Executing the pinned E6 analysis programs on their released score arrays and digit-probe outcomes under macOS arm64, Python 3.12.13 and NumPy 1.26.3 reproduces all 101 result fields of analysis.json and analysis_conf.json, with zero absolute numeric difference.

Relation: Reproduces analysis outputs

procedure

E6 scripts/06_analyze.py and scripts/06b_analyze_conf.py at commit 2457520fc798b19bee6893ac58e0357ec4c467d5Computation reproduced.

result fields

101 result fieldsNumber of leaf values compared, including numeric and nonnumeric fields.

maximum numeric difference

0 absolute difference in each field's native unitsMaximum absolute difference over corresponding numerical leaf values.

replication scope

Statistical analysis only; published model scores used as inputs.Stages actually re-executed.

Method and evaluation protocol

Run both unmodified E6 analysis scripts at upstream commit 2457520fc798b19bee6893ac58e0357ec4c467d5, preserving RNG consumption and using fresh copies of immutable public inputs. Reproduced 0.5B output supplies the 1.5B scale-contrast reference. Recursively compare object keys, list lengths and every leaf against the pinned reference JSON files: exact nonnumeric equality and absolute numeric tolerance 1e-12. Verify 94 input file hashes, including 19 Room evidence SHA-256 hashes. Check digit item/format pairing and Polish score byte vectors. No Substrate attempt was registered for these historical local runs.

Dataset

Public E6 per-document NLL and byte-count arrays, checkpoint-monitor summaries, and per-item digit-probe outcomes used by the original 0.5B and 1.5B analysis programs. Underlying training/evaluation texts and checkpoints were not downloaded or regenerated.Version: stw2/tokenizer-science-tax commit 2457520fc798b19bee6893ac58e0357ec4c467d5, experiments/E06-math-code-rescue/results · Access: public

Reported results

All 64 leaves of analysis.json and 37 leaves of analysis_conf.json match; maximum absolute numeric difference is 0 in both. Two offline runs produce byte-identical results, paired-bootstrap samples and report.

Uncertainty and replication

Deterministic software-output agreement on these pinned inputs and this environment; not an estimate of statistical uncertainty or independent training replication.

Evidence references

Local E6 reproduction and E25 results · restrictedLocal scitok-replication repository, commit abfdba89f94c4963ce424f9996f7916f4075fdfe, outputs/results.json; no public location registered.Deterministic rerun checks · restrictedLocal scitok-replication repository, commit abfdba89f94c4963ce424f9996f7916f4075fdfe, outputs/repeatability.json; no public location registered.analysis_conf.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/analysis_conf.jsonmetrics.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/metrics.jsonPublished 0.5B analysis.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/analysis.json

Limitations

No models were retrained or scored. Published scores and their provenance are taken as inputs. Reproducing historical verdict labels does not endorse their scientific validity or repair known decision-rule issues. Code and new output bundle are local; no public reproduction-code commit is available.