Jump to a findings section
Exact behavior held; neither adaptive selection nor incomplete custody advanced the claim.
Across four held-out blocks, native-restored and equally informed cached-prefix representations returned 128/128 exact paired decisions. The native-restored representation had lower measured logical wall time at every one of 24 block-horizons, so the frozen selector chose it every time, matched the always-native policy exactly, improved on the best static policy by 0%, and beat both static policies in 0/4 blocks. The pilot therefore retained its null hypothesis: no adaptive selector utility was demonstrated. The exact frozen tree passed 13 focused checks plus a formal replay of 70 implementation pins and 29 artifact pins. HEATWAKE separately recorded that a host reboot interrupted its raw collection after three of seven required full UTC days. Fifty closed receipts were retained, the unreceipted final fragment was excluded, and a fresh successor contract preserved the same admission thresholds before any committed launch evidence. A 123,520-byte three-file public sample passed 7 focused intake and custody checks, but is quarantined for offline adapter smoke only. No source admission, model fit, final-test opening, signal, edge, profitability, order, trade, or production result followed.
Program record
Waggle + Kea
Verified in scope: 128/128 exact paired decisions, information-necessity controls, Kea qualification, restricted continuation, and exact frozen-tree replay. Rejected: adaptive selector utility on this pilot. Not established: general timing, overall efficiency, cross-model transfer, unique semantics, production value, or authority.
Open programProgram record
HEATWAKE
Verified: the prior raw collection reached three of seven full UTC days before a host reboot, 50 closed receipts remain retained, and the unreceipted fragment is excluded. A small public sample passed bounded adapter-smoke controls. Not supported: complete custody, source admission, model eligibility, a forecast, signal, edge, profitability, trading, or production operation.
Open programAtomic claim deltas.
Swipe horizontally to compare all four columns.
| Claim | Prior status | New evidence | Current status |
|---|---|---|---|
| A task-sufficient selector can outperform always using one representation. | Untested. Earlier work established exact paired behavior but did not test adaptive representation selection. | Four held-out blocks and six horizons per block produced 128/128 exact paired decisions. The native-restored representation was lower at all 24 block-horizons, so the selector chose it throughout, matched always-native, and beat both static policies in 0/4 blocks. | Rejected for this feasibility pilot; no adaptive utility demonstrated. |
| Task-sufficient native and readable representations preserve exact downstream decisions. | Exact parity supported on earlier bounded paired fixtures. | The new source-separated workload returned 128/128 exact paired decisions. Thirty-two replacement and 32 no-state controls each produced zero target passes, and the restricted consumer continued only from the qualified projection. | Supported on this one frozen workload, model, runtime, and host. |
| HEATWAKE can complete the existing seven-day raw collection window. | Two of seven full UTC days recorded at the prior daily cutoff; collection incomplete. | A host reboot interrupted the window after three full UTC days and 50 closed receipts. The unreceipted final fragment was excluded, and the fixed window became impossible to complete. | Rejected for that window; a fresh preregistered window is required. |
| A pinned public sample can support bounded adapter development without becoming research evidence. | No alternate public sample had passed the bounded intake controls. | A 123,520-byte, three-file raw sample passed 7 focused contract, hash, schema, size, create-once, and refusal checks. It contains no derived features, labels, resolved outcomes, or identity-bearing tables. | Accepted for offline parser, event-time, availability-time, gap, and duplicate smoke only; not admitted for training, evaluation, labels, or final testing. |
Claims, controls, receipts, reruns.
Frozen information cutoff
The review fixed one committed Waggle/Kea snapshot and one committed HEATWAKE snapshot at July 31, 2026 at 09:08:25 EDT. Uncommitted, active, lock-held, and later source-writer bytes were excluded.
Task-sufficient paired comparison
Four held-out blocks used the same pinned model and runtime, equal source information, fixed expected decisions, and horizons of 1, 2, 4, 8, 16, and 32. The selector could observe only frozen pre-execution features.
Preregistered selector rule
Support required the selector to beat both always-native and always-readable policies in aggregate and in at least three of four held-out blocks while retaining exact quality and at most 5% oracle regret. Uniform native dominance satisfied the regret gate but falsified adaptive utility.
Negative and information controls
Thirty-two causal-replacement and 32 no-state control decisions produced zero target passes. Six failed frozen trials were preserved rather than rewritten; the accepted seventh trial used new sources after a separate recovery gate.
