Skip to content
Research library
Findings update 005 July 27, 2026 Locally source-verified

Native reuse beat repeated full text, not cached text

Waggle + Kea now supports a bounded six-branch native-state reuse and independently qualified handoff mechanism, but the strongest matched controls still reject an overall efficiency claim. HEATWAKE now has a frozen successor admission audit and a reliable read-only operator view, but no completed collection, admitted source, model result, signal, edge, or trading result.

Waggle + Kea

6/6 · five informed arms

Every informed arm returned exact outcomes; Kea qualified all six handoffs and the no-state consumer abstained.

Strongest timing result

20.285 s vs 12.850 s

Resident native beat repeated full text but remained slower than the cached-prefix text control.

HEATWAKE

Audit frozen · 0/7 days

The successor admission method is fixed, but collection is incomplete and no downstream claim advanced.

Review status
Source-verified local research update; not peer reviewed or independently validated
Author
William Keenan
Publisher
K&E Studios Research
Information cutoff
July 27, 2026 at 13:36:01 EDT
Evidence scope
Frozen committed snapshots, matched controls, immutable aggregate receipts, zero-model replays, and local release verification
Jump to a findings section
Executive finding

The mechanism advanced; the strongest efficiency claim did not.

One frozen local model state was restored once into a resident service and used for six source-separated branches. Resident native, cached-prefix text, repeated resident full text, fresh native, and fresh full text each returned 6/6 exact outcomes; Kea qualified all six outputs, a projection-only consumer continued, and the no-state consumer abstained. Resident native became faster than repeated full-text reconstruction at branch three and finished in 20.285 seconds versus 36.141 seconds, but it remained slower than cached-prefix text at 12.850 seconds and fresh native at 17.881 seconds. This supports the mechanism and one narrow break-even, not overall efficiency. Separately, HEATWAKE froze the exact V4 source-admission audit before the collection outcome and corrected a read-only observer that had mistaken two legal reconnect states for identity drift. The accepted observation still reported 0/7 complete UTC days, raw source not admitted, and every model, signal, edge, profitability, trading, and production claim false.

What changed

Atomic claim deltas.

Swipe horizontally to compare all four columns.

ClaimPrior statusNew evidenceCurrent status
One native-state service can support multiple source-separated tasks through Kea.Supported across three tasks in a resident context service.Five informed arms each returned 6/6 exact outcomes; one resident native restore served all six branches, Kea qualified 6/6 outputs, and a projection-only consumer continued.Supported across these six bounded fixtures.
Native reuse can beat repeated full-text reconstruction as branches accumulate.Rejected in the three-task service result.Resident native first became faster at branch three and retained the advantage through branch six: 20.285 s end-to-end versus 36.141 s for repeated resident full text.Supported only against this repeated full-text control after branch three.
Native reuse is the most efficient informed service in this experiment.Rejected.Cached-prefix text completed in 12.850 s and fresh native in 17.881 s, both faster than resident native at 20.285 s.Rejected.
Kea can independently qualify the handoff without becoming an authority source.Supported across the earlier three-task projection.Kea qualified six decisions against a separately pinned trust root; forged ledger lineage was rejected, the restricted consumer continued from the projection, and the no-state control abstained.Supported across these six bounded fixtures.
HEATWAKE has a fixed successor source-admission method before seeing the collection result.A V4-specific adapter was the next required dependency.The V4 audit contract now fixes terminal-receipt, closed-segment, seven-day, coverage, freshness, custody, and refusal checks without changing scientific thresholds or enabling automatic admission.Method frozen; outcome not run and source not admitted.
HEATWAKE's current collector changes its science or trading evidence level.Raw custody only at 0/7 complete days.A read-only operator receipt verified an advancing sole collector and legal reconnect states, but still reported 0/7 complete days, no terminal receipt, and closed downstream gates.No upgrade; raw custody and ABSTAIN.
How this was verified

Claims, controls, receipts, reruns.

1

Frozen information cutoff

The review fixed one committed Waggle/Kea snapshot and one committed HEATWAKE snapshot at July 27, 2026 at 13:36:01 EDT. Uncommitted work and later commits were excluded even when source work continued during review.

2

Matched five-arm comparison

The same six frozen tasks and expected values were evaluated through resident native, resident cached-prefix text, repeated resident full text, fresh native, and fresh full text. A no-state arm tested abstention.

3

Zero-model evidence replay

The Waggle/Kea service, adversarial, formal-replay, prior-checkpoint, typecheck, and build gates were rerun without additional model loads, forward or decode steps, provider calls, external calls, or authority effects.

4

HEATWAKE separation

The review verified the frozen admission contract, preregistration, operator receipts, focused admission tests, service tests, browser contracts, typecheck, lint, and build. It did not read active raw events, run admission, construct features, fit a model, open the final test, or execute a trade.

