# Lookahead post: evidence bundle

Prepared September 10, 2026 for Residual Thoughts from the recorded July 20
lookahead experiment. Research source files are copied byte for byte. This
bundle packages existing outcomes; it records no new model run.

- PREREG.md: frozen protocol and dated amendments.
- REPORT_LADDER.md: original result report, retained without editorial changes.
- outcome_ladder.json: recorded per-seed fp32 outcome evaluations and baseline.
- analysis_ladder.json: recorded statistical analysis and score diagnostics.
- gate_*.json: saved estimator, reproduction and capture gate evidence.
- src/: original E6 experiment/analysis scripts, for inspection.
- make_lookahead_blog_figures.py: post renderer; requires numpy and matplotlib.
  Run `python make_lookahead_blog_figures.py` here to regenerate the figures
  from the two JSON inputs. It neither loads a model nor recomputes inferential
  statistics. It checks the plotted paired differences against the analysis.
- SHA256SUMS.txt: hashes of the files actually included in this bundle.
- original_artifact_manifest.txt: the historical manifest for larger research
  artifacts, retained for provenance. It is not the bundle's file inventory.

This is sufficient to reproduce the post's plotted values and inspect its
reported analysis. It is not a complete training reproduction: model checkpoints,
datasets, score arrays, masks and parent-program dependencies are not included.
The source scripts assume the original multi-directory research layout.

The report's broad registration summary must be read alongside its detailed
Amendment 2 disclosure: repeat-stability thresholds were set after the repeat
check was observed. The post preserves that distinction. The missing persisted
L1/L4 sketch-convergence margins are also disclosed in the report and post.

Numbers use the E6 fp32 outcome convention throughout. They should not be
silently pooled with the previous article's differently evaluated outcomes.