docs: the format catalog count is a live invariant, not a fixed number - #3186
Merged
Merged
Conversation
The SSOT in gHashTag/t27 grew from 83 records (arXiv:2606.09686v2, 2026-06-22) to 109 on 2026-08-11 (TNF, BNF and GF-T ladders, commit b928725). This README still quoted 83 as if it were the catalog's name. The number now appears only as a dated fact; the catalog is referred to without a count so the text does not go stale again when the SSOT grows. arXiv v3 (submit/8037924) is retitled 'Golden Ruler: A Numeric Format Catalog with Bit-Exact Conformance Vectors'. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
This was referenced Sep 4, 2026
Merged
tri skill lost: the audit that found nothing in SKILL.md, and two bare headings in docs/NOW.md
#3206
Merged
This was referenced Sep 5, 2026
…that was never made The catalog grew to 109 formats on 2026-08-11 (b928725) and every pack index moved with it: bit-precise 69 -> 89, self-consistent 6 -> 0, structural 8 -> 20, witnessed 10. conformance/vectors/README.md still listed the 83-pack set of 2026-06-14, so its tables, totals and SHA-256 list are re-derived from INDEX_all_formats.json and the live pack files. The positioning doc, the two gf256 witness notes and the tnf-gfternary skill quoted 83 as the catalog's name; they now carry it as a dated measurement with the live count beside it. The catalog paper is cited under its v3 title, "Golden Ruler: A Numeric Format Catalog with Bit-Exact Conformance Vectors for FP8, BF16, MXFP4, and Microscaling Formats" (arXiv:2606.09686, v3 announced 7 Sep 2026). Hardware: no die of any TRI-NET chip exists. The Tiny Tapeout TTSKY26a and TTSKY26b submissions were withdrawn and refunded, so the README no longer calls the three chip repos "silicon variants", the positioning diagram's oracle is the Artix-7 (XC7A200T) FPGA prototype, and the skill's "Verified against silicon" section is verified on that prototype. tools/check_catalog_count.py: PAPER_DECLARED_COUNT stays 83 on purpose -- the live arXiv entry still returns v2 today, and the constant's own rule is that it moves only when the fetched entry shows it. The pending 83 -> 109 bump and its evidence command are recorded in the comment. Gate run: SSOT == fresh regen == 109, exit 0; --self-check 6/6. The TNF manuscript is under review at Microprocessors and Microsystems (MICPRO-D-26-00839, submitted 3 Sep 2026); no arXiv preprint exists. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This was referenced Sep 5, 2026
… in June, and six living lines still carried the old count Fix-up after review of 839e669. README.md: the Hugging Face catalog dataset was described as "the format-catalog SSOT (109 records as of 2026-09-04)". The mirror is dataset v3.0, last commit 2026-06-13 ("v3: count=83 truth"), 83 rows on datasets-server, and it was not touched on 2026-09-04. It is now described as a snapshot of 2026-06-13 that lags the live SSOT and is due for re-publishing; the CI-invariant sentence is attached to the SSOT only. Verified: curl https://huggingface.co/api/datasets/playra/numeric-format-catalog (lastModified 2026-06-13T15:12:13Z) and https://datasets-server.huggingface.co/size?dataset=playra/numeric-format-catalog (num_rows 83). Hard-coded count removed from living lines, keeping the dated qualifier: .claude/skills/tnf-gfternary.md (three sites: "the catalogue as then registered", "as it stood then", "as it stood when measured"), conformance/witness/gf256/README.md and SEPARATION_BOUND.md ("71 of the formats then catalogued"; "it has grown since this witness was written"), conformance/vectors/gen_all_formats.py (the gate enforces "the catalog-count invariant; the count is live, not fixed"; "every pack file"). Oracle wording: gen_all_formats.py and conformance/vectors/README.md said gf16 has a "silicon oracle" / is the "silicon-oracle anchor". The oracle is the Artix-7 XC7A200T FPGA prototype; no die exists. Now "FPGA oracle" / "FPGA-oracle anchor, XC7A200T", matching docs/POSITIONING_CONFORMANCE_LAYER.md. tnf-gfternary.md heading: "Verified on the Artix-7 FPGA (XC7A200T) prototype" asserted board execution the source does not state. docs/theory/TNF_ARTICLE_RU.md:968 says the synthesized node is compared against exact integer arithmetic in Z[phi]; the heading now says exactly that: "Verified against the synthesized node (exact integer reference in Z[phi])". tools/check_catalog_count.py: the paper count is keyed by arXiv version (PAPER_DECLARED_BY_VERSION = {v1: 84, v2: 83, v3: 109}, PAPER_LIVE_VERSION = "v2"), so the 7 Sep 2026 switch is a one-token change and the superseded title is no longer quoted verbatim. The evidence rule is unchanged: the live version moves only when the curl shows it (fetched 2026-09-05: still 2606.09686v2, updated 2026-06-22T12:28:45Z). The gate output now names the version it compares against. Gate exit 0 (SSOT == fresh regen == 109), --self-check 6/6, --strict-paper exit 3 as designed. docs/theory/TNF_ARTICLE_RU.md bibliography entry 5: title token and version only, "Golden Ruler: a numeric format catalog with bit-exact conformance vectors for FP8, BF16, MXFP4, and microscaling formats. arXiv:2606.09686v3". Body lines 108 and 247 ("каталог из 83 форматов", "Из 83 записей") are the manuscript's own measurement statements and are left as they are. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-09-05 04:22:41 UTC
Summary
Seal Status
|
This was referenced Sep 5, 2026
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-09-05 04:52:28 UTC
Summary
Seal Status
|
This was referenced Sep 5, 2026
…PR itself had no NOW entry and no issue Fix-up after review of PR #3186 (Refs #2466). - conformance/vectors/crosswalk_sw_hw.py loaded INDEX_all_formats.json and then hard-coded 83 into the header it writes; the total is now len(rows) (dry run in a temp dir: "(=109)"); the July #199 HW snapshot keeps its 71/83, dated. Schema v0.2: totals.ssot_total added, hw_ids_outside_ssot83 -> hw_ids_outside_ssot (no consumer in the tree). - specs/numeric/golden_sieve.t27: the W778 comment says which catalogue it measured (83 on 2026-08-15) and that the count is a CI invariant (109 at v3). Resealed with a freshly built t27c: seal --save then --verify, all five hashes MATCH; only spec_hash and sealed_at moved, the four gen_hash_* are byte-identical. - docs/theory/IGLA-FORMAL-RESULTS.md: a reading note in the header -- in this ledger "silicon" is the FPGA die on the board an entry names, never a fabricated chip (TTSKY26a/26b withdrawn before fabrication and refunded), and each entry quotes the catalogue count as it stood when measured; the W602 line carries the qualifier inline. - docs/theory/TNF_ARTICLE_RU.md: reference 11 pointed at hackaday.io/project/183153 (404, no Wayback snapshot); it now points at hackaday.io/project/164907 (Ternary Computing Menagerie, 200, hosts "Ternary27 Standard.pdf"), which the corrected English bibliography in docs/reports/upstream/tnf_paper.FIXED.tex already carries. The provenance header notes that the 83 in the body is the translated revision's count (83 = 11 + 72 kept) and that the catalogue is 109 at v3. - docs/now/2026-09-05-the-catalog-count-is-a-live-invariant.md: the entry the required gates check-now-freshness and check were red for. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Contributor
PR DashboardGenerated at: 2026-09-05 06:46:07 UTC
Summary
Seal Status
|
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-09-05 07:07:32 UTC
Summary
Seal Status
|
gHashTag
enabled auto-merge
September 5, 2026 07:45
Contributor
PR DashboardGenerated at: 2026-09-05 07:47:06 UTC
Summary
Seal Status
|
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
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.
Refs #2466 — the
tools/check_catalog_count.pycomment this PR rewrites is the paper-divergence WARN that issue is about; the WARN itself stays until arXiv v3 is live (announces 7 Sep 2026), so the issue is referenced, not closed.The SSOT in gHashTag/t27 grew from 83 records (arXiv:2606.09686v2, 2026-06-22) to 109 on 2026-08-11 (TNF, BNF and GF-T ladders, t27 commit b928725). This README still quoted 83 as if it were the catalog's name.
The number now appears only as a dated fact, and the catalog is referred to without a count, so the text does not go stale again when the SSOT grows. Companion change: arXiv v3 (submit/8037924) is retitled 'Golden Ruler: A Numeric Format Catalog with Bit-Exact Conformance Vectors' and adds a section on the three ladders.
Second commit (2026-09-05, 839e669): the rest of the living docs on this surface.
conformance/vectors/README.mdre-derived fromINDEX_all_formats.jsonand the pack files: 69/6/8/83 -> 89/0/20/109 (witnessed 10), SHA-256 list regenerated for all 109 packs, the 26 TNF/BNF/GF-T packs of b928725 added to the tables, the self-consistent section closed (the six wide GoldenFloat rungs were promoted throughconformance/witness/).docs/POSITIONING_CONFORMANCE_LAYER.md,conformance/witness/gf256/{README,SEPARATION_BOUND}.md,.claude/skills/tnf-gfternary.md: 83 becomes a dated measurement next to the live count; the catalog paper is cited under its v3 title, "Golden Ruler: A Numeric Format Catalog with Bit-Exact Conformance Vectors for FP8, BF16, MXFP4, and Microscaling Formats" (arXiv:2606.09686, v3 announced 7 Sep 2026).tools/check_catalog_count.py:PAPER_DECLARED_COUNTdeliberately stays 83 until the live arXiv entry shows v3 (the constant's own rule); the pending 83 -> 109 bump and its verification command are documented in place.python3 tools/check_catalog_count.py-> exit 0 (SSOT == fresh regen == 109);--self-check6/6.Not touched (historical):
specs/numeric/golden_sieve.t27,ERRATA_*.md,docs/reports/*.Third commit (2026-09-05, 32bf814): fix-up after review.
