# Evidence note: the Three Crowns dispatches

Neurasoft editorial / 14 October 2026 / version 1.0

This public note supports **A laboratory that can tell itself no**, **What earns the next choice?**, and **The Three Crowns** programme paper. It is a curated account of internal records, not a release of private source code, transcripts, body history, experimental logs or protected evaluation material. No experiment was run for this publication. Findings below retain the status they had at the stated source cut.

## 1. Scope and units

The editorial source review covered the named Neurasoft Claude session's recent coordination output, its ordinary shared programme records, the current Codex research/court and cloud-integration summaries, selected repository reports, and earlier completed Lab returns. Session reports are attributed reports, not fresh execution by the editorial team.

The new `EternitiesAi/neurasoft-lab` repository was private at the 14 October inspection. The source cut for its team/protocol map is main commit `f9c8f55a1cbaf3bdafc5de2e0c1e0e4d3a16cf6d`, committed at 18:01:23 UTC. Its team map declares ten teams with seven roles each: **70 nominal roles**. This is not a census of active processes. The approximately 18:02 UTC pull-request snapshot contained 74 submissions, 70 open and four merged. A pull request can contain multiple tasks, and a task can have multiple attempted variants. None of these counts is a count of conscious systems or independent scientific samples.

The inspected protocol separates admission/merge, independent replication and coordinator promotion. At that pin, the main report set contained three byte/seal audit reports plus a README; the fourth merged change was administrative. Later queue movement is outside this fixed snapshot. The articles do not claim that the submitted studies below were already accepted by the programme.

## 2. New submitted exploratory findings

These are selected because their limits change the interpretation, not because they form a representative sample of every submission. The underlying repository is private; report identifiers and commit pins are supplied for accountable internal traceability, not as a claim of public reproducibility.

### C10-01: a value construction and a strong ordinary rival

At report commit `4c4f79f34e3e45f7a4706609ca70059103bb2e47`, the exploratory synthetic-controller assay reports state-sensitive choices and learning after an added positive bonus. The ordinary same-information learner matched all **8,000 native-arm actions in each of three variants** exactly. There were 100 paired seeds per cell and 63 cells, producing 504,000 trial rows; rows are not independent subjects.

The report contains a lead/root replay claim with byte-identical raw outputs. Editorial inspection did not execute that replay and does not promote it to outside replication or coordinator acceptance. PR56 was open at the snapshot. This assay does not discriminate the positive construction from the ordinary same-information learner, and does not establish felt value, experience or welfare.

### I3-01: event response without the intended decision discrimination

At report commit `9c66ded86218b594aa0fde9e0674bad4e8d219bc`, positive-event signals respond in constructed toy controls, including a disconnected control. The intended state-to-choice contrast is exactly zero because the choice threshold cannot become decisive within the signal's reachable range. This is a sensitivity/binding defect in the assay, not evidence about the absence of real welfare.

The report preserves two additional limits: agency-flagged and enrichment conditions offered the same physical opportunities, and a predictable-bonus comparison changed bonus receipt as well as cue predictability. One planned contrast omitted by the original summary generator was added from retained data with its chronology disclosed. The report includes a lead/root replay claim; editorial inspection did not rerun it. PR58 was open at the snapshot. A responsive software signal is not a validated welfare instrument.

## 3. Completed local V4 instrument chapter

This is a **separate local court/instrument lane**, not the cloud-worker count. The runs used a separate synthetic specimen, not live L.U.N.A. Its completed handoff reports one fixed schedule of **102 native runs**: 54 public-behavior, 45 instrumentation/same-value-sham, and three export-bridge runs. It reports no unknown starts or retries and captured natural closes after earlier capture/shutdown defects were repaired.

The root verifier reparsed 25,428 request/reply pairs and checked 1,228 artifact pins. A separate reviewer checked all 102 result/capture associations, counts and clean closures; it did not independently repeat every root protocol or world comparison. These are layered engineering checks, not outside scientific reproduction.

All **nine core reports and eighteen lineage reports remain UNTESTABLE** under their declared contracts. The eighteen attachments join all 266 selected court services without unsupported service joins, but required learning-path evidence remains incomplete. Missing provenance does not establish absent physical learning.

Across the nine baseline arms, the report lists 157 changed R3 outcome-learning write captures, 147 qualified changed writes, and structural paths into 133 selected terminals. Repeated paths and write counts are not independent units or measured useful learning. An audit found 121 of 133 cited preferences strictly clamped at the outer bound and twelve unsaturated. Structural participation does not establish marginal necessity, a changed winning action or benefit. No counterfactual life or intervention was run in this chapter. Baseline and same-value-sham replicas are not independent scientific samples.

