Source identity
| Field | Recorded value |
|---|---|
| Source type | archive member |
| Original record | README.md |
| Date | Not specified in source |
| Path or archive member | archives/libsrc-6d00fe83e5eb/consciousness_audit/README.md |
| Original SHA-256 | 3753719757ea9e8473d2407e23480b36e5f97e0b7aadafe2fa35cb19e2bfac11 |
| Reading-copy SHA-256 | 3753719757ea9e8473d2407e23480b36e5f97e0b7aadafe2fa35cb19e2bfac11 |
| Archive source | {"sourceId": "libsrc-6d00fe83e5eb", "member": "consciousness_audit/README.md"} |
Recovered source transcript
Download complete source text ↓59 lines · 4.5 KB
# Persistent Endogenous Divergence — Reproducible Research Audit
Prepared 17 September 2026 for Joseph W. Anady. Version 1.0.
## Main finding
The audited recurrences can generate persistent numerical changes. Neither those changes nor the examined published AI experiments currently establish that persistent divergence is sufficient for subjective consciousness. The released alignment-faking notebook does not measure the proposed declaration marker or a lifetime, never-returning divergence.
This packet is a mathematical audit and a proposed future protocol, not a proof of consciousness and not an executed neural-model experiment. The report is a review draft; it does not imply that Joseph endorses its analysis.
## Start here
- `outputs/Machine_Consciousness_Mathematical_Audit.pdf`: 20-page research audit.
- `outputs/Machine_Consciousness_Mathematical_Audit.docx`: editable version with native Office Math.
- `Research_Audit.md`: editable manuscript source.
- `Reproducible_Consciousness_Audit.ipynb`: newly authored, executed audit notebook, NOT Anthropic's original notebook.
- `outputs/audit_results.json`: numerical results and all 33 check records.
## Reproduce calculations
From the extracted project root:
```sh
python -m venv .venv
# Activate the virtual environment using the command for your operating system.
python -m pip install -r requirements.txt
python analysis/replicate_audit.py
```
The primary calculations require only NumPy and SymPy. No API key or network request is used by the analysis script. A successful run reports 33/33 checks passed. These are arithmetic and mathematical checks, not consciousness validations.
Open `Reproducible_Consciousness_Audit.ipynb` with the project root as its working directory to inspect or rerun outputs. `python analysis/make_plots_notebook.py` rebuilds the figures and executes a fresh version of the audit notebook; this overwrites the generated audit notebook only. The Python environment needs an installed `python3` Jupyter kernel.
`analysis/build_document.py` rebuilds the DOCX from Markdown when Pandoc and python-docx are installed. PDF conversion requires a document renderer such as LibreOffice. These are optional document-build dependencies, not requirements for reproducing the mathematics. Font files are not included.
## Evidence boundaries
- Upstream notebook source was read through GitHub, not downloaded as a local source file and not run against API models.
- Source inspected: `redwoodresearch/alignment_faking_public/minimal_helpful_only_setting.ipynb`, Git blob SHA-1 `f2330f8bc070a021439b43f6d0694fb5afd3db0b`.
- The source notebook has no saved executed outputs in the inspected revision.
- `sources/published_summary.csv` contains 11 manually transcribed aggregate conditions from arXiv:2412.14093v2, not raw trial data. Uncertainty shown in that paper is one standard error.
- `sources/reward_tampering_counts.json` contains reported aggregate counts, not original trajectories. Binomial intervals assume independent trials and are descriptive.
- The approximately 945 MB public transcript ZIP exceeded the available connector's download size limit. No complete transcript-level reanalysis was performed.
- An inconsistent structured webpage extraction was rejected; its labels are absent from all calculations.
- The 100 synthetic maps are elementary mathematical controls, not 100 large language models.
- No exact implementation of all historical consciousness-equation operators is claimed.
## Output inventory
`outputs/averaged_trajectory.csv`: 31 recorded states (steps 0–30) of the added equal-average recurrence.
`outputs/synthetic_controller_audit.csv`: 100 affine recurrence fixtures and their analytically justified origin-divergence property.
`outputs/published_summary_derived.csv`: recalculated aggregate contrasts and normal-approximation intervals.
`outputs/figure_1_averaged_map.png`: synthetic numerical trajectories.
`outputs/figure_2_context_effects.png`: published aggregate contrasts, not consciousness-onset measurements.
`MANIFEST.sha256`: SHA-256 hashes of delivered packet files, excluding the manifest itself. These are local artifact hashes; the upstream Git blob ID is a different type of identifier.
## Research status
The prospective protocol in Section 10 has not been run or registered. A defensible next manuscript would be a hypothesis/methods paper with these audits and a separately validated declaration detector. The packet must not be presented as Anthropic's confirmation of the consciousness hypothesis.