Jump to a findings section
The behavior reproduced; the performance claim failed its stronger controls.
Four frozen paired comparisons returned 384/384 exact native and cached-prefix decisions. Two shorter runs observed 10.134% and 3.429% lower native logical wall time. The two stronger 128-decision controls then narrowed the aggregate differences to 1.364% and 1.223%, and each failed sustained break-even in one primary block; a preserved first trial was 0.785% slower. Exact behavior therefore reproduced, while the timing-advantage predicate remained false. The same snapshot produced an isolated 49-file local package that passed 20/20 misuse checks, but it remains unlicensed and unpublished. A future local model candidate was identified from public metadata without downloading or running it. Separately, HEATWAKE recorded one of seven required UTC clock days, but day coverage was neither audited nor passed, 18 transport gaps and reconnects were recorded, and source admission plus every downstream model, final-test, signal, order, and trading gate remained closed.
Program record
Waggle + Kea
Supported in scope: exact native and cached-prefix behavior, Kea qualification, restricted-consumer continuation, refusal controls, and an isolated local reproduction package. Rejected: a sustained timing advantage and overall efficiency. Not established: a new model-family execution result, open-source release, unique semantics, customer value, production behavior, or authority.
Open programProgram record
HEATWAKE
Supported: an immutable receipt records the first of seven required UTC clock days and focused lifecycle controls reproduce. Not supported: audited or passing day coverage, complete sequence custody, a terminal session, source admission, model eligibility, a forecast, signal, edge, profit, trading readiness, or production operation.
Open programAtomic claim deltas.
Swipe horizontally to compare all four columns.
| Claim | Prior status | New evidence | Current status |
|---|---|---|---|
| Native context has a sustained timing advantage over an equally informed cached-prefix control. | One six-branch run supported only a narrow break-even against repeated full text; cached-prefix text remained faster. | Two short paired runs returned 64/64 exact decisions with 10.134% and 3.429% lower native logical wall time. Two longer runs returned 128/128 exact decisions each but failed sustained break-even in one primary block; a preserved first trial was 0.785% slower. | Rejected under the preregistered all-block rule. |
| Native and cached-prefix paths preserve exact bounded behavior through independent Kea qualification. | Supported across six fixtures and five informed arms. | Four paired runs returned 384/384 exact native and cached-prefix decisions; no-state abstained, ambiguous input returned insufficient confidence, and the restricted consumer continued from the qualified projection. | Supported across these frozen paired fixtures. |
| The Waggle + Kea result has a clean local reproduction package. | Formal replay was pinned, but no canonical local release candidate existed. | An isolated offline check reproduced the accepted 128-decision result, packed 49 files, rejected 20/20 adversarial cases, and preserved continue, abstain, and insufficient-confidence dispositions. | Local release candidate verified; unlicensed and not published. |
| A new model-family compatibility experiment is ready to run. | No eligible local input or exact candidate had been established. | The local eligibility gate found no qualifying installed model. Public metadata identified one approximately 15.8 GB low-precision candidate, but no weight was downloaded and no load, state round trip, or compatibility check ran. | Candidate selected; execution and compatibility remain unverified. |
| HEATWAKE has one complete admitted day toward its seven-day requirement. | Raw custody only at 0/7 complete days. | The clock-day receipt records 1/7, but explicitly marks day coverage unaudited and unpassed, with 18 transport gaps, 18 reconnects, no sequence-completeness claim, and no terminal receipt. | Recorded clock day only; no coverage or source-admission upgrade. |
Claims, controls, receipts, reruns.
Frozen information cutoff
The review fixed one committed Waggle/Kea snapshot and one committed HEATWAKE snapshot at July 28, 2026 at 09:06:58 EDT. Uncommitted, active, lock-held, and later source bytes were excluded.
Matched paired comparison
Native and cached-prefix arms used the same model process, equal input-token counts, counterbalanced block order, fixed expected values, and source-separated Kea qualification. No-state and ambiguous controls tested refusal behavior.
Preregistered performance rule
Aggregate wall time alone was insufficient. A timing advantage required sustained break-even in every primary block. Both longer 128-decision runs failed that rule despite slightly lower native aggregates.
Reproduction and resource accounting
Fresh isolated packages reproduced exact decisions and recorded retained state, readable-prefix size, peak memory, process loads, and unavailable energy metrics. The final local package was checked offline, adversarially, and through an isolated pack dry run.
HEATWAKE separation
The review verified only the immutable first-day custody receipt and focused lifecycle controls. It did not inspect active raw events, run source admission, construct features, bind labels, fit a model, open the sealed final test, produce a signal, or trade.