Program 01

Waggle + Kea

Six-branch native-state reuse with independent qualification

Full program

Verdict

Supported in scope: a bounded six-task native-state reuse result was independently qualified by Kea and handed to a restricted consumer. The native arm beat repeated full-text reconstruction after branch three. Not supported: superiority to cached-prefix text or fresh native, overall efficiency, learned language, semantic advantage, production behavior, deployment, or authority.

The positive result became narrower and more useful. Native reuse avoided rebuilding the full readable prefix for each branch and crossed a repeat-full-text break-even, but a properly engineered cached-prefix text service remained the fastest informed arm. The result therefore establishes a bounded mechanism and comparative boundary rather than a general efficiency advantage.

Finding 1.1 Supported in scope

One restore served six independently qualified branches

6 tasks · 5/5 informed arms at 6/6 exact

The resident native service restored one frozen state into one context and served six source-separated suffixes. Kea qualified all six decisions, the restricted consumer continued from the projection alone, and the no-state consumer abstained.

Boundary: Exact parity across every informed arm establishes neither unique native semantics nor higher output quality. The result is limited to one local model, one host, six frozen fixtures, and one accepted run.

Finding 1.2 Supported in scope

Native reuse crossed one narrow break-even

Branch 3 break-even · 20.285 s vs 36.141 s at branch 6

Resident native first became faster than repeated resident full-text reconstruction at branch three and preserved that advantage through the sixth branch.

Boundary: This comparison is against rebuilding the full readable prefix in a new context for every branch. It is not the strongest informed baseline and does not establish overall efficiency.

Finding 1.3 Negative result

The strongest controls remained faster

12.850 s cached text · 17.881 s fresh native · 20.285 s resident native

Cached-prefix text reused one readable prefix and remained the fastest informed service. Six cache-warmed fresh native processes also completed before resident native.

Boundary: Process order was frozen, but operating-system cache and thermal state were not controlled. The first native model load was much slower than later loads, so no causal explanation or adjusted timing claim is made.

Finding 1.4 Boundary finding

The audit boundary resisted forged lineage

13/13 attacks rejected · no-state ABSTAIN

Kea rejected forged lineage against a separately pinned trust root. The restricted consumer received only the approved projection, while the no-state control abstained.

Boundary: Kea verifies the bounded handoff and can refuse it; Kea does not grant execution authority, prove source truth beyond the frozen contract, or turn the result into a customer or production claim.

Not demonstrated

  • No learned language, neuralese, hidden-state semantic transfer, unique capability, or semantic or quality advantage.
  • No superiority to cached-prefix text or fresh native, and no token, Credit, provider-cost, memory, storage, energy, or overall-efficiency savings.
  • Startup order, operating-system cache, and thermal state were not controlled; direct energy was unavailable.
  • No customer result, production reliability, deployment, external action, or authority execution.

Next legitimate gates

  1. 1. Freeze process order and cache-state controls before measurement, then repeat the branch curve.
  2. 2. Extend the preregistered branch horizon to test whether native reuse ever crosses the cached-prefix text control.
  3. 3. Keep Kea qualification, separate trust roots, forged-lineage refusal, and no-state abstention mandatory.
Program 02

HEATWAKE

Frozen source-admission audit with incomplete raw custody

Full program

Verdict

Supported: the successor source-admission method was frozen before outcome access, the read-only operator remained available across legal reconnect states, and every reviewed effect stayed within its zero-authority boundary. Not supported: completed coverage, source admission, model eligibility, a forecast, signal, edge, profit, trading readiness, or production operation.

The scientific result did not advance. The methodological dependency did: the exact audit that may inspect the completed successor session is now frozen and cannot silently weaken coverage or custody rules after seeing the outcome. A separate operator repair corrected a false 503 during legal reconnect states without changing collection bytes, thresholds, or handoff gates.

Finding 2.1 Supported in scope

The V4 admission audit was frozen before outcome

11/11 admission checks · automatic admission disabled

The contract fixes terminal-session, exact closed-receipt pairing, seven complete UTC days, daily coverage, timestamp freshness, duplicate and crossed-book refusal, continuity, and unchanged science thresholds.

Boundary: The preregistration remains WAIT. The audit has not run on a terminal session, no result exists, and a passing future source audit would establish only raw-source eligibility for a later separate stage.

Finding 2.2 Supported in scope

The read-only operator stopped misclassifying legal reconnects

97 service tests · 117 frontend tests · POST 405

The observer now accepts two contract-defined ABSTAIN transitions: no active segment during reconnect backoff and an unseeded new segment awaiting its first snapshot. Malformed identity, hashes, counters, process ownership, effects, and handoff state still refuse.

