gHashTag · gHashTag · Oct 4, 2026 · Sep 21, 2026 · Oct 4, 2026 · Oct 4, 2026
@@ -1,5 +1,5 @@
 ; GENERATED by apps/website/scripts/onboarding-from-spec.mjs (gHashTag/trinity)
-; from specs/catalog/onboarding.t27, sha256 b0b993292013f0de28d9111de91d529eeac2101f1510d5a7d46d994c37bb828a
+; from specs/catalog/onboarding.t27, sha256 9a45a5d4f2d95131b5d9253315b64dcd19f7712f793a62e19c4dcf2a0af543f4
 ; Served as https://t27.ai/agents.t27 and https://t27.ai/llms.txt -- the same bytes at both addresses.
 ; Do not edit either file: edit the spec and re-run the generator.
 
@@ -104,7 +104,8 @@ pub const REQUIRES_RUNNING_OUR_CODE : bool = false;
 pub const MEASURED_AT : str = "2026-10-04";
 pub const SPEC_COUNT : u16 = 1778;
 pub const SPEC_LINES : u32 = 490448;
-; Health as the manifest reports it. The three add up to SPEC_COUNT, and the test below says so.
+; Health as the manifest reports it, over the files that are modules. Those three plus
+; HEALTH_NOT_MODULE add up to SPEC_COUNT, and the test below says so.
 ; These moved a long way on 2026-09-21, and the cause was the instrument, not the corpus: the
 ; wasm served here until then was built from no committed source, and it accepted anything.
 ; Fed a Zig file, a Markdown file or a string literal left open, it answered typecheck.ok and
@@ -119,6 +120,18 @@ pub const SPEC_LINES : u32 = 490448;
 pub const HEALTH_OK : u16 = 1370;
 pub const HEALTH_WARN : u16 = 232;
 pub const HEALTH_FAIL : u16 = 176;
+; A file with the .t27 extension is not always a t27 module, and a document is not a broken
+; spec. The compiler's own classifier -- the rule `t27c classify` has always used, shared with
+; the wasm bridge since gHashTag/t27#4541 -- answers that question before the parser is asked,
+; and sync-t27-specs writes its verdict per file as sourceKind plus a notSource total. The
+; manifest served today was written by a wasm without the classifier, so it carries no
+; notSource and every file counts as a module: this is 0 until the next resync, and the gate
+; compares it against manifest.notSource.total rather than trusting it. When the classifier was
+; first run over the corpus on 2026-09-21 it found 121 of 1454 files that were not modules, and
+; 31 of those compiled cleanly -- whether a file is a module and whether it compiles are two
+; questions, which is why this is a fourth group and not a fourth health state.
+pub const HEALTH_NOT_MODULE : u16 = 0;
+pub const HEALTH_NOT_MODULE_NOTE : str = "not-a-module is the classifier's verdict, not a health state; the manifest's notSource.byKind splits it and notSource.byHealth shows these files compile or fail on their own";
 ; Repositories the manifest was built from, and the worlds the scan found on GitHub.
 pub const REPO_COUNT : u8 = 14;
 pub const WORLD_COUNT : u8 = 9;
@@ -128,9 +141,10 @@ pub const WORLD_COUNT : u8 = 9;
 ; Do not report the second count as the first, and do not read the total as that many broken specs.
 ; A file with the .t27 extension is not always t27: Zig source, Markdown prose, an older
 ; `spec Name { }` dialect and fixtures this repository damages on purpose all sit in the first
-; group. This file no longer prints a count per class. It carried one until 2026-09-21, hand
-; typed, adding to 154 beside a sentence that said 145 -- a breakdown nothing checked, which is
-; the one kind of number this document is not allowed to hold.
+; group until the manifest carries the classifier's verdict, after which they leave for
+; HEALTH_NOT_MODULE above. This file no longer prints a count per class. It carried one until
+; 2026-09-21, hand typed, adding to 154 beside a sentence that said 145 -- a breakdown nothing
+; checked, which is the one kind of number this document is not allowed to hold.
 pub const HEALTH_FAIL_NOTE : str = "health=fail covers both a spec that did not parse and a spec whose backend failed; the manifest's backendFailures separates them";
 pub const HEALTH_FAIL_UNPARSED : u16 = 169;
 ; What is left once the unparsed are set aside was seven specs on 2026-09-21 that parsed,
@@ -214,8 +228,14 @@ pub const ASKS_FOR_CREDENTIALS : bool = false;
 pub const ASKS_TO_ACT_ALONE : bool = false;
 
 test health_accounts_for_every_spec {
-    assert HEALTH_OK + HEALTH_WARN + HEALTH_FAIL == SPEC_COUNT;
+    // Four groups, not three. The health states are counted over modules; the rest of the
+    // corpus is HEALTH_NOT_MODULE. Leaving the fourth term out would let this assert hold
+    // over 1333 files while SPEC_COUNT says 1454 -- the sum would simply be wrong, quietly.
+    assert HEALTH_OK + HEALTH_WARN + HEALTH_FAIL + HEALTH_NOT_MODULE == SPEC_COUNT;
     assert HEALTH_FAIL > 0;
+    // Not every .t27 is a module. Zero is allowed: a manifest written before the classifier
+    // crossed the wasm boundary has no notSource, and the gate holds this to that manifest.
+    assert HEALTH_NOT_MODULE < SPEC_COUNT;
     // The qualifier is part of the count, not a footnote to it: each of these is a SUBSET of
     // fail, so a reading of HEALTH_FAIL that ignores them is reading a smaller number wrong.
     // A subset may be the whole: on 2026-09-24 every fail is unparsed, because the specs that

@@ -1,5 +1,5 @@
 ; GENERATED by apps/website/scripts/onboarding-from-spec.mjs (gHashTag/trinity)
-; from specs/catalog/onboarding.t27, sha256 b0b993292013f0de28d9111de91d529eeac2101f1510d5a7d46d994c37bb828a
+; from specs/catalog/onboarding.t27, sha256 9a45a5d4f2d95131b5d9253315b64dcd19f7712f793a62e19c4dcf2a0af543f4
 ; Served as https://t27.ai/agents.t27 and https://t27.ai/llms.txt -- the same bytes at both addresses.
 ; Do not edit either file: edit the spec and re-run the generator.
 
@@ -104,7 +104,8 @@ pub const REQUIRES_RUNNING_OUR_CODE : bool = false;
 pub const MEASURED_AT : str = "2026-10-04";
 pub const SPEC_COUNT : u16 = 1778;
 pub const SPEC_LINES : u32 = 490448;
-; Health as the manifest reports it. The three add up to SPEC_COUNT, and the test below says so.
+; Health as the manifest reports it, over the files that are modules. Those three plus
+; HEALTH_NOT_MODULE add up to SPEC_COUNT, and the test below says so.
 ; These moved a long way on 2026-09-21, and the cause was the instrument, not the corpus: the
 ; wasm served here until then was built from no committed source, and it accepted anything.
 ; Fed a Zig file, a Markdown file or a string literal left open, it answered typecheck.ok and
@@ -119,6 +120,18 @@ pub const SPEC_LINES : u32 = 490448;
 pub const HEALTH_OK : u16 = 1370;
 pub const HEALTH_WARN : u16 = 232;
 pub const HEALTH_FAIL : u16 = 176;
+; A file with the .t27 extension is not always a t27 module, and a document is not a broken
+; spec. The compiler's own classifier -- the rule `t27c classify` has always used, shared with
+; the wasm bridge since gHashTag/t27#4541 -- answers that question before the parser is asked,
+; and sync-t27-specs writes its verdict per file as sourceKind plus a notSource total. The
+; manifest served today was written by a wasm without the classifier, so it carries no
+; notSource and every file counts as a module: this is 0 until the next resync, and the gate
+; compares it against manifest.notSource.total rather than trusting it. When the classifier was
+; first run over the corpus on 2026-09-21 it found 121 of 1454 files that were not modules, and
+; 31 of those compiled cleanly -- whether a file is a module and whether it compiles are two
+; questions, which is why this is a fourth group and not a fourth health state.
+pub const HEALTH_NOT_MODULE : u16 = 0;
+pub const HEALTH_NOT_MODULE_NOTE : str = "not-a-module is the classifier's verdict, not a health state; the manifest's notSource.byKind splits it and notSource.byHealth shows these files compile or fail on their own";
 ; Repositories the manifest was built from, and the worlds the scan found on GitHub.
 pub const REPO_COUNT : u8 = 14;
 pub const WORLD_COUNT : u8 = 9;
@@ -128,9 +141,10 @@ pub const WORLD_COUNT : u8 = 9;
 ; Do not report the second count as the first, and do not read the total as that many broken specs.
 ; A file with the .t27 extension is not always t27: Zig source, Markdown prose, an older
 ; `spec Name { }` dialect and fixtures this repository damages on purpose all sit in the first
-; group. This file no longer prints a count per class. It carried one until 2026-09-21, hand
-; typed, adding to 154 beside a sentence that said 145 -- a breakdown nothing checked, which is
-; the one kind of number this document is not allowed to hold.
+; group until the manifest carries the classifier's verdict, after which they leave for
+; HEALTH_NOT_MODULE above. This file no longer prints a count per class. It carried one until
+; 2026-09-21, hand typed, adding to 154 beside a sentence that said 145 -- a breakdown nothing
+; checked, which is the one kind of number this document is not allowed to hold.
 pub const HEALTH_FAIL_NOTE : str = "health=fail covers both a spec that did not parse and a spec whose backend failed; the manifest's backendFailures separates them";
 pub const HEALTH_FAIL_UNPARSED : u16 = 169;
 ; What is left once the unparsed are set aside was seven specs on 2026-09-21 that parsed,
@@ -214,8 +228,14 @@ pub const ASKS_FOR_CREDENTIALS : bool = false;
 pub const ASKS_TO_ACT_ALONE : bool = false;
 
 test health_accounts_for_every_spec {
-    assert HEALTH_OK + HEALTH_WARN + HEALTH_FAIL == SPEC_COUNT;
+    // Four groups, not three. The health states are counted over modules; the rest of the
+    // corpus is HEALTH_NOT_MODULE. Leaving the fourth term out would let this assert hold
+    // over 1333 files while SPEC_COUNT says 1454 -- the sum would simply be wrong, quietly.
+    assert HEALTH_OK + HEALTH_WARN + HEALTH_FAIL + HEALTH_NOT_MODULE == SPEC_COUNT;
     assert HEALTