Skip to content

docs(paper): MICPRO R1 — path roots, scale convention, log count - #742

Merged
gHashTag merged 1 commit into
mainfrom
docs/micpro-r1-artefact-roots
Sep 5, 2026
Merged

gHashTag merged 1 commit into
mainfrom
docs/micpro-r1-artefact-roots

Conversation

@gHashTag

@gHashTag gHashTag commented Sep 4, 2026

Copy link
Copy Markdown
Owner

Revision R1 of MICPRO-D-26-00839, prepared 2026-09-05. Not submitted to the journal; that is the owner's account.

What a sibling report called 15 dead artefact links were 14 live files under research/arxiv_tnf/measurements/, cited relative to the paper's own directory. The one real defect was that the paper never named the root. Fixed with one paragraph.

Changes (all in tnf_paper.tex, plus a record in SUBMISSION.md):

  • Where the record files are: repository, both path roots, submission tree a0fb006 (tex last changed 64400a1).
  • Lloyd–Max paragraph: no longer a "ceiling"; sample described as the code does it (strided, ≤80k values per tensor, research/block/block_bound.py); scale rule stated as s = 2^⌈log2(amax/top)⌉, with the spec's rule and its 23.5380 cited from MXFP4_SCALE_CONVENTION_2026-08-11.md.
  • "2026 literature varies the transform": now names HBQ (arXiv:2609.00450), M2XFP (arXiv:2601.19213), VS-Quant (arXiv:2102.04503), shared microexponents (arXiv:2302.08007); four \bibitems.
  • Log count: "zero *.log files" was false at a0fb006 (five e2m11_add seed logs in measurements/pnr_logs/); withdrawal note named, and it retracts itself; fpga/phiscale = 163 files.

No table number changed. tectonic builds clean, 147 pp. tools/check_paper_numbers.py flags none of the new literals (its three pre-existing unit-carrying flags, 62.69/64.16/71.90 MHz, are untouched).

🤖 Generated with Claude Code

…e scale convention, correct the log count

- paragraph 'Where the record files are': measurements/ paths resolve from
  research/arxiv_tnf/, research/ and tools/ from the repository root;
  submission tree a0fb006, tnf_paper.tex last changed in 64400a1
- Lloyd-Max: not a ceiling on perplexity; the 18,850,950 values are a strided
  <=80k-per-tensor sample (block_bound.py), not whole blocks; scale rule
  s=2^ceil(log2(amax/top)) stated, spec rule gives 23.5380 on the same run
  (MXFP4_SCALE_CONVENTION_2026-08-11.md)
- '2026 literature varies the transform' now names HBQ, M2XFP, VS-Quant and
  shared microexponents, which vary the scale field; four bibitems added
- 'git ls-files finds zero *.log' was false at a0fb006: five e2m11_add seed
  logs under measurements/pnr_logs/; withdrawal note named and marked as
  self-retracting; fpga/phiscale holds 163 files, not 145
- SUBMISSION.md: journal submission record (MICPRO-D-26-00839)

No table number changed. Built clean with tectonic (147 pp).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
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