repro/numerics: measured L4/L5 GF16 vs bf16/fp16 accuracy oracle (Closes #1072) - #1073
Closed
gHashTag wants to merge 1 commit into
Closed
repro/numerics: measured L4/L5 GF16 vs bf16/fp16 accuracy oracle (Closes #1072)#1073gHashTag wants to merge 1 commit into
gHashTag wants to merge 1 commit into
Conversation
Add repro/numerics/{nmse_gf16.py,nmse_manifest.json,README.md} and fill the
L4/L5 TBD cells in docs/NUMERICS_VALIDATION.md with measured, reproducible
round-trip NMSE/ULP numbers for GF16 (E6M9) vs bfloat16 and float16, per
docs/GF16_BFLOAT16_NMSE_PROTOCOL.md. Host-measured, unsealed => informational,
not a silicon certifying claim. D_PHI is an identity sanity check.
Closes #1072
gHashTag
force-pushed
the
repro/numerics-l4-l5-oracle
branch
from
June 14, 2026 11:38
7af9080 to
aa395e9
Compare
Contributor
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Owner
Author
|
Этот PR полностью перекрыт уже смерженным #1075 (коммит 2962148 на master). #1075 был собран поверх ветки #1073 и squash-merge'нут, поэтому master уже содержит все изменения из #1073 ПЛЮС сертифицирующий NMSE-манифест и seal-hash honesty. Мердж #1073 (aa395e9) сейчас УДАЛИЛ бы repro/numerics/nmse_manifest_protocol_v1.json и repro/numerics/validate_manifest.py и откатил бы улучшения Makefile/docs. Измеренный L4/L5 oracle живёт на master. Работа отслеживается в #1072. Закрываю как superseded. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes #1072.
Fills the L4 (differential) and L5 (comparative) TBD cells in
docs/NUMERICS_VALIDATION.mdwith measured, reproducible round-trip accuracy numbers for GoldenFloat GF16 (E6M9) versus bfloat16 and IEEE float16, perdocs/GF16_BFLOAT16_NMSE_PROTOCOL.md.What's added
repro/numerics/nmse_gf16.py— round-tripsreal -> format -> realthrough the GF16 codec SSOT (conformance/gf16_ref.py),ml_dtypes.bfloat16, andnumpy.float16; computes NMSE, NMSE ratios, max-abs-err, a ULP-like (kappa-approx) metric, and an overflow/saturation rate overD_NORM,D_LOG,D_RELU,D_PHI,D_DEEP, plus a documented wide-range extensionD_WIDE. Enforces the L5 identity witness and aborts non-zero on failure or negative NMSE.repro/numerics/nmse_manifest.json— protocol section-6 manifest (seed 2718281, 2,000,000 samples/distribution).repro/numerics/README.md— measured results table + honest interpretation.docs/NUMERICS_VALIDATION.md— TBD cells in sections 4/5/6 replaced with measured values; section 11 + closing line updated.Measured headline (host, unsealed)
D_WIDE(log2|x|~U(-28,28)) fp16 overflows 21.4% of samples; GF16 (max ~4.29e9) and bf16 lose none — the honest exponent/mantissa tradeoff (protocol section 3.3), matching IBM DLFloat (DOI 10.1109/ARITH.2019.00023) and Popescu et al. (arXiv:2103.15940).Honesty (R5-HONEST)
Host-measured, unsealed => informational, not a silicon certifying claim.
D_PHIis an identity sanity check, not a superiority claim. ASCII-only. L6 SSOT untouched.Closes #1072.