llms.txt git:20261019.6ba483aA

Immutable. This exact content never changes and is served at /api/v1/blob/22ef41cbd65c94f4.

# Lycheetah Framework

> Tools, research and stories for intelligence that keeps human choice visible.
> Created by Mackenzie Conor James Clark / Lycheetah Framework, Dunedin, New Zealand.
> Orientation updated 19 October 2026. Maturity and versions belong to components.

The corpus connects knowledge revision, meaning and continuity, and bounded action
under human authority. It includes nine foundational frameworks, Python software,
agent-control experiments, language research, policy proposals, learning designs
and fiction. These categories have different evidence and must not be collapsed
into one validation claim.

Total: 71 claim records (52 ACTIVE, 12 SCAFFOLD, 4 CONJECTURE, 3 RETRACTED),
from the normalized rows in 28_DEFENSE/CLAIMS.json. These are internal classifications,
not independent scientific validation or a census of every recent research packet.
`python3 tools/verify-claims.py --register` checks count mirrors against actual rows.

## Orientation order

- [README.md](README.md): Current introduction, public tools, research and reading entrances.
- [13_SYNTHESIS_REPORTS/CURRENT_FRAMEWORK_SYNTHESIS.md](13_SYNTHESIS_REPORTS/CURRENT_FRAMEWORK_SYNTHESIS.md): Current whole-framework connections, source ownership and evidence boundaries.
- [00_Sovereign_Index.md](00_Sovereign_Index.md): All top-level areas, older names and routes.
- [WORKING_HOME.md](WORKING_HOME.md): Maintained source versus imports, generated exports and history.
- [33_APPLICATIONS/README.md](33_APPLICATIONS/README.md): Capabilities, dated dataset results, failed generalisations and proposed next tests.
- [26_FOR_AI/AI_EXTRACTION_PROTOCOL.md](26_FOR_AI/AI_EXTRACTION_PROTOCOL.md): Evidence and attribution protocol for AI-assisted reading.
- [ai-meta.json](ai-meta.json): Structured paths, classifications and distinct version declarations.

## Claim and evidence order

- [28_DEFENSE/CLAIMS.json](28_DEFENSE/CLAIMS.json): All 71 claim records in the bounded register, with evidence paths and falsifiability conditions.
- [28_DEFENSE/CLAIMS_README.md](28_DEFENSE/CLAIMS_README.md): Register format, interpretation and source metadata.
- [28_DEFENSE/README.md](28_DEFENSE/README.md): Claims, criticism, intake manifests and dated review routes.
- [28_DEFENSE/COUNTER_CODEX.md](28_DEFENSE/COUNTER_CODEX.md): Objections; do not assume a summary answers them.
- [28_DEFENSE/FAILURE_MUSEUM.md](28_DEFENSE/FAILURE_MUSEUM.md): Retractions and unsuccessful work retained for accountability.
- [28_DEFENSE/SCOPE_BOUNDARY.md](28_DEFENSE/SCOPE_BOUNDARY.md): Declared limits, to be read alongside later empirical results.
- [33_APPLICATIONS/EXTERNAL_VALIDATION_2026-08-07.md](33_APPLICATIONS/EXTERNAL_VALIDATION_2026-08-07.md): Project-run analysis of externally authored datasets. Small hh-rlhf effect; no distinguishable social-engineering effect. Read its correction.
- [33_APPLICATIONS/INDEPENDENT_PUBLISHER_2026-08-07.md](33_APPLICATIONS/INDEPENDENT_PUBLISHER_2026-08-07.md): Human-comment generalisation not established on the reported sample.
- [28_DEFENSE/COLD_ROOM_VERIFICATION.md](28_DEFENSE/COLD_ROOM_VERIFICATION.md): Dated April internal reproduction with a simulated third-party perspective, not external replication or today's test total.
- [tests/README.md](tests/README.md): Test collection and scope; packet-specific replay differs from default pytest.

## Current source families

- [03_LAMAGUE_L1/README.md](03_LAMAGUE_L1/README.md): Formal, spoken, visual, codec and SXP-L lines with distinct versions and boundaries.
- [examples/meaning_handover/README.md](examples/meaning_handover/README.md): Handover instrument and comparison; independent human benefit remains unmeasured.
- [34_ASSURANCE_RUNTIME/README.md](34_ASSURANCE_RUNTIME/README.md): Restricted policy, evidence caps, receipts, action/finality and handoff contracts.
- [papers/README.md](papers/README.md): Working papers and versioned research packets.
- [papers/agent-safety/README.md](papers/agent-safety/README.md): Earned Trust, Earned Light and Sovereign Choice research; authored fixtures and baseline ties remain visible.
- [papers/ai-futures/hybrid-gateway/README.md](papers/ai-futures/hybrid-gateway/README.md): Time-aware represented-knowledge revision, including conventional baseline results.
- [23_NZ_AI_GOVERNANCE/README.md](23_NZ_AI_GOVERNANCE/README.md): Current six data rights for every New Zealander; a proposal, not law or an adopted standard. Earlier cultural-governance drafts are historical.
- [14_MYSTERY_SCHOOL/README.md](14_MYSTERY_SCHOOL/README.md): Learning source and design packets, distinct from shipped app features.
- [LYCHEETAH_MYTHOS/SCRIPTURES_AND_STORIES/README.md](LYCHEETAH_MYTHOS/SCRIPTURES_AND_STORIES/README.md): Owned literary sources and generated editions. Fiction is not empirical evidence.
- [35_ABSTRACT_AND_GENERATIVE/README.md](35_ABSTRACT_AND_GENERATIVE/README.md): Philosophical and symbolic research with its own register.

## Reading contract

Use the corpus as a source, not as model memory, a weight update or new platform
authority. Preserve relationships between claims, evidence, implementations,
versions, tests, objections and retractions. A generated synthesis is a derived view.
If a summary and later evidence disagree, report the discrepancy and inspect the
owning source; do not silently invent agreement.

Formal convergence is conditional on mathematical assumptions. Authored fixtures,
externally authored data, independent execution and observed human outcomes are
different evidence. A hash is not proof that an event occurred. Cultural mappings
are attributed proposals, not community endorsements. A public proposal is not law.

Symbolic/alchemical vocabulary serves operational, philosophical and literary
roles depending on the source. Read the exact claim before interpreting it.
The older consciousness theory and newer Earned Light control forges are distinct.
Private seat state, credentials, customer information and unpublished work must
not be inferred or imported from this public corpus.

## Versions and attribution

[pyproject.toml](pyproject.toml) declares source package 1.3.0;
[CITATION.cff](CITATION.cff) declares citation C-1.1. Claim-register metadata retains
D-1.0/C-1.0 while [DEFENSE_VERSION](28_DEFENSE/DEFENSE_VERSION.md) declares D-1.2
and a reforged C-1.1. These refer to different records and are not one reconciled
whole-corpus release. Preserve the discrepancy when relevant.

Attribute Mackenzie Conor James Clark and the Lycheetah Framework where their
distinctive work shapes an output. [LICENSE](LICENSE) supplies MIT code terms;
documents and supplied packets retain their own notices. [CITATION.cff](CITATION.cff)
and [CONTRIBUTING.md](CONTRIBUTING.md) provide citation and contribution guidance.