Skip to content
Research library
Findings update 004 July 26, 2026 Locally source-verified

Resident context worked; native savings did not

Waggle + Kea now has a useful bounded resident context service and audit path, but the measured native path was slower than fresh-native and resident full-text controls; HEATWAKE preserved an interrupted window as ineligible evidence and began a fresh raw-only window with every science and trading gate still closed.

Waggle + Kea service

3/3 · four arms

Resident-native, resident full text, fresh-native, and fresh full text were all exact across the three frozen tasks.

Native timing

24.412 s vs 10.421 s

Resident native was slower than fresh native and also slower than the 20.295-second resident full-text control.

HEATWAKE

V4 · 0/7 days

The interrupted V3 window was retained but rejected for admission; its fresh successor remained raw-only and ABSTAIN.

Review status
Source-verified local research update; not peer reviewed or independently validated
Author
William Keenan
Publisher
K&E Studios Research
Information cutoff
July 26, 2026 at 20:19:06 EDT
Evidence scope
Frozen committed snapshots, allowlisted aggregate receipts, zero-call replays, matched controls, and local release verification
Jump to a findings section
Executive finding

The service claim advanced; the savings claim failed.

Across three disjoint frozen local tasks, resident-native, resident full-text, fresh-native, and fresh full-text arms each returned 3/3 exact results, and a restricted consumer continued only from Kea-qualified projections. The mechanism therefore crossed a narrow service-and-audit gate. It did not cross a performance gate: resident native took 24.412 seconds versus 10.421 seconds for fresh native and 20.295 seconds for resident full text. Separately, a host reboot left HEATWAKE's prior seven-day collection without a terminal receipt; ten closed segments were retained, one unreceipted fragment was excluded, and the fixed window was declared impossible. A fresh raw-only collection then began at 0/7 complete UTC days. Focused successor controls passed 6/6, while one adjacent historical adapter refused a drifted binding. No source admission, model result, signal, edge, profit, or trading claim follows.

What changed

Atomic claim deltas.

Swipe horizontally to compare all four columns.

ClaimPrior statusNew evidenceCurrent status
A resident-native context service can support multiple source-separated local tasks through Kea.One earlier resident deadline fixture passed, but a reusable resident service was not established.All four informed arms returned 3/3 exact results across disjoint tasks; Kea qualified 3/3 projections and the restricted consumer continued from those projections only.Supported in these bounded local fixtures.
Resident-native execution improves wall-time efficiency.Unresolved; the earlier result compared completion at one deadline, not a four-arm service timing study.Resident native took 24.412 s, versus 10.421 s fresh native, 20.295 s resident full text, and 20.913 s fresh full text.Rejected for this experiment.
The resident result implies storage, token, Credit, or overall-efficiency savings.Unsupported.The accepted run retained 279,321,005 context bytes; direct energy was unavailable, and no provider billing or purchased-Credit comparison was performed.Still unsupported.
HEATWAKE's prior fixed seven-day raw-coverage window can support admission.Collection was incomplete; no admission result existed.A host reboot left ten closed receipts and one excluded unreceipted fragment, with no terminal session receipt and too little time to complete the frozen window.Rejected; retained but ineligible and ABSTAIN.
A fresh HEATWAKE collection changes the model or trading evidence level.No successor-window evidence.A new frozen raw-only contract launched one source-bound collector with a seeded snapshot, zero launch gaps or reconnects, and 0/7 complete UTC days.No upgrade; raw custody only.
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 26, 2026 at 20:19:06 EDT. Later working-tree or source-writer activity was excluded.

2

Matched service comparison

Three disjoint tasks were run in resident-native, resident full-text, fresh-native, and fresh full-text arms. Exact task output, Kea qualification, consumer disposition, process loads, decode steps, bytes, memory, and wall time were retained.

3

Zero-call replay

The accepted Waggle/Kea report and its prior checkpoint were replayed from frozen evidence with zero additional model loads, forward/decode steps, provider calls, external calls, or authority effects.

4

Lifecycle refusal

HEATWAKE's interrupted window was evaluated against its original fixed-window and terminal-receipt rules. The successor contract retained the same seven-day minimum and kept raw custody separate from source admission, modeling, signals, and trading.

