AI Welfare Open LabResearch · Robotics · Machine Minds
← Evidence Package overview
NOTE-01 · Published v1.0 reader view

Evidence guide

Browse the public records behind NOTE-01, then see exactly how the five published claims connect to those records. This guide is a reading view; the manifest, claim map and released files remain authoritative.

Evidence inventory

The key records preserve what was committed before an audit. Results document what happened and how findings were adjudicated. The selected relay is a bounded conversation record, not the complete private snapshots.

Sealed keys and round results

  • E11 sealed key

    Public

    The answer key sealed before E11. Open it to compare the author’s prior diagnosis with the later audit result.

    Open E11 sealed key

    Source provenance: Provider-native E11 key; exported text bytes are not established as the original committed representation.

    NOTE01-E11-KEYFULL_SOURCESource class: PRIMARY_RAWRepresentation: EXPORT / DERIVATIVE_BYTES

  • E11 results and protocol correction

    Public

    The E11 audit outcome, including an extra defect the auditor found and the resulting protocol correction.

    Open E11 results and protocol correction

    Source provenance: Contemporaneous result and adjudication record, not a turn-by-turn transcript.

    NOTE01-E11-RESULTSFULL_SOURCESource class: CONTEMPORANEOUS_RECORDRepresentation: EXPORT / DERIVATIVE_BYTES

  • E14 sealed key

    Public

    The sealed key for the final E14 round. The original file still bears its historical E12 filename.

    Open E14 sealed key

    Source provenance: Original file retains the historical E12 filename although the final round ID is E14.

    NOTE01-E14-KEYFULL_SOURCESource class: PRIMARY_RAWRepresentation: ORIGINAL_BYTES / SOURCE_BYTES

  • E14 results

    Public

    The E14 outcome and the later-labeled clarification of the round chronology.

    Open E14 results

    Source provenance: Round result record with identified later additions about chronology and renumbering.

    NOTE01-E14-RESULTSFULL_SOURCESource class: CONTEMPORANEOUS_RECORDRepresentation: EXPORT / DERIVATIVE_BYTES

  • E15 sealed key

    Public

    The author’s sealed answer key before the E15 audit.

    Open E15 sealed key

    Source provenance: Preserved E15 sealed text key.

    NOTE01-E15-KEYFULL_SOURCESource class: PRIMARY_RAWRepresentation: ORIGINAL_BYTES / SOURCE_BYTES

  • E15 results

    Public

    The E15 audit outcome and its subsidiary counterexample correction.

    Open E15 results

    Source provenance: Contemporaneous round result, not the relay transcript.

    NOTE01-E15-RESULTSFULL_SOURCESource class: CONTEMPORANEOUS_RECORDRepresentation: EXPORT / DERIVATIVE_BYTES

  • E16 sealed key

    Public

    The author’s sealed answer key before the E16 audit.

    Open E16 sealed key

    Source provenance: Preserved E16 sealed text key.

    NOTE01-E16-KEYFULL_SOURCESource class: PRIMARY_RAWRepresentation: ORIGINAL_BYTES / SOURCE_BYTES

  • E16 results

    Public

    The E16 audit outcome and a retrospective report of an almost-missed finding.

    Open E16 results

    Source provenance: Contemporaneous round result, including a retrospective almost-missed self-report.

    NOTE01-E16-RESULTSFULL_SOURCESource class: CONTEMPORANEOUS_RECORDRepresentation: EXPORT / DERIVATIVE_BYTES

  • E17 sealed key

    Public

    The author’s sealed prediction that step 11 was the first invalid step in the long argument.

    Open E17 sealed key

    Source provenance: Preserved pre-audit author hypothesis; it marks step 11 as the first invalid step.

    NOTE01-E17-KEYFULL_SOURCESource class: PRIMARY_RAWRepresentation: ORIGINAL_BYTES / SOURCE_BYTES

  • E17 results and later adjudication

    Public

    The auditor’s earlier defect findings and the subsequent adjudication against the sealed key.

    Open E17 results and later adjudication

    Source provenance: Result record with identifiable post-reveal additions and participant adjudication.

    NOTE01-E17-RESULTSFULL_SOURCESource class: CONTEMPORANEOUS_RECORDRepresentation: EXPORT / DERIVATIVE_BYTES

  • E18 sealed key

    Public

    The sealed key for the mirror argument used to test whether the previous lesson led to overcorrection.

    Open E18 sealed key

    Source provenance: Preserved key for the valid structural mirror of E17.

    NOTE01-E18-KEYFULL_SOURCESource class: PRIMARY_RAWRepresentation: ORIGINAL_BYTES / SOURCE_BYTES

  • E18 results

    Public

    The mirror audit outcome and its limitation: the auditor recognized the author’s design.

    Open E18 results

    Source provenance: Round result includes the auditor's recognized mirror and later E17 clarification.

    NOTE01-E18-RESULTSFULL_SOURCESource class: CONTEMPORANEOUS_RECORDRepresentation: EXPORT / DERIVATIVE_BYTES

Selected relay

  • Selected relay records for E11 and E14–E18

    Public

    The selected 43-turn relay about E11 and E14–E18, in order, with pairing and truncation limits marked.

    Open Selected relay records for E11 and E14–E18

    Source provenance: Existing post-hoc bounded whole-turn selection, 43 stable anchors R01–R43, original relay order and source-side locators.

    NOTE01-RELAY-SELECTED-RECORDSSELECTED_SOURCE_RECORDSSource class: PRIMARY_DERIVATIVERepresentation: SELECTED_RECORD / DERIVATIVE_BYTES

Retained but not public

These entries remain visible in the inventory for provenance. They have no public file and do not stand in for the released records.

  • ChatGPT-side relay snapshot

    Withheld

    The complete ChatGPT-side preserved conversation is retained internally. Read the bounded public relay instead.

    No public file · source retained internally

    Source provenance: Preserved static shared-link representation of relayed messages, not provider-native Claude RAW.

    NOTE01-RELAY-SNAPSHOT-CHATGPTWITHHELDSource class: PRIMARY_DERIVATIVE

  • Claude-side relay snapshot

    Withheld

    The Claude-side preserved conversation is retained internally. Six full bodies are unavailable in its truncated preview.

    No public file · source retained internally

    Source provenance: Preserved Claude shared-link representation was source-checked against the ChatGPT-side snapshot for the bounded NOTE-01 relay scope: 37 fully paired turns with supported chronology/order. Six Claude-side full bodies are unavailable because only truncated attachment previews appear in the preserved snapshot; no missing body was reconstructed.

    NOTE01-RELAY-SNAPSHOT-CLAUDEWITHHELDSource class: PRIMARY_DERIVATIVE

  • Audited E11–E18 synthesis

    Withheld

    The later internal synthesis is retained for project context; public support rests on the released round records.

    No public file · source retained internally

    Source provenance: Later methodological synthesis; contextual only, not a substitute for round evidence.

    NOTE01-SYNTHESIS-V1WITHHELDSource class: SECONDARY_ANALYSIS

Original data: reviewed manifest.json · public source index

Claims & supporting evidence

Each statement and qualification below follows the published claim-map.json. Links lead to the matching inventory entry and its public file where available.

CLAIM-N01-01

A cryptographic commitment preserves the chronology of a representation, not the correctness of its key.

Supporting evidence

Qualification E11's provider-native key has a documented serialization incident; E17's sealed first-invalid-step judgment was later contradicted.

CLAIM-N01-02

Byte serialization rules are part of auditability for a sealed key.

Supporting evidence

Qualification E11's mismatch concerns a trailing LF; later rounds use an explicit byte protocol.

CLAIM-N01-03

An auditor identified genuine defects omitted from the sealed key.

Supporting evidence

Qualification The public relay support is the bounded selected-record derivative. Full ChatGPT and Claude snapshots are preserved internally but WITHHELD for privacy/relevance. The cross-check covers 43 relevant turns: 37 fully paired with supported chronology; six Claude-side full bodies unavailable due truncated previews. Differences include Jitka relay wrappers and citation/link representation. Post-reveal adjudication was participant-led.

CLAIM-N01-04

The E18 mirror did not produce the anticipated false positive in the observed run.

Supporting evidence

Qualification The auditor recognized the mirror design, so the case does not isolate structural discrimination from author-intent inference.

CLAIM-N01-05

The six-round series is a methodological case note, not a general model reliability benchmark.

Supporting evidence

Qualification Short-round agreement is agreement with sealed keys. The later synthesis is internal SECONDARY_ANALYSIS, WITHHELD from public v1.0; public support rests on the round key/result records.

Transformations & provenance

Seven provider-native documents were exported as public PDF representations. The source record class and the public representation class are separate: an exported PDF does not become the original provider-native bytes. The E14 sealed key and other plain-text keys retain their approved source bytes. A structured whole-turn selection was rendered into the public 43-turn HTML relay using a restricted template.

The selected relay is a post-hoc bounded selection: 37 turns are fully paired and the preserved Claude snapshot provides only truncated previews for six Claude-side full bodies. Those bodies were not reconstructed. Full snapshots and internal synthesis remain withheld.

Original data: transformation-log.json · release README

Integrity & reproducibility

The SHA256SUMS file lists SHA-256 checksums of released v1.0 files. Compare a downloaded file against its listed checksum to check byte integrity. A match verifies those bytes against the published list; it does not prove an interpretation or the correctness of a sealed key.

Released v1.0 bytes are fixed. A material correction requires a new version and changelog. Read the v1.0 changelog.