Frozen-tree replay
A clean archive of the exact commit passed 13 focused checks, verified 70 implementation and 29 artifact pins, recomputed Kea qualification and consumer output, and used zero new model, provider, network, or authority effects.
HEATWAKE fail-closed custody
A lifecycle interruption cannot be repaired by relabeling an unreceipted fragment. The prior session remains retained but ineligible; the successor contract keeps the seven-day threshold and science gates unchanged and requires a fresh window.
Bounded public-sample intake
Only three hash-pinned raw tables totaling 123,520 bytes were accepted. Oversized acquisition, clockless promotion, derived features, labels, resolved outcomes, identity-bearing data, training, evaluation, and source admission were refused.
Waggle + Kea
Exact task-sufficient handoff with no adaptive selector value
Verdict
Verified in scope: 128/128 exact paired decisions, information-necessity controls, Kea qualification, restricted continuation, and exact frozen-tree replay. Rejected: adaptive selector utility on this pilot. Not established: general timing, overall efficiency, cross-model transfer, unique semantics, production value, or authority.
The selector did not fail because it made bad choices; it failed because the tested environment never presented a choice. The native-restored representation had lower measured logical wall time at every horizon, so the selector reproduced the always-native policy exactly. That preserves a narrow exactness result while rejecting the stronger claim that pre-execution features added adaptive value.
Exact task behavior survived both representations
128/128 exact paired decisions
Native-restored and equally informed readable-prefix paths returned the expected held-out decisions across four blocks, and Kea qualified the restricted handoff.
Boundary: This is one source-separated numerical coordination workload on one pinned local model, runtime, and host. It does not establish unique semantics, broader reasoning, or production reliability.
The selector collapsed to one static policy
24/24 horizons native · 0/4 blocks improved
The native-restored representation had lower descriptive logical wall time at every tested horizon. The selector therefore chose it every time, matched always-native, and recorded zero uplift over the best static policy.
Boundary: Uniform dominance makes the 10,000-resample uncertainty interval degenerate; it does not supply heterogeneous regimes, confirmatory inference, or independent replication.
The information controls remained necessary
0/32 replacement · 0/32 no-state target passes
Neither causally replaced information nor an absent handoff produced a target pass, while the qualified projection continued.
Boundary: These controls support information necessity only for the frozen task. They do not establish a minimum language, hidden-state semantics, compression, or cross-task transfer.
Replay checked the result without another model run
13 checks · 70 implementation pins · 29 artifact pins
The exact frozen tree rehashed sources, recomputed qualification and consumer output, and preserved no-state abstention with zero model or external calls.
Boundary: Local replay verifies internal consistency, not independent validation or peer review. Six earlier failed trials remain part of the record.
Not demonstrated
- No adaptive selector advantage outside this feasibility pilot and no heterogeneous-regime crossover.
- No general speed, token, Credit, provider-cost, memory, storage, network, energy, economic, or overall-efficiency claim.
- No learned language, hidden-state semantic transfer, minimum sufficient language, unique capability, or higher output quality.
- No cross-model, cross-runtime, cross-hardware, production, customer, deployment, or authority result.
- No independent replication, confirmatory inference, or causal explanation for the measured timing differences.
Next legitimate gates
- 1. Run the already-preregistered heterogeneous-regime crossover only on new frozen sources, requiring both representations to win at least one regime and the selector to generalize out of sample.
- 2. Stop this selector line if a second frozen study again produces uniform static dominance.
- 3. Any efficiency claim additionally requires matched memory, storage, energy, and end-to-end cost accounting against the strongest readable control.
HEATWAKE
Interrupted custody retained; every downstream gate remains closed
Verdict
Verified: the prior raw collection reached three of seven full UTC days before a host reboot, 50 closed receipts remain retained, and the unreceipted fragment is excluded. A small public sample passed bounded adapter-smoke controls. Not supported: complete custody, source admission, model eligibility, a forecast, signal, edge, profitability, trading, or production operation.
Elapsed time did not become admissible evidence. The reboot made the fixed collection window impossible to complete, so the only valid recovery is a fresh preregistered window with unchanged thresholds. The public sample improves parser and clock tooling only; it cannot replace prospective custody or enter any scientific stage.
The interrupted window failed closed
3/7 full days · 50 closed receipts
The retained session records three full UTC days and 50 closed receipts before a host reboot interrupted the non-restartable collector.
Boundary: The unreceipted final fragment is excluded. The session has no terminal receipt, cannot complete its fixed window, and is not eligible for source admission or model input.
Recovery preserved the original thresholds
7-day requirement unchanged · 0 rows copied
A fresh successor contract retained the existing coverage and science gates, required a new custody root and next-midnight window, and copied no rows from the interrupted session.
Boundary: The frozen cutoff contains no committed successor launch receipt or coverage result. A contract is not an active or completed collection.
A public sample qualified only for adapter smoke
123,520 bytes · 3 files · 7 checks
Hash, schema, size, clock, create-once, oversized-repository, and clockless-promotion controls passed for a deliberately small raw sample.
Boundary: The sample is non-representative and quarantined. It has no training, evaluation, calibration, label, final-test, source-admission, edge, profitability, or trading use.
Every science and execution gate stayed closed
0 admissions · 0 fits · 0 final opens · 0 orders
The reviewed snapshot records no admitted source, model fit, sealed-test access, signal, position, order, trade, capital, or spend effect.
Boundary: These exact zero effects do not establish broad security, market value, or production reliability. The current policy action remains ABSTAIN.
Not demonstrated
- No complete seven-day collection, terminal receipt, passing coverage audit, admitted raw source, or model-input eligibility.
- No feature or label construction, calibration, model fit, candidate selection, or sealed final-test result.
- No forecast, signal, market edge, profitability, execution result, trading readiness, or production reliability.
- The public sample is non-representative and accepted only for offline parser, clock, gap, and duplicate smoke.
- No raw market payload, identity-bearing table, private receipt, reconstruction detail, credential, order, trade, capital, or spend data is published.
Next legitimate gates
- 1. Verify a committed successor launch receipt before describing a fresh collection as active.
- 2. Require seven new qualifying full UTC days, exact terminal custody, and the separately frozen source-admission audit before any data becomes model eligible.
- 3. Run the bounded public-sample adapter smoke without emitting a training, evaluation, label, or performance artifact.
Falsifiers and revision conditions.
- 1Revise the selector conclusion only if a new preregistered heterogeneous study contains at least one frozen regime favoring each representation and the selector beats both static policies out of sample.
- 2Revise exact handoff scope only after a different model, runtime, host, and task family reproduces the same qualification and refusal behavior.
- 3Revise HEATWAKE custody only after a new fixed window reaches seven qualifying full UTC days with exact terminal custody and a passing separate source-admission audit.
- 4Revise the public-sample boundary only after rights, causal clocks, representativeness, gap handling, and a separately preregistered scientific role all pass without opening the sealed final test.
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 selector result is a feasibility pilot on one local model, runtime, host, and deterministic numerical workload. The 10,000-resample interval is descriptive and degenerate because the selector never departed from always-native.
- 3The observed 10.813% aggregate native timing difference is not a general timing or efficiency claim. Native retained state was materially larger than the readable prefix, and direct energy was unavailable.
- 4Six failed frozen selector trials are retained. The accepted seventh trial used new sources only after a separately recorded recovery gate; failed outputs were not rewritten into the accepted result.
- 5The interrupted HEATWAKE session has no terminal receipt. Its unreceipted final fragment is excluded, and its three elapsed full days are not source admission or scientific evidence.
- 6The public dataset sample is deliberately tiny and non-representative. Publisher metadata and a declared license do not independently establish custody, causal fitness, or permission for scientific use.
- 7The citation record is first-party attribution metadata. No DOI, archive deposit, external repository publication, peer review, or independent endorsement exists.
- 8This verification made no paid request, provider call, new model run, source-program write, final-test open, order, trade, capital change, credential change, or authority grant.
Rechecked from the source trees.
Waggle + Kea focused controls
PASSThirteen checks covered exact H1 qualification on a synthetic crossover, tied-cost refusal, no-state abstention, information leaks, content-ID forgery, negative zero, authority smuggling, projection tamper, and a fresh restricted-consumer CLI.
Waggle + Kea formal replay
PASSThe exact frozen commit verified 70 implementation pins and 29 accepted artifacts, rehashed sources, recomputed Kea qualification and consumer output, and used zero model loads, decode steps, provider calls, external calls, or authority effects.
HEATWAKE interruption controls
PASSThe focused suite verified the interruption record, excluded fragment, retained closed receipts, unchanged thresholds, fresh-start law, source bindings, secret-free collection path, and zero model or execution effects.
Public-sample intake controls
PASSContract, schema, hash, size, create-once, oversized-repository, clockless-promotion, metadata-only catalog, and zero-science-effect checks passed; the sample remains quarantined.
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