# Sīdus E1 RC4 — working research record

Version 0.2 · 6 November 2026. Prepared for Ben Amuwo with AI-assisted evidence review.

The paper adds a reading guide to the 6 November numerical audit. It preserves the RC4 decision, thresholds, populations and open interruption history. The older canonical claim, charter and freeze manifest are unchanged. The author authorized publication of this working record and companion on sidus.quest on 6 November 2026. It is not a peer-reviewed article, an independent replication, or a new scientific run.

## Read the evidence at its level

- The original guarded analyzer reproduced the final adjudication byte-for-byte.
- A separate, source-aware calculation matched 481 comparisons using the same SciPy statistical library. This is arithmetic reproduction, not an independent experiment.
- Local identity checks matched 95/95 bound records. The audit checked the recorded ledger and backup payloads; hashes do not establish independently authenticated chronology.
- The current environment verifier passed, but a historical distribution-verifier refusal and later restart remain unresolved. The author recalls an unplanned stop and does not know the cause.
- Primary U is high-capacity; low-capacity replacement is a descriptive diagnostic. Invalid primary components contribute zero in the all-seed confirmation population.
- Fresh checkpoint-derived readout reconstruction and reference-certificate regeneration were not performed. Earlier technical reviewers were AI, and unaffiliated human review remains pending.

## How the companion is derived

`EVIDENCE_INDEX.json` binds the selected evidence files and compact display data. `scripts/build_e1_data.py` selects recorded final fields and audited component summaries; it does not fit a model, generate seeds or change hypotheses. Rate shares normalize each seed’s rate vector, and displayed averages average those shares. V compares per-seed L1 errors, not distance to a mean allocation.

The seed selector reads existing seeds 9201–9701. The formula panel shows raw values, validity and the resulting composite contribution. The dashed budget panel uses invented teaching values and predicts no return. Individual one-sided 95% t lower bounds are not two-sided 95% confidence intervals. Numerical underflow is labelled rather than reported as probability zero.

The paper’s supplement indexes the full local audit bundle retained separately. This website carries only selected records: no ledger, sealed payloads or private keys. Large files are requested only through evidence links, not background visualizer fetches. No paid service or model call is used by the companion.

## Remaining decisions

The current working-draft release includes the stated author and AI assistance. Formal submission, archival licensing/citation choices and subsequent revisions remain author decisions. The unexplained interruption must remain disclosed unless actual recovery evidence resolves it. Independent replication planning remains NOT AUTHORIZED FOR EXECUTION. This outcome-exposed companion is not a clean-room specification.