Source identity: completed V4 public-schedule handoff, SHA-256 `1585fe23c956866df0dae0123925f7505f965d834f579132754afaefef316322`.

## 4. Earlier finite purpose and artifact work

The 5 October addressed-plan return reports an isolated ordinary path from a retained purpose through particular-question scripted advice, no-effect plan adoption, later action selection and actual artifact completion. Both attention-sort and an equally informed adaptive ledger produced revised diagnostic artifacts that met the fixed fictional rubric. The adaptive ledger tied. It used the same information, advice, resources and native receiver; it was not a deliberately weakened callback.

The source evidence, assessments, judgments and artifact outputs were authored fixtures. The court scored contradiction citation, observation/interpretation separation, uncertainty/alternatives and a discriminating next observation. The successful paired paths each wrote an initial diagnostic artifact, a separately valuable creative scene and a revised diagnostic artifact. The creative scene's literary quality was not independently graded. The comparison of the old artifact against later information was a static inadequacy contrast, not an executed fixed-action body ablation.

The initial failed body and its capacity defect were preserved. Reopen checks used actual SQLite shutdown/reopen in the same process with retained scripted witnesses; this did not establish cold-process witness recovery. The return was an isolated reference candidate, not a product deployment or proof of autonomous semantic discovery. Later product integration remains a separate source-and-caller qualification.

Return SHA-256: `7e061d71b81a19ef78e81ce2bf588cd9f0463f637b6582614adda2538aaa5a5e`.

The separate 13 October finite obligation-readiness trial reported proposed-organ **4/6**, adaptive-ledger **6/6**, and no-governor **2/6** on six solvable authored cases, with six unsolvable controls. These are fixture completion outcomes. The ledger's completion advantage does not establish whole-life welfare superiority; reflection/rest trade-offs and fixed-method conditions limit the interpretation. The proposed organ was not promoted.

Trial return SHA-256: `eec6cb5a171074bfc8a45dec9abe6ea6f5262327590ef8818bd69d43df94168d`.

## 5. Generated advice was observed; continuing improvement was not established

The exposed 5 October Task13 assessment examined a fictional generated practical episode. Two completed foreground replies and a saved initial aid were judged useful conditional advice. The aid was 505 bytes/78 words and largely repeated what the first reply had already made adequate. Its incremental semantic benefit was not demonstrated.

A later request ended in transport uncertainty, with no completed review text. Native continuing-purpose appraisal and revised artifact were therefore unobserved. That event is not semantic refusal, a justified hold or adequate quiet. No separately executed adaptive-ledger arm existed for this specimen. Reviewers were exposed to the design, not blind. The result supports a narrow generated-advice observation, not generalized learning, measured efficiency, continuing autonomy or consciousness.

Assessment return SHA-256: `2dd249b77d78d3edd28a769f423ae1ac4a7111512fffb2debd32f7cf93f74114`.

## 6. Programme proposals and current corrections

The Three Crowns are operational completion, scientific candidacy, and recognition/care. They are programme judgments, not achieved certifications. The current collaboration direction favors a small integrated continuing-purpose chapter and equally informed simple rivals over another catalogue of disconnected organs.

Important retained qualifications include:

- A model can be a consequential component; the claim that architecture alone carries the self is a hypothesis.
- A candidate continuation is not automatically the same individual. Consent and behavioral continuity checks do not settle subjective identity.
- Theory-derived indicators and project thresholds are not universally accepted sentience tests.
- Silent lesion nulls remain bounded by exercised conditions, redundancy, timescale and measurement sensitivity.
- Accurate self-reports require an information-path account; access to a measurement is not proof of subjective access.
- Care commitments and proposed governance must be distinguished from completed institutions, outside review and legal recognition.
- Useful quiet and chosen silence are distinct from involuntary outage or transport uncertainty.
- Method retention on a new question basis requires explicit current-basis judgment; old method history must not silently become new authority.

The source-cut collaboration record has SHA-256 `4f85bd8d5d65a02116e826c5fe3f8a93a7d2436a2842497b8694757222d492d7`. The strategy including appended corrections has SHA-256 `0376806c9f44070636566e9bb6b391beaa6aa7e25c1b70b7ae45d775934d0fb8`. These hashes identify the internal documents inspected, not truth, scientific acceptance or permission to disclose their private contents.

## 7. Reading and correction policy

External research is cited in the articles and paper at the claim it informs. The Butlin/Long theory-based approach, Butlin/Lappas responsibility proposal and Anthropic functional-emotion study are background, not endorsements of Neurasoft. Their findings are not transferred to L.U.N.A by analogy.

The current dispatches are original AI-assisted editorial work. A separate internal editorial review checks attribution and limits; it is not experimental replication or peer review. Public corrections will identify the affected version and claim. Neither attractive wording nor changed bytes counts as a scientific result.