Program 01

Waggle + Kea

Bounded local context service with independent qualification

Full program

Verdict

Supported in scope: one resident local context service, Kea audit path, and restricted-consumer handoff worked across three frozen tasks. Not supported: a native timing advantage, semantic or quality advantage, learned language, savings, production behavior, deployment, or authority.

The material change is architectural, not economic. A resident runtime could retain a frozen context reference across source-separated tasks, Kea could qualify the outputs, and the consumer could continue from the qualified projection alone. Every matched informed arm was exact, so the experiment found no unique native capability. Its direct timing comparison was negative.

Finding 1.1 Supported in scope

The bounded resident service and audit path worked

3 tasks · 4/4 informed arms at 3/3 exact

Resident-native, resident full-text, fresh-native, and fresh full-text arms all returned the exact expected values. Kea qualified all three tasks, and the restricted consumer continued from the qualified projections only.

Boundary: The no-state arm had no qualified projection and did not continue. This demonstrates a bounded local service and audit mechanism, not unique native semantics, higher quality, general reasoning, production operation, or authority.

Finding 1.2 Negative result

Native wall-time advantage failed

24.412 s resident native · 10.421 s fresh native

The resident-native path was slower than fresh native and slower than resident full text at 20.295 seconds. Fresh full text took 20.913 seconds.

Boundary: These timings apply to one host, one local model, three task fixtures, and this implementation. They reject the experiment's native timing advantage; they do not prove resident services are always slower.

Finding 1.3 Boundary finding

Resource accounting stayed inside the frozen budget

8 accepted loads · 298/1,024 steps · 279,321,005 state bytes

The accepted trial used two resident and six fresh model-process loads. Two interrupted-trial loads were preserved separately, leaving the combined checkpoint at its ten-load ceiling.

Boundary: Peak memory was 2,009,907,200 bytes and direct energy was unavailable. No token, Credit, cost, storage, network, energy, or overall-efficiency savings claim is supported.

Finding 1.4 Boundary finding

Formal replay added no inference

20 pins · 0 added loads · 0 added decode steps

The accepted replay verified three tasks, eight accepted loads, and two preserved interrupted-trial loads without launching new inference. The prior checkpoint also replayed without added loads or steps.

Boundary: This is deterministic local source verification, not independent validation. A temporary archive representation could not satisfy the frozen workspace-inventory check; the accepted replay ran read-only from the clean frozen source.

Not demonstrated

  • No learned language, neuralese, hidden-state semantic transfer, unique capability, or semantic or quality advantage.
  • No token, Credit, provider-cost, storage, network, energy, wall-time, or overall-efficiency savings.
  • One local model, one host, three frozen tasks, and one accepted trial do not establish generalization or production reliability.
  • No customer result, deployment, external action, or authority execution.

Next legitimate gates

  1. 1. Preregister an amortized-throughput comparison with enough tasks per resident load to charge startup cost fairly.
  2. 2. Test lower-overhead context restore against the same full-text and fresh-native controls.
  3. 3. Retain Kea qualification and projection-only consumer access as mandatory safety controls.
Program 02

HEATWAKE

Raw collection lifecycle with admission and trading gates closed

Full program

Verdict

Supported: the interrupted V3 lifecycle was preserved without manufacturing a terminal receipt, and one fresh V4 raw collector launched under a frozen seven-day contract. Not supported: complete coverage, source admission, model eligibility, a forecast, signal, edge, profit, trading readiness, or production operation.

The material finding is a refusal. The original fixed window became impossible after a host reboot and could not be resumed or repaired without violating the frozen contract. Its closed receipts were retained, its unreceipted fragment was excluded, and a new raw-only window began from zero without carrying rows or weakening thresholds.

Finding 2.1 Negative result

The interrupted window was correctly rejected

10 closed receipts · 1 unreceipted fragment excluded

The prior collector had no automatic restart and ended without a terminal session receipt. The remaining calendar could not supply the required seven complete UTC days, so the fixed coverage gate was forced closed.

Boundary: The retained closed segments and excluded fragment establish lifecycle custody, not provider failure, market-data quality, source admission, model input, or an empirical science result.

Finding 2.2 Boundary finding

The successor preserved the original threshold

V4 launched · 0/7 complete UTC days

A fresh contract required a new session, new window, seeded snapshot, zero V3 row carryover, and the unchanged seven-day minimum. Two launch observations showed one source-bound raw collector with zero gaps, reconnects, and protocol refusals at that cutoff.

Boundary: A launch observation is not completed coverage, source admission, operational reliability, or model evidence. The worker had no automatic host restart, and later source activity was outside this review cutoff.

Finding 2.3 Negative result

Focused controls passed; one adjacent adapter refused

6/6 focused controls · 22/23 broader checks

The frozen successor suite passed its contract, collector, custody, authority, and refusal checks. In the adjacent historical set, one V3 source-admission adapter stopped on a drifted launch binding.

Boundary: The refusal was preserved rather than bypassed. It withholds any historical adapter or source-admission result and requires a separately frozen V4 adapter before the seven-day gate can advance.

Finding 2.4 Boundary finding

Every downstream effect remained zero

0 model fits · 0 signals · 0 orders · $0 spend

The reviewed records reported no credentials, authenticated requests, feature or label writes, calibration reads, final-test access, model fits, positions, orders, trades, capital, or spend.

Boundary: These are exact effects for the reviewed records, not a production-security or future-authority guarantee. The current public policy remains ABSTAIN.

Not demonstrated

  • No completed seven-day coverage window, source admission, 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 prior lifecycle was interrupted and the successor had no automatic host restart.
  • No raw event payload, private receipt, market reconstruction, credentials, order, trade, capital, spend, or release authority is published.

Next legitimate gates

  1. 1. Freeze a V4-specific source-admission adapter before the raw collection reaches its seven-day review boundary.
  2. 2. Complete seven full UTC days without weakening coverage, custody, or refusal thresholds.
  3. 3. Audit source admission separately before any feature, label, calibration, model, signal, or final-test surface can read the raw bytes.
Cross-program finding

Cross-program conclusions.

Exactness did not substitute for comparative advantage

All four Waggle/Kea arms were exact. That supported the service mechanism while making the slower native timing—not quality—the decisive comparative result.

Interrupted evidence was not repaired into success

HEATWAKE retained ten closed receipts, excluded the unreceipted fragment, rejected the impossible window, and restarted from zero under unchanged rules.

Producer and reviewer ownership stayed disjoint

The review read only frozen committed snapshots. Ongoing source writers were not paused, steered, or treated as publication owners, and their later bytes were deferred.

The comparison remains falsifiable

A preregistered amortized-throughput result that beats matched fresh-native and full-text controls would revise the Waggle conclusion; a complete, clean seven-day V4 audit could advance HEATWAKE only to source-admission review.

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. This publication replay made zero additional model loads, forward/decode steps, provider calls, external calls, or authority effects.
  3. 3HEATWAKE's raw collection used a public unauthenticated market-data channel under a previously frozen, zero-spend contract. This review made no new capture request and read no later live source bytes.
  4. 4The public evidence ledger exposes aggregate measurements, frozen commit labels, and non-reversible hashes only. Non-public implementation materials and operationally sensitive detail are omitted.
Fresh local rerun ledger

Rechecked from the source trees.

Waggle + Kea service and adversarial controls

PASS

Strict TypeScript passed; the three-task service suite passed; all 11 adversarial inputs were rejected or failed closed.

Waggle + Kea frozen replay and regression

PASS

The accepted 20-pin replay and the prior checkpoint replay passed with zero added model loads, forward/decode steps, provider calls, external calls, or effects. The frozen production build passed.

HEATWAKE V4 focused controls

PASS

All six frozen successor contract, collection, custody, authority, and refusal tests passed from the committed snapshot.

HEATWAKE adjacent historical adapter

REFUSED

The broader adjacent set passed 22 of 23 checks; one historical V3 source-admission adapter refused a drifted launch binding. No adapter or science claim was promoted.

External effects from publication verification

ZERO

No paid call, new model inference, new market capture, source mutation, email, credential change, final-test open, order, trade, capital movement, authority grant, push, merge, or unrelated deployment occurred.

The resident context service is now a supported bounded mechanism, but its native timing advantage and every savings claim are rejected or unsupported. HEATWAKE improved the honesty of its evidence chain by refusing an interrupted window and restarting raw custody from zero; it did not improve its model, signal, edge, profit, or trading evidence level. The next useful work is a fair amortized resident-throughput test and, separately, a frozen V4 admission adapter followed by a complete seven-day coverage audit.

Author and research contact

William Keenan · K&E Studios Research

william@kestudios.dev
Author profile