Boundary: This is a local loopback operator-availability result. It changed no collection evidence, research threshold, scientific conclusion, or trading authority.

Finding 2.3 Boundary finding

Collection remained incomplete

0/7 complete UTC days · 0 protocol refusals

The accepted bounded receipt observed one advancing collector, fifteen closed segments of sixteen started, fourteen retained reconnect gaps, and no protocol refusal while the seven-day window remained at zero complete days.

Boundary: Frame and segment counters are custody observations, not evidence of complete coverage, data quality, source admission, model eligibility, signal quality, operational reliability, or market edge.

Finding 2.4 Boundary finding

Every downstream claim remained closed

0 fits · 0 final-test opens · 0 orders · $0 spend

The frozen contract and accepted receipts reported no feature or label writes, calibration reads, model fits, final-test access, signals, positions, orders, capital, or spend.

Boundary: These are exact effects for the reviewed snapshots, not a broad production-security guarantee. The current decision remains ABSTAIN.

Not demonstrated

  • No completed seven-day collection, terminal session receipt, source-admission result, admitted model corpus, or final-test result.
  • No forecast, calibrated direction, signal, market edge, profitability, execution result, or trading readiness.
  • No production reliability claim; the operator result is local loopback only and the collector has retained reconnect gaps.
  • No raw payload, private receipt, reconstruction detail, credential, order, trade, capital, spend, or release authority is published.

Next legitimate gates

  1. 1. Preserve the sole collector through seven complete UTC days and its exact terminal receipt.
  2. 2. Run the already-frozen source-admission audit without reading active or unreceipted files and without changing any threshold.
  3. 3. Only a passing source audit may open a separate, still-gated feature and model-input decision; it cannot open the final test or trading.
Cross-program finding

Cross-program conclusions.

The strongest baseline changed the conclusion

Native reuse beat repeated full-text reconstruction, but cached-prefix text remained faster. Reporting both prevents a real mechanism from being misreported as a general efficiency win.

The admission rule was frozen before the outcome

HEATWAKE's V4 audit now fixes the exact terminal, custody, coverage, freshness, and refusal rules while the collection is still incomplete.

Operator availability was separated from scientific evidence

A false 503 was corrected only in the read-only projection. The raw collector, thresholds, receipts, admission state, model gates, and ABSTAIN decision did not change.

Later source activity was excluded

Both programs continued independently. The public review stayed bound to the preregistered committed snapshots and deferred later commits to a future cycle.

Limitations and disclosures

Limits on the reported findings.

  1. 1All verification described here is local source verification by K&E Studios Research, not peer review or independent validation.
  2. 2The accepted Waggle/Kea experiment used one local open-weight model, one workstation, six deterministic fixtures, and one accepted timing run. No statistical confidence interval or direct energy measurement is claimed.
  3. 3Operating-system cache and thermal state were not controlled. The frozen process order produced a large startup-order difference, which is reported rather than subtracted after the result.
  4. 4HEATWAKE uses a public unauthenticated market-data channel under a frozen zero-spend collection contract. Raw payloads, private receipts, market reconstruction detail, and active collector state are not published.
  5. 5This publication replay made no new model loads, forward or decode steps, paid requests, market-capture requests, feature writes, model fits, final-test opens, orders, trades, capital changes, or authority grants.
Fresh local rerun ledger

Rechecked from the source trees.

Waggle + Kea mechanism and refusal controls

PASS

The six-task service suite passed; all 13 adversarial mutations were rejected; strict TypeScript and the production build passed.

Waggle + Kea frozen replay

PASS

The 23-pin formal replay verified six tasks, 15 accepted process loads, and 690 accepted forward/decode steps with zero added model work, calls, or effects. The C15c prior replay also passed.

HEATWAKE source-admission contract

PASS

All 11 V4 admission tests passed in the declared dependency environment. The frozen preregistration remains WAIT and no admission result exists.

HEATWAKE operator receipts and service

PASS

Both write-once operator receipts deep-verified; all 97 operator-service tests passed; the focused V4 browser-contract file passed 2/2.

HEATWAKE frontend

PASS

All 36 frontend files and 117 tests passed, followed by ESLint, strict TypeScript, and the optimized build.

Waggle + Kea now has stronger evidence for a bounded native-state reuse and independently audited handoff mechanism, including a repeat-full-text break-even after three branches. It still does not have evidence for superiority to cached-prefix text, overall efficiency, learned language, or production behavior. HEATWAKE now has its successor admission method frozen and a reliable local read-only operator, but its evidence level remains raw custody at 0/7 complete days with ABSTAIN. The next legitimate results require a cache-controlled longer-horizon Waggle comparison and, separately, a completed HEATWAKE terminal session followed by the already-frozen source audit.

Author and research contact

William Keenan · K&E Studios Research

william@kestudios.dev
Author profile