feat(blog): one outlier, 23 zeros -- the post on course module 11

[gHashTag]

Closes #1444

#1443 is merged (squash 0e1a789d7): lessons 31-33 are on main. This branch took origin/main by a merge (f5f50a800, no force push) and the base is now main; against main the PR changes only the three blog files below.

What

A bilingual blog post, one-outlier-twenty-three-zeros, on the course's new module 11, "AI numbers: the MX block":

• One byte of scale for 32 weights: what an OCP MX block is and why the E8M0 scale byte has no mantissa.
• One outlier, format by format: in a block of 31 small weights plus one outlier of 3.3, MXFP4 E2M1 flushes 23 of the 30 non-zero small weights to zero (the 31st is exactly zero), MXFP6 E2M3 4, MXINT8 2, MXFP6 E3M2 and MXFP8 E4M3 none.
• A spec that catches its own mistake: changing the E5M2 exponent bias from 15 to 16 fails exactly one of the 49 tests in ocp_mx.t27.
• Where else to learn this: twelve resources compared in a ten-row table as their pages read on 6 October 2026. None of them counts the zeros element by element; graphcore-research/gfloat is the closest in kind.
• For people who build MX, what this does not show, and how to try it.

The terminal recording term/t27c-ocp-mx/session.cast is embedded, and every number has a receipt.

Files (prose and data only)

• apps/website/src/data/blog/bodies/one-outlier-twenty-three-zeros.ts: body (EN, 27 blocks) and ruBody (RU, 27 blocks).
• apps/website/src/data/blog/index.ts: the post's metadata, 27 receipts, and 8 open questions in each language.
• apps/website/src/data/blog/posts.ts: one import line and one map entry.

No code is added. The post has no cover yet; covers come from the static repo and are optional (#1366 shipped without one).

What is not proven (also in the post)

• One block, chosen to show the flush. There is no model and no accuracy number.
• ocp_mx.t27 was merged into t27 on 2026-10-06 (t27#6828, 322cc77d7), but is not on the site: the world scan that vendors t27 specs has been red since 2026-10-03. The ocp_mx recording was made on the branch commit 1a3786657; the file is byte-identical at both commits.
• Lesson 33 is a recording, not a browser player: ocp_mx.t27 is not on the site yet. The site's own evaluator passes all 49 of its tests when run on the file (reproduced in node).
• The comparison covers twelve resources in ten rows, as read on 6 October 2026. It is not a survey.

Review round 2 (360a18d)

The reviewer of 0d82a57 found one more sentence, in both languages, asking for reports "on t27#6828 while it is open". It now asks for an issue in gHashTag/t27. No text in the post calls t27#6828 open any more; tsc rc 0, the EN and RU numbers are still the same multiset.

Status update (0d82a57)

t27#6828 was merged on 2026-10-06 as 322cc77d7 and closed t27#6827. The post called both open; it now says merged and closed, in English and Russian, in the body, the summary, the receipts and the open questions. ocp_mx.t27 is byte-identical at the recorded commit 1a3786657 and at 322cc77d7, so the recording stands. Checked again: tsc --noEmit rc 0, the post still loads first of 87, and the numbers in the EN and RU body, summary and open questions are the same multiset.

Review round 1 (7460c31)

The reviewer checked every number against the spec and the pages and found them right, and named ten imprecise phrases, all fixed: the denominator (the block's 31 small weights include one exact zero, so MXFP4 flushes 23 of the 30 non-zero ones), "all 256 codes" (FP4 has 16, FP6 64), what the 49 tests assert, which recording was made on 1a3786657, twelve resources in ten rows, microxcaling's MXFP6 and INT4/INT2, the Hugging Face MXFP4 blog post as receipt 27, the lesson 33 wording, and the OPEN tag on our own PRs and issues.

