordHash": "1d97ec60dd42c497c6269e1c43b821c58b742c222828d1619d76316461812c0f", "status": "PASS_SAFE_LOCAL_FIXTURE_JOURNEY_EMPIRICAL_GATE_CLOSED" }, "supportedClaim": "The existing fixture machinery and operator contract can execute and recover coherently through one bounded local journey.", "notSupported": [ "Empirical market model", "Calibrated directional signal", "Market edge", "Profitability", "Trading readiness", "Production readiness" ] }, { "id": "empirical-readiness", "implementedMachinery": [ "source admission", "offline Hyperliquid normalization adapter", "causal multiscale feature volumes", "900-second leakage-safe target split", "baseline evaluation", "candidate evaluation", "walk-forward evaluation", "negative controls", "paper economics", "signal envelope", "safe local operator" ], "empiricalState": { "admittedThirdPartyBytes": 0, "empiricalSamples": 0, "empiricalTrainingRuns": 0, "empiricalEvaluationRuns": 0, "empiricalInferenceRuns": 0, "empiricalFinalTestOpens": 0, "commercialPublicRightsCleared": false }, "storageBoundary": { "selectedArchiveBytesBeforeExtraction": 75610200000, "internalFreeBytes": 68453412864, "deficitBytes": 7156787136, "requiredExternalVolume": "encrypted APFS SSD, at least 1 TB capacity and at least 850 GB free", "eligibleExternalDiskObserved": false, "archiveRequestMade": false }, "supportedClaim": "Empirical machinery has implementation and test evidence, but no empirical result exists." }, { "id": "operator", "result": { "scope": "safe local research operator", "syntheticStates": 3600, "depthLevels": 96, "views": "1 second through 1 hour", "dataClock": "SYNTHETIC", "forecastProvenance": "MOCK", "policyAction": "ABSTAIN", "actionRail": "DISABLED", "webTests": 78 }, "supportedClaim": "The negative research state is operable and inspectable locally.", "notSupported": [ "Research validity from UI completeness", "Live data", "Trading readiness", "Production readiness" ] } ], "programNonClaims": [ "No admitted live or historical market corpus", "No empirical fit or generalization", "No validated forecast, market edge, or profitability", "No order, trade, capital deployment, or production readiness", "No cleared commercial or public redistribution rights for the selected historical source" ] } ], "verification": { "performedAt": "2026-07-15T21:01:57-04:00", "environment": "local macOS research workstation", "runs": [ { "program": "Waggle + Kea", "command": "npm run test:waggle-kea-qwen3-heldout-context-gate", "result": "PASS", "observed": "3 continued, 2 abstained, 2 zero-model preflights, 6 local processes, 76 decode calls, 0 paid/external/authority effects; recovery added 0 inference calls" }, { "program": "Waggle + Kea", "command": "npm run test:waggle-kea-qwen3-14b-native-fanout", "result": "PASS", "observed": "protocol and report hashes matched; break-even 4; stateBytes 279321005; immutable replay used 0 model calls and 0 effects" }, { "program": "Waggle + Kea", "command": "npm run test:waggle-kea-latent-equivalence", "result": "PASS", "observed": "9 checks; 8/8 bit and semantic equivalence; follow-up ineligible; 0 calls" }, { "program": "Waggle + Kea", "command": "npm run test:waggle-kea-library-api", "result": "PASS", "observed": "build/import, exact replay and correction, fixture-only boundaries, and package artifact verified" }, { "program": "Waggle + Kea", "command": "npm run typecheck", "result": "PASS" }, { "program": "Waggle + Kea", "command": "npm run build", "result": "PASS" }, { "program": "HEATWAKE", "command": "bash scripts/verify.sh", "result": "STOPPED_HARNESS_DEPENDENCY", "observed": "The legacy verifier discovered a pytest module, but its base Python environment did not declare pytest. No product assertion failed before the stop. Focused suites were then run in compatible environments." }, { "program": "HEATWAKE", "command": "PYTHONPATH=src:ml:. python3 -m unittest tests.test_signal_promotion tests.test_signal_walk_forward -v", "result": "PASS", "observed": "9/9 tests" }, { "program": "HEATWAKE", "command": "PYTHONPATH=:src:ml:. python3 -m pytest -q tests/test_empirical_paper_economics.py tests/test_empirical_acquisition_runner.py", "disclosure": "The non-semantic temporary absolute path was redacted; the pytest version and executable test targets are preserved.", "result": "PASS", "observed": "16/16 tests; pytest was installed only into a temporary verification directory because system Python supplied the required torch runtime" }, { "program": "HEATWAKE", "command": "./apps/operator-service/.venv/bin/pytest -q apps/operator-service/tests", "result": "PASS", "observed": "operator service suite passed" }, { "program": "HEATWAKE", "command": "npm run verify", "workingDirectory": "apps/operator", "result": "PASS", "observed": "lint, typecheck, 78 tests across 25 files, and Next.js production build" } ], "verificationEffects": { "paidProviderCalls": 0, "emailSends": 0, "orders": 0, "trades": 0, "deploymentsFromResearchHarnesses": 0, "authorityGrants": 0, "heatwakeFinalTestOpens": 0, "newEmpiricalBytesAdmitted": 0 } }, "releaseNonClaims": [ "Not peer reviewed", "No universal agent language", "No native-state compression or overall-efficiency claim", "No HEATWAKE empirical market result", "No forecast, market edge, profitability, trading, or production claim", "No customer adoption, revenue, or commercial validation claim" ] }