README.md: the HF catalog dataset is dataset v3.0, a snapshot of 2026-06-13 that lags the live SSOT (verified: HF API lastModified 2026-06-13T15:12:13Z, datasets-server num_rows 83, last commit "v3: count=83 truth"); the earlier "109 records as of 2026-09-04" wording was wrong and is gone. The CI-invariant sentence now belongs tospecs/numeric/formats_catalog.t27only..claude/skills/tnf-gfternary.md(3),conformance/witness/gf256/README.md,conformance/witness/gf256/SEPARATION_BOUND.md(2),conformance/vectors/gen_all_formats.py(2).gen_all_formats.pyandconformance/vectors/README.mdis now "FPGA oracle" / "FPGA-oracle anchor, XC7A200T" (no die exists; matchesdocs/POSITIONING_CONFORMANCE_LAYER.md).tnf-gfternary.mdheading: "Verified against the synthesized node (exact integer reference in Z[φ])", which is whatdocs/theory/TNF_ARTICLE_RU.md:968states; the previous heading asserted board execution the source does not state.tools/check_catalog_count.py: paper count keyed by arXiv version (PAPER_DECLARED_BY_VERSION,PAPER_LIVE_VERSION = "v2"), so the 7 Sep 2026 switch is a one-token change and the superseded title is no longer quoted verbatim; evidence rule unchanged (curl 2026-09-05 still returns 2606.09686v2). Gate exit 0,--self-check6/6,--strict-paperexit 3 as designed.docs/theory/TNF_ARTICLE_RU.mdbibliography entry 5: title token and version only (Golden Ruler, arXiv:2606.09686v3). Body lines 108 and 247 are the manuscript's own measurement statements, left as they are.Fourth commit (2026-09-05, 8e686a2): fix-up after the second review.
docs/now/entry and referenced no t27 issue, socheck-now-freshness,checkandcheck-linked-issuewere red. Addeddocs/now/2026-09-05-the-catalog-count-is-a-live-invariant.md(both NOW gates pass locally with the CI scripts,scripts/ci/now-sync-gate-diff.shandtools/check_now_entry_shape.py), and this body now carriesRefs #2466.conformance/vectors/crosswalk_sw_hw.py: it read the SW total fromINDEX_all_formats.jsonand then printed a hard-coded 83 into its own header; it now prints the count it read (temp-dir dry run:(=109); the July Ring 052: Lotus phase automation — brain seal refresh pipeline #199 HW snapshot keeps its 71/83, dated; schema v0.2,totals.ssot_totaladded,hw_ids_outside_ssot83->hw_ids_outside_ssot, no consumer in the tree). The committedCROSSWALK_sw_hw.mdJuly snapshot is not regenerated.specs/numeric/golden_sieve.t27: the W778 comment is pinned to its run (83 formats on 2026-08-15; the count is a CI invariant, 109 at v3). Resealed with a freshly builtt27c(seal --save, then--verify: all five hashes MATCH; onlyspec_hashandsealed_atchanged, the fourgen_hash_*are identical).docs/theory/IGLA-FORMAL-RESULTS.md: a reading note in the header — "silicon" in this ledger is the FPGA die on the board an entry names, never a fabricated chip (TTSKY26a/26b withdrawn before fabrication and refunded), and each entry quotes the catalogue count as it stood when measured; the W602 line carries the qualifier inline. The wave-numbered entries themselves are not rewritten.docs/theory/TNF_ARTICLE_RU.md: reference 11 pointed at hackaday.io/project/183153 (404, no Wayback snapshot) and now points at hackaday.io/project/164907 (Ternary Computing Menagerie, HTTP 200, hosts "Ternary27 Standard.pdf"), whichdocs/reports/upstream/tnf_paper.FIXED.texalready carries; the provenance header notes that the body's 83 is the translated revision's count (83 = 11 + 72 kept) and that the catalogue is 109 at v3. Reference 28's DOI 10.1145/3649329.3656502 resolves at Crossref (FQP, DAC '24); only dl.acm.org refuses non-browser fetches — unchanged.tools/check_catalog_count.py:49submit/8037924: verified 2026-09-05 against the arXiv notification "arXiv replacement submit/8037924 for 2606.09686" (e-prints@arxiv.org, received 4 Sep 2026) — kept.🤖 Generated with Claude Code