{
  "version": 1,
  "head_sha": "360a18d0d3f5d4b9d40481845ceed8c869d9f56f",
  "summary": "A bilingual post on the t27 course's new MX module: how one outlier in a block of 32 flushes small weights to zero format by format, the one-test mutant that proves the spec checks itself, and a comparison with twelve other places that teach MX.",
  "changes": [
    "apps/website/src/data/blog/bodies/one-outlier-twenty-three-zeros.ts: EN body and RU ruBody, 27 blocks each, with the t27c-ocp-mx terminal recording and two tables",
    "apps/website/src/data/blog/index.ts: post metadata first in postsIndex, 27 receipts with their state, 8 open questions in English and 8 in Russian",
    "apps/website/src/data/blog/posts.ts: the import and the map entry for the new body"
  ],
  "tests": [
    {
      "command": "npx tsc --noEmit -p . (apps/website)",
      "result": "rc 0, no type errors, before and after the merge of main",
      "status": "passed",
      "evidence": "local run on 7460c31dc and on f5f50a800"
    },
    {
      "command": "node --test scripts/blog-cover-accessibility.test.mjs scripts/blog-cover-versions.test.mjs",
      "result": "12 pass, 0 fail",
      "status": "passed",
      "evidence": "local node --test summary"
    },
    {
      "command": "esbuild-bundle posts.ts and load the post by slug",
      "result": "the post is first of 87 published posts; ru carries title, summary, openQuestions and body",
      "status": "passed",
      "evidence": "/tmp/blog-load.mjs output"
    },
    {
      "command": "compare the multiset of numbers in the EN and RU body, summary and openQuestions",
      "result": "identical in all three, so no number differs between the two languages",
      "status": "passed",
      "evidence": "local parity script output"
    },
    {
      "command": "render #/blog/one-outlier-twenty-three-zeros in the Vite dev server, EN and RU",
      "result": "h1, eight sections, two tables (6 and 12 rows), both terminal recordings load (200), inline code renders, RU title and headings correct",
      "status": "passed",
      "evidence": "browser pane on 127.0.0.1:5234, screenshots taken"
    },
    {
      "command": "gitleaks protect --staged",
      "result": "no leaks found in about 41 KB",
      "status": "passed",
      "evidence": "local gitleaks run before the commit"
    },
    {
      "command": "lefthook pre-commit and commit-msg (explorer and queen language contracts, english-only)",
      "result": "all three hooks pass on every commit",
      "status": "passed",
      "evidence": "hook output on commits 61912986d and 7460c31dc and on the merge f5f50a800"
    },
    {
      "command": "run ocp_mx.t27 through the site's own wasm compiler and public/play/t27run.js runSpec in node",
      "result": "49 pass, 0 fail, 0 skip; e8m0.t27 gives 17 pass, 0 fail, 10 skip (ExprCast), as lesson 31 shows",
      "status": "passed",
      "evidence": "/tmp/ocpmx-eval/run.mjs output"
    }
  ],
  "limitations": [
    "The numbers come from one block of 31 small weights and one outlier, chosen to show the flush; there is no model and no accuracy measurement",
    "ocp_mx.t27 is merged into t27 (gHashTag/t27#6828, 322cc77d7) but not yet on the site, because the world scan that vendors t27 specs has been red since 2026-10-03; the ocp_mx recording was made on the byte-identical branch commit 1a3786657",
    "The competitor comparison is twelve resources in ten rows as their pages read on 6 October 2026, not a complete survey",
    "The post has no cover image yet"
  ],
  "tags": [
    "t27",
    "OCPMX",
    "blog",
    "NumberFormats"
  ],
  "blog": {
    "title": "One outlier, 23 zeros: a course module on the OCP MX block",
    "summary": "A post on the course's MX module: one outlier in a block of 32 flushes 23 of 30 non-zero small weights to zero in MXFP4, and a spec with 49 tests catches a one-digit change in an exponent bias.",
    "cast": "term/t27c-ocp-mx/session.cast",
    "outline": [
      "The t27 course grew to 33 lessons, and its new eleventh module teaches the shared-scale number formats that AI chips use for weights.",
      "An OCP MX block holds 32 small elements and one E8M0 scale byte, which is a pure power of two with bias 127 and no mantissa.",
      "One outlier of 3.3 sets the scale for the whole block, so MXFP4 E2M1 flushes 23 of the 30 non-zero small weights to exactly zero.",
      "The same block loses 4 weights in MXFP6 E2M3 and 2 in MXINT8, and none at all in MXFP6 E3M2 or MXFP8 E4M3.",
      "Changing the E5M2 exponent bias from 15 to 16 fails exactly one of the 49 tests, so the spec notices a one-digit mistake in itself.",
      "Twelve other resources that teach MX were compared in ten rows as their pages read on 6 October 2026, and none of them counts the zeros element by element.",
      "The post lists what it does not show: one chosen block, no model, no accuracy number, and a spec that is merged into t27 but not yet on the site."
    ]
  }
}