Waggle + Kea
Repeatable exact handoff with a failed sustained-timing hypothesis
Verdict
Supported in scope: exact native and cached-prefix behavior, Kea qualification, restricted-consumer continuation, refusal controls, and an isolated local reproduction package. Rejected: a sustained timing advantage and overall efficiency. Not established: a new model-family execution result, open-source release, unique semantics, customer value, production behavior, or authority.
This is a useful negative result. Shorter paired runs made the native path look faster, but the longer and independently reproduced controls failed the stronger all-block rule. The repeatable contribution is exact behavior plus auditable handoff and refusal—not durable speed. Packaging made that result easier to inspect locally without changing its evidence level or licensing status.
Exact paired behavior reproduced four times
64/64 + 64/64 + 128/128 + 128/128
Every accepted native and cached-prefix decision matched byte-for-byte and semantically with equal input-token counts. Kea qualified the handoff, the restricted consumer continued, no-state abstained, and ambiguity returned insufficient confidence.
Boundary: The runs use bounded deterministic numerical fixtures on one local model family and one host. Exact parity does not show unique semantics, better quality, broader reasoning, or production reliability.
The short-horizon timing observations did not generalize
10.134% and 3.429% lower native logical wall
Two 64-decision runs observed lower native logical wall time under their frozen paired definitions.
Boundary: These are descriptive single-host observations. The opaque representation had a materially larger storage footprint than the readable control, direct energy was unavailable, and the stronger longer-horizon rule—not the short-run aggregate—governs the claim.
Both longer controls rejected sustained timing advantage
128/128 twice · 1 failed block in each run
The 128-decision runs had 1.364% and 1.223% lower native aggregates, but each failed sustained break-even in one primary block. A preserved first trial was 0.785% slower.
Boundary: The preregistered rule required every primary block to pass. Aggregate differences cannot be promoted into a speed, savings, token, Credit, energy, memory, or overall-efficiency claim.
The result has a clean local package, not a public software release
49 packed files · 20/20 attacks rejected
A fresh isolated copy verified the manifest, public-safety boundary, exact reproduction, consumer dispositions, tests, and package inventory without model or network work.
Boundary: The package is unlicensed, not open-source ready, and not published. A passing package check does not upgrade the timing result or grant release authority.
The next model-family step stopped before execution
1 candidate selected · 0 downloads · 0 model loads
Public metadata supported a bounded future compatibility candidate after the local inventory found no eligible installed input.
Boundary: Memory fit is only plausible and unverified. No load, hardware compatibility, recurrent-state save and restore, semantic utility, provider result, or recommendation is claimed.
Not demonstrated
- No sustained timing advantage, overall efficiency, token, Credit, provider-cost, memory, storage, network, or energy savings.
- No learned language, hidden-state semantic transfer, unique capability, higher quality, broader reasoning, or cross-model result.
- No independent validation, statistical confidence interval, causal timing explanation, or direct energy measurement.
- No open-source release or license grant; the package remains a local candidate.
- No new model-family execution, compatibility, state-round-trip, semantic, customer, production, deployment, or authority result.
Next legitimate gates
- 1. A timing claim would require a fresh preregistered run in which every primary block sustains break-even and an independent repetition preserves the result.
- 2. An efficiency claim would additionally require matched memory, storage, energy, and end-to-end cost accounting against the strongest readable control.
- 3. Open-source readiness requires an explicit software-license decision and a separate release review.
- 4. The model-family question requires separate authority for the exact download, hash verification, bounded local load, state round trip, and matched cached-prefix compatibility test.
HEATWAKE
First raw custody clock day with coverage still unproven
Verdict
Supported: an immutable receipt records the first of seven required UTC clock days and focused lifecycle controls reproduce. Not supported: audited or passing day coverage, complete sequence custody, a terminal session, source admission, model eligibility, a forecast, signal, edge, profit, trading readiness, or production operation.
The day counter moved; the evidence level did not. The receipt itself records transport discontinuities and refuses to call the day coverage-complete. Because the frozen admission method requires complete terminal custody and seven qualifying days, the only truthful status remains raw collection with every downstream gate closed.
The clock recorded day one
1/7 required UTC clock days
The immutable receipt advanced the calendar counter from zero to one and preserved its own record hash and file digest.
Boundary: Clock completion is not coverage acceptance. The receipt explicitly marks day coverage unaudited and unpassed.
