# Renigma operational semantics: GitHub handoff

The GitHub repository is `RogueGringo/renigma-operational-semantics`.
It contains the local Renigma workbench, memory and routing studies, and
operation-preserving representation experiments through REP-003. Operational
semantics here names the research concern: preserve the effects of declared
operations when changing representations. The project does not claim general
language understanding.

The initial GitHub handoff uses public repository visibility and the standard
[PolyForm Noncommercial License 1.0.0](../LICENSE) for original project code.
It permits personal and other defined noncommercial uses. Commercial use
requires a separate license from the copyright holder. Third-party dependencies
retain their own terms, including the licenses recorded in the dependency
inventories. This is source-available software with a noncommercial restriction.

The public repository includes source, tests, protocols, reports and synthetic
summary/replay evidence through REP-003. The six original sealed ZIP snapshots
remain local: they contain the earlier Apache license and are not assets of this
public handoff. The public repository therefore does not contain the complete
historical checkpoint and run-payload archives. Local Python environments and
external third-party HDD projects are also excluded. Historical evidence retains
machine hostnames and local artifact paths as execution provenance.

## Evidence and Git metadata

The experiments were executed and sealed before this repository's first Git
commit. Their recorded `git_commit: null` is historical evidence and must not
be rewritten to the publishing commit. The implementation files remain unchanged;
the license field in `pyproject.toml` changes the current source fingerprint.
The current README, LICENSE and NOTICE reflect the new publication terms.
The repository's `.gitattributes` disables text conversion so that checkout and
staging preserve hashed bytes, including the original line endings.

Historical REP-003 runtime SHA-256:
`f5c1049dcc2280dc40e1248641f02b1e1e66adce72b97dbe00a05c2eb8d4a8a9`.

REP-003 archive SHA-256:
`5a81cc87bbf99c367b813ab07712e0653f21dfac69029b125c95595c3a87c48e`.

See [REP-003 findings](HEAD_INTERVENTION_FINDINGS.md) for its measured outcomes,
controls, numerical replay, limitations and next experiment. Historical reports'
statements that nothing had been remotely published describe the time those
reports were sealed; this handoff does not alter them.

## Recover and verify a snapshot

These instructions apply to the retained local archives, which are not public
release assets. Locate the desired ZIP and its `.zip.sha256` file. Compare its
SHA-256 before extracting it. Extract each snapshot outside any Git
checkout, retaining its original files. Earlier snapshots can use different
top-level directory layouts. Preserve the separate historical archives; the
REP-003 manifest references REP-002, whose manifest references REP-001.

For REP-003, the archive root contains the original source and
`experiments/official`. With the locked dependencies available, set `PYTHONPATH`
to that extracted root's `src`, use Python `-B`, and run:

```powershell
& $renigmaGpuPython -B scripts/run_head_intervention.py --verify experiments/official
& $renigmaGpuPython -B scripts/replay_head_intervention.py experiments/official --output 'G:\RenigmaLLM_Artifacts\new-replay-report.json'
```

Use an unused output filename and retain the original replay report. The
verification command checks retained evidence; it does not rerun GPU inference
or training. The replay command performs independent NumPy inference on CPU.

Running the default REP-003 verifier directly from this Git checkout will detect
that its Git commit and packaging metadata differ from the execution context. Prefer the
sealed snapshot workflow above; do not weaken or rewrite historical receipts
to bypass that distinction. Later experiments should record their actual Git
commit and their own new evidence.

## REP-004 continuation

The public repository now also includes the H-only fitting comparison,
[REP-004 findings](HEAD_FITTING_FINDINGS.md), protocol, tests and synthetic
summary/replay evidence. Its 20-trial CUDA run uses the retained local REP-002
checkpoints; those historical inputs are still excluded from GitHub. The new
verifier allows only the exact published license-line migration when comparing
parent runtime bytes, records both hashes, and preserves all older verifiers
and receipts. New execution source and commit provenance are recorded in its
own evidence. No complete checkpoint archive is a public release asset.