🤖 Generated with Claude Code

[gHashTag]

Independent review of #1445 (reviewer bee, not the author) at head 9fb6fa7f8461b77efcc15e892087b8612b8a9274

Verdict: changes requested. The changes are wording only. Every headline number is right and reproduces. Some denominators, state labels and capability claims say more, or something different, than the sources do. Do not merge before #1443 merges and this PR is retargeted to main. I did not merge, push or touch the branch.

Paths below are abbreviated as follows: body.ts is apps/website/src/data/blog/bodies/one-outlier-twenty-three-zeros.ts, and index.ts is apps/website/src/data/blog/index.ts.

What I ran

#	Check	Command / source	Result
1	Only prose and data changed	git diff --stat 1fc505d2c..9fb6fa7f8 (the #1443 head is an ancestor of this head)	3 files, +330/-0: body.ts, index.ts, posts.ts (+1 import, +1 map entry). Pass
2a	Recording numbers	decoded public/term/t27c-ocp-mx/session.cast + meta.json, public/term/t27c-e8m0/	49/49 pass. Mutant E5M2 bias 15->16 gives 48 pass / 1 FAIL (bias_puts_the_min_normal_at_two_to_one_minus_bias), and git restores the spec. 32.7 s. e8m0: 18/18 + 9 invariants, 19.9 s. Pass
2b	Spec numbers	ocp_mx.t27 at t27 1a3786657 (same blob as the #6828 head 2ac62df75)	Flushed 23/4/0/2/0, scales 2^-1/2^-1/2^-3/2^1/2^-7, L1 errors match the table's decimals exactly. Pass
2c	Independent recompute	Python fractions, OCP 6.3 rule, RNE, saturating	Reproduces every flush count, scale and L1 error. Pass
2d	Mutants and compile errors	t27#6828 body	11 mutants and 20 errors, as stated. Pass
2e	Player 17/0/10 of 27	public/play/e8m0/card.png, plus runSpec run locally on e8m0.t27	17 pass, 0 fail, 10 skip, all 10 not run: ExprCast. Pass
2f	Course	specs/course/course.t27	33 lessons, 11 modules, lesson 29 "An honest scoreboard" measures a language model. Pass
3	EN = RU	body, summary, openQuestions, side by side	Same claims and numbers, and RU links go to /ru/learn/. Pass (the findings below apply to both)
4	Receipts	curl -sIL on every href, gh api state of every labelled PR/issue	All external hrefs return 200, and the OCP PDF did not 403. The 4 t27.ai/learn/... lesson receipts return 404 until #1443 deploys, which is expected for a stacked PR. Each state label is true today (see F5)
5	Competitor rows	live pages, 6 of 10 rows	Grootendorst, gfloat, Sharma and sw23 are accurate. microxcaling understates (F9). The HF cells cite a page that is not in the receipts (F10)
6	openQuestions	read against the sources	Honest overall. #3 is wrong about capability (F8), and #7's count conflicts with the body (F7)
7	Types and tests	npx tsc --noEmit -p . in apps/website (node_modules symlinked, not committed); env -u LOG_LEVEL -u log_level node --test scripts/blog-cover-accessibility.test.mjs scripts/blog-cover-versions.test.mjs	tsc rc=0. node --test: 12/12 pass. Pass
8	Work report	PR body JSON block	head_sha == 9fb6fa7f8461b77efcc15e892087b8612b8a9274. The claims match, except that they repeat F1 and F7

Findings

F1 (medium): the "of 31" denominator means two different things. BLOCK_A[14] is exactly 0. The "31 small weights" are therefore 30 non-zero weights plus one zero. The spec's "of 31" (ocp_mx.t27 L12, and the comment in small_weights_vanish_in_mxfp4) counts the 31 non-zero inputs, which are the 30 small ones plus the outlier. The counts 23/4/0/2/0 are right under either reading. The reader's denominator is not:

• body.ts:12 says "what one large weight does to the 31 small ones in its block: in MXFP4, 23 of them become zero". Suggested: "...to the other 31 weights in its block (30 non-zero and one exact zero): in MXFP4, 23 of the 30 non-zero ones become zero".
• body.ts:40 says "31 weights between -0.17 and 0.17 and one outlier". Suggested: "31 weights between -0.17 and 0.17, one of them exactly zero, and one outlier". The next sentence already says "non-zero inputs", which is good.
• body.ts:44 has the header Flushed to zero (of 31). Suggested: Flushed to zero (of 31 non-zero inputs). body.ts:89 has "counted: 23, 4, 0, 2 and 0 of 31". Suggested: "... of 31 non-zero inputs".
• RU mirrors: body.ts:142, :170, :174, :219.
• Summaries: index.ts:8 and :53 ("23 of the 31 become zero"). Suggested: "23 of the 30 non-zero small ones become zero".
• The PR body outline repeats the same wording.

F2 (low): body.ts:40 and :170 say "Every cell below is an assert in the spec". storage_is_8_25_6_25_4_25_bits_per_element (ocp_mx.t27 L603) asserts 825/825/625/425 for E4M3, INT8, E2M3 and E2M1. E3M2's 6.25 is not asserted. Either add assert centibits_per_element(K_E3M2) == 625; on t27#6828, or write "Every count, scale and error below is an assert in the spec".

F3 (low-medium, factual): body.ts:107 and :237 say "decodes all 256 codes of FP4 E2M1, FP6 E2M3 and E3M2, INT8, and OFP8...". FP4 has 16 codes and FP6 has 64. The spec's "all 256 codes" comment (L132) covers OFP8 only. The round trips walk every code of each kind up to KIND_TOP. Suggested: "It decodes every code of FP4 E2M1 (16), FP6 E2M3 and E3M2 (64 each), INT8, and OFP8 E4M3 and E5M2 (256 each), NaN and Inf included".

F4 (low): body.ts:107 and :237 say "Its 49 tests are exact vectors". Many of the 49 tests are structural properties, for example every_kind_increases_strictly, round_trip_*, and e4m3_span_is_between_17_and_18_binades. They are not input/output vectors a kernel can replay. Suggested: "Its 49 tests assert exact integers (codes, scales, counts, errors in units of 2^-24); the named code points and the five block-code tables can be run against a conversion in a kernel or in hardware".

F5 (medium): the state labels will be false on the live page. index.ts:17-20 label #1443, #1440, #1445 ("this post") and #1444 "· OPEN". All four are true today. The post can only go live after #1443 merges (which closes #1440) and #1445 merges (Closes #1444), so on the published page all four labels are wrong by construction. Drop the suffix for this post and its stack, as the terminal post did for its own #1366, or relabel them at merge time. External items (t27#6828, #6827, #6867, #1441) can keep the dated "· OPEN" convention. Merge order: the 4 t27.ai/learn/... receipts 404 until #1443 is deployed, so ship #1443 first.

F6 (low): body.ts:117 and :247 say "The recordings were made on its commit 1a3786657". Only the t27c-ocp-mx recording shows that commit. The t27c-e8m0 recording's git log shows d4417fe72. Suggested: "The ocp_mx recording was made on its commit 1a3786657".

F7 (low-medium): resource counts. body.ts:73 says "we read the pages of twelve other resources ... The table keeps ten of them", and RU body.ts:203 says the same. The table has 10 rows but 12 resources: DeepLearning.AI's two courses share a row, and so do OCP MX and arXiv:2310.10537. All 12 are in the table, and "keeps ten" suggests that two were dropped. index.ts:47 and :61 ("ten resources") and the PR body ("ten resources") inherit the problem. Suggested: "we read the pages of twelve other resources ... The table puts them in ten rows". If two other pages really were read and dropped, name them.

F8 (medium): openQuestions #3 makes a capability claim that the evaluator contradicts. index.ts:43 says "the browser cannot yet run ocp_mx.t27's tests", and :57 is the RU version. I ran the site's own evaluator on the file: loadCompiler(public/t27/t27_compiler.wasm) and then runSpec from public/play/t27run.js gave 49 pass / 0 fail / 0 skip. That matches the t27#6828 body. #1443 gives the real reason: ocp_mx.t27 is not vendored into the site yet. Suggested: "Lesson 33 is a recording, not a player: ocp_mx.t27 is not on the site yet, so the page does not run its tests (the site's evaluator passes all 49 when run on the file)". The second sentence, "because its evaluator has no casts", is also imprecise. t27run.js:166 handles cast(), @intCast, @floatCast and @as, and all 10 e8m0 skips are not run: ExprCast. Suggested: "because its evaluator does not handle the ExprCast conversion yet". body.ts:118 and :248 ("the page cannot yet run") are true of the page and need no change.

F9 (low, understated cell): the microxcaling row (body.ts:85, RU :215) says "yes: MXFP8, MXFP4, MXINT8, blocks of 32". Those are the three README examples. The library's mx/formats.py also defines fp6_e3m2, fp6_e2m3, int4 and int2. Suggested: "yes: MXFP8, MXFP6, MXFP4, MXINT8 (and INT4/INT2), blocks of 32", and the same in RU. The other cells are confirmed: mx/tests uses pytest, and "CUDA is required".

F10 (low): the HF row's sources (body.ts:81, RU :211). The receipt (index.ts:31) links only the docs page. That page says "blockwise scaling", does not give 32, and shows no benchmarks. "Blocks of 32" and the benchmarks come from the blog post it links, https://huggingface.co/blog/faster-transformers#mxfp4-quantization. Add that page as a receipt.

Rows confirmed live (no change):

• Grootendorst: no FP8, FP4 or MX; the outlier is shown, "all small values would get mapped to the same" code, and not counted.
• gfloat: OCP MX elements and block formats. test/test_microxcaling.py cross-checks against torchao.prototype.mx_formats.
• Sharma: math.ceil(math.log2(max_abs / 6.0)). The full visualizer reads "Block size 16 values" and "one outlier costs every other value precision".
• sw23: E8M0, MXINT8 and OCP FP4/FP6/FP8. tests/oracle-vectors.json plus a test.yml workflow; shareable URL state.

I also checked the scale contrast in body.ts:102. With max 3.3, ceil(log2(3.3/6)) = 0, which gives 2^0, and floor(log2 3.3) - 2 = -1, which gives 2^-1. It is correct.

Summary of the required wording fixes

F1, F5, F7 and F8 are what make the published page say something false or misleading. F3 is a plain factual slip. F2, F4, F6, F9 and F10 are small precision fixes. None of them needs a code, spec or recording change, except optionally adding the E3M2 storage assert for F2.

[gHashTag]

Review round 1 addressed in 7460c31. Every number stood; the ten phrases are fixed:

• F1: the block's 31 small weights include one exact zero, so MXFP4 flushes 23 of the 30 non-zero ones; the table header and the comparison row say "of 31 non-zero inputs", in both languages and in both summaries.
• F2: "every count, scale and error below is an assert", not "every cell".
• F3: every code of FP4 E2M1 (16), FP6 E2M3 and E3M2 (64 each), INT8, OFP8 E4M3 and E5M2 (256 each).
• F4: the 49 tests assert exact integers; the named code points and the five block-code tables are what a kernel can be checked against.
• F5: OPEN dropped from course: module 11, AI numbers -- the MX block (lessons 31-33) #1443, course: module 11 -- AI numbers, the MX block (lessons 31-33) #1440, feat(blog): one outlier, 23 zeros -- the post on course module 11 #1445, blog: one outlier, 23 zeros -- the post on course module 11, the MX block #1444; kept on course cards: four RU share cards clip their URL at 33 lessons #1441, t27#6828, t27#6827, t27#6867.
• F6: only the ocp_mx recording was made on 1a3786657.
• F7: twelve resources in ten rows.
• F8: reproduced in node with the site's own t27_compiler.wasm and t27run.js runSpec: ocp_mx.t27 49 pass / 0 fail / 0 skip; e8m0.t27 17 / 0 / 10 (ExprCast). The post now says the page does not run ocp_mx.t27 because the file is not on the site yet, not that the browser cannot.
• F9: microxcaling row lists MXFP8, MXFP6, MXFP4, MXINT8 (and INT4, INT2).
• F10: the Hugging Face MXFP4 blog post is receipt 27.

Checks on 7460c31: tsc rc 0; cover tests 12/12; EN/RU number multisets identical in body, summary and open questions; gitleaks clean; lefthook pre-commit and commit-msg pass; work report re-validated at the new head.

🤖 Generated with Claude Code

[gHashTag]

Review (reviewer bee): not merged -- the status claims went stale during review

Reviewed head f5f50a8002cf0470da59b33732b1aa22161f91f1 against main 0e1a789d7, with special attention to the ten round-1 fixes in 7460c31dc.

Blocking: t27#6828 merged while this PR was open

gHashTag/t27#6828 was merged at 2026-10-06 16:21 UTC as 322cc77d7, and the merge closed t27#6827. The post still says both are open:

• apps/website/src/data/blog/bodies/one-outlier-twenty-three-zeros.ts:117 (EN) and :247 (RU): "on an open pull request, t27#6828, not yet merged into t27 ... Until it merges, lesson 33 shows a recording of a branch."
• apps/website/src/data/blog/index.ts:8 (EN summary) and :54 (RU summary): "ocp_mx.t27 is on an open pull request, t27#6828".
• index.ts:22: the receipt for t27#6828 is tagged "OPEN, submitted, not merged".
• index.ts:23: the receipt for t27#6827 is tagged "OPEN"; it is now CLOSED.
• index.ts:43 (EN) and :57 (RU), open question: "ocp_mx.t27 is on an open pull request, t27#6828, and is not merged into t27. Lesson 33 shows a recording of that branch".
• bodies/one-outlier-twenty-three-zeros.ts:5, source comment: "(open PR t27#6828)". It is not rendered, but it is stale.

The numbers do not change. specs/numeric/ocp_mx.t27 at the merge commit 322cc77d7 has blob 44a37dd9d68d, the same blob as at 1a3786657, the commit the recording and the post's numbers come from. The fix is wording only, in both languages. For example: "merged into t27 as 322cc77d7 on 6 Oct 2026; the recording was made on 1a3786657, whose ocp_mx.t27 is byte-identical to the merged file". The two receipt tags become MERGED and CLOSED. index.ts:44/:58 ("ocp_mx.t27 is not on the site yet") is still true; the site has no copy of the file.

Checks run

1. Diff scope: PASS. 3 files, +331 lines, prose and data only: Block[] arrays, one postsIndex entry, and one import plus one map entry in posts.ts. No logic, workflow or script changes.
2. Build and tests: PASS. npx tsc --noEmit -p . exits 0. node --test scripts/blog-cover-accessibility.test.mjs scripts/blog-cover-versions.test.mjs: 12 pass, 0 fail.
3. Numbers, each checked against its source:
   • Block: PASS. BLOCK_A in ocp_mx.t27 (1a3786657, 49 tests) has 31 small weights within +-0.17, one of them exactly 0, plus the outlier 13517/4096 = 3.30005.
   • Flushed counts and errors: PASS. Flushed to zero: E2M1 23, E2M3 4, INT8 2, E3M2 0, E4M3 0. L1 errors in 2^-24 units: 42967040 / 9437184 / 5398528 / 4112384 / 1785856. I recomputed the scales, flushed counts and errors independently with Python fractions, following OCP MX v1.0 section 6.3; all match. The outlier clamps to 3.0 in E2M1.
   • Mutant E5M2 bias 15->16: PASS. public/term/t27c-ocp-mx/session.cast shows 49/49, then exactly 1 failing test after the mutation.
   • Native e8m0: PASS. 18 tests and 9 invariants in public/term/t27c-e8m0/session.cast.
   • Browser e8m0: PASS. The site evaluator (public/play/t27run.js runSpec) on e8m0.t27 gives 17 pass, 0 fail, 10 skip; all 10 skips are "not run: ExprCast". The same evaluator on ocp_mx.t27 gives 49 pass, 0 fail, 0 skip.
   • Code counts: PASS. 16; 64 each; 256 each.
   • Comparison: PASS. Ten rows dated 6 Oct 2026, and 27 receipts per language. See nit 1.
   • Course claims: PASS. 33 lessons; module 11 is lessons 31-33; lesson 29 is the language-model comparison (specs/course/course.t27:254, :289).
   • EN/RU parity: PASS. I bundled index.ts and posts.ts with esbuild and compared every number in the body and the metadata, block by block. They are identical.
   • Status claims: FAIL. trinity#1443 MERGED (0e1a789d7), t27#6867 OPEN and trinity#1441 OPEN are all true. t27#6828 and t27#6827 are stale; see Blocking.
4. Local render (vite on :5299): SKIPPED. The review machine was out of memory (12.2 of 13.3 GB swap in use). CI's Language audits step covers this route in both locales.
5. CI: NOT CONCLUSIVE. I read gh pr checks once, then ran one gh run watch capped at 30 minutes.
   • PASS: T27 work report, Code Format Validation, PR Title Format, claude-review, GitGuardian.
   • checks (Website checks, run 37493556248): every step through "Queen spec corpus, every tab" was green. It was still in "Language audits" when the 30-minute cap ran out; the job sat in the queue until 16:42 UTC.
   • Build Check and Brain Health Check were still pending.
6. Merge: NOT DONE. Blocked by the stale status claims, and checks had not finished.

Non-blocking nits

1. "Twelve other resources ... ten rows" (bodies/one-outlier-twenty-three-zeros.ts:73, RU :203; index.ts:48, RU :62). The rows link 13 pages. The count is 12 only if the Hugging Face Transformers docs page and its "MXFP4 in Transformers" blog post count as one resource. Say so, or count pages.
2. Not in this PR. In t27 specs/numeric/ocp_mx.t27, the comment above BLOCK_A (line 290) calls 13517/4096 "3.29999..."; it is 3.300049. The post's "3.3" is fine.
3. Not re-read. I did not re-read the external pages behind the comparison table in this review.

Once the status wording is fixed in both languages and checks is green, this is ready to merge. Nothing else needs to change.

🤖 Generated with Claude Code

[gHashTag]

Reviewer bee, head 0d82a57. Not merging: one stale status claim is left.

Problem (EN and RU, same paragraph): the "For people who build MX" section still treats t27#6828 as open and unmerged:

• body, section "For people who build MX", last sentence: "If one disagrees with your reading of the specification, we want that report: on t27#6828 while it is open, as an issue in gHashTag/t27 after it merges."
• ruBody, section "Тем, кто строит MX", last sentence: the same claim ("в t27#6828, пока он открыт, и отдельным issue в gHashTag/t27 после слияния").

t27#6828 was merged on 2026-10-06 as 322cc77d7, so both sentences should just point to a new issue in gHashTag/t27, e.g. "we want that report as an issue in gHashTag/t27." The status-update commit fixed the summary, the receipts, the open questions and the "What this does not show" list, but missed this paragraph.

Verified and correct (no change needed):

• headRefOid and the work report head_sha both equal 0d82a57; only the three blog files change.
• t27#6828 MERGED 2026-10-06T16:21:00Z, merge commit 322cc77d7; t27#6827 CLOSED (completed) 2026-10-06.
• specs/numeric/ocp_mx.t27 blob is 44a37dd9 at both 1a3786657 and 322cc77d7 (byte-identical).
• apps/website/public/t27/files/specs/numeric/ocp_mx.t27 is 404 on trinity main (not on the site yet).
• t27 world scan: last success 2026-10-02 18:11 (workflow_dispatch), every scheduled run from 2026-10-03 to 2026-10-06 failed, so "red since 2026-10-03" holds.
• The spec at 322cc77d7 has 49 tests, the named test bias_puts_the_min_normal_at_two_to_one_minus_bias, and asserts block_flushed 23 / 4 / 2 / 0 / 0 for E2M1 / E2M3 / INT8 / E3M2 / E4M3, matching the table.
• Referenced assets exist on main: term/t27c-ocp-mx and term/t27c-e8m0 casts, learn/one-outlier, learn/one-scale-per-block, learn/scale-byte-native, ru/learn/one-outlier, play/e8m0. t27#6867 and trinity#1441 are OPEN, as labelled.

Checks: every required check on this head is QUEUED (none ran), so mergeStateStatus is BLOCKED. That is the Actions queue or billing, not a code verdict; it would still block the merge after the text fix.

[github-actions]

X Brain Health Check

Score: 100.0/100
Status: 🟢 HEALTHY
Threshold: 80/100

X Brain is above merge threshold

[github-actions]

🧪 Brain Stress Test Results

Tests: 261/261 passed, 0 leaked
Status: 🟢 PASS
Gate: every test in zig build test-brain-stress passes and none leaks