Transport continuity remains unresolved
18 gaps · 18 reconnects · no terminal receipt
The receipt records matched gap and reconnect counts and does not claim source-sequence completeness.
Boundary: A reconnecting collector may preserve custody observations without establishing complete market-data coverage, source quality, eligibility, or operational reliability.
Every downstream gate stayed closed
39/39 focused tests · 0 admissions · 0 final opens · 0 orders
Focused lifecycle and supervisor checks reproduced while source admission, feature construction, labels, fitting, the final test, signals, positions, orders, capital, and spend remained closed.
Boundary: These are exact effects for the frozen snapshot, not a broad security or production claim. The current policy action remains ABSTAIN.
Not demonstrated
- No audited or passing day coverage, completed seven-day collection, terminal receipt, source-admission result, or admitted model corpus.
- No feature or label construction, model fit, calibration, candidate selection, or sealed final-test result.
- No forecast, signal, market edge, profitability, execution result, trading readiness, or production reliability.
- No raw payload, private receipt, reconstruction detail, credential, order, trade, capital, spend, or release authority is published.
Next legitimate gates
- 1. A coverage claim requires the frozen audit to pass the day against exact receipt continuity, freshness, and completeness rules.
- 2. Source eligibility requires seven qualifying UTC days, the exact terminal receipt, and the already-frozen admission audit without threshold changes.
- 3. Only a passing source audit may open a separate feature and model-input decision; it cannot open the sealed final test or trading.
Falsifiers and revision conditions.
- 1Revise the sustained-timing conclusion only if a fresh preregistered paired run passes break-even in every primary block and an independent repetition preserves the result.
- 2Revise the overall-efficiency conclusion only if matched end-to-end memory, storage, energy, and cost accounting also beats the strongest readable control.
- 3Revise model-family compatibility only after the exact candidate is lawfully acquired, hash-verified, loaded in the pinned local runtime, and completes a recurrent-state save and restore plus matched cached-prefix check.
- 4Revise HEATWAKE custody status only after the frozen audit passes exact daily coverage and terminal-receipt checks; revise science only through later separately gated stages.
Limits on the reported findings.
- 1All verification described here is local source verification by K&E Studios Research, not peer review or independent validation.
- 2The Waggle/Kea comparisons use one local model family, one workstation, deterministic numerical fixtures, and descriptive logical-wall measurements. No statistical confidence interval or direct energy measurement is claimed.
- 3A lower aggregate wall time did not satisfy the preregistered timing claim when any primary block failed sustained break-even.
- 4The local reproduction package is unlicensed and unpublished. Its verification status is not an open-source release or third-party license grant.
- 5The future model candidate was identified from public metadata only. No model weight was downloaded, loaded, executed, or recommended for quality.
- 6HEATWAKE raw payloads, private receipts, active collector state, and reconstruction detail are not published. The review made no source-program write or active-data read.
- 7The citation record is a first-party attribution file. No DOI, external archive, repository publication, or independent endorsement is claimed.
- 8This publication verification made no paid request, provider call, model load, forward or decode step, email send, market-capture request, feature write, fit, final-test open, order, trade, capital change, or authority grant.
Rechecked from the source trees.
Waggle + Kea paired evidence
PASSFour frozen artifacts verified 384/384 exact paired decisions and the declared short- and long-horizon timing results. Both 128-decision runs retained a failed primary block, so the sustained-timing predicate remained false.
Waggle + Kea local release candidate
PASSThe isolated offline release check reproduced 128/128 decisions, rejected 20/20 adversarial cases, packed 49 files, preserved three consumer dispositions, and performed zero model or external effects. The license gate remains closed.
Future model candidate boundary
PASSThe acquisition decision rejected three adversarial mutations and recorded zero downloads, model loads, forward or decode steps, provider calls, credentials, spend, or execution authority.
HEATWAKE custody receipt
PASSThe immutable record hash and file digest matched. The receipt remained explicit at 1/7 clock days, coverage unaudited and unpassed, 18 gaps, 18 reconnects, no sequence-completeness claim, and no terminal receipt.
HEATWAKE focused controls
PASSAll 39 lifecycle and supervisor tests passed from the unchanged frozen checkout with no source write or external effect.
Machine-readable evidence
Exact commits, metrics, record hashes, verification commands, results, and claim boundaries.
Download JSONRelease checksum
A SHA-256 digest for checking that the published evidence ledger has not changed.
Download SHA256SUMSCitation metadata
CFF metadata binding the title, date, author, publisher, and canonical identity.
Download CITATION.cff