Skip to content

docs: run-004 reproduced the retracted verdict on the mojibake snapshot - #68

Merged
ronaldtse merged 1 commit into
mainfrom
docs/tiny-mojibake-findings
Aug 28, 2026
Merged

docs: run-004 reproduced the retracted verdict on the mojibake snapshot#68
ronaldtse merged 1 commit into
mainfrom
docs/tiny-mojibake-findings

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Documents the run-004 forensics: the "clean-label re-run" spec was
still pointing at the Aug-23 snapshot; its best/ decodes as mojibake
and its gate eval reproduces the retracted 83.08 to four decimals with
the teacher reproducing 1.3205 (final_eval.json = provenance).
RESULTS.md gets the finding ahead of the client-tier section;
PUBLICATION-NOTES.md gets the methodological lesson (label provenance
must be content-hashed into run records; identical numbers are only
reproducibility evidence when the input pipeline is versioned).
run-005 (labels regenerated live from the r6 teacher) is the honest
falsification test, in flight.

Identical-to-4-decimals (83.0797 vs 83.08) + live mojibake decode of
best/ = same poisoned data, not a capacity result. run-005 is the real
clean-label test.
@ronaldtse
ronaldtse merged commit f680717 into main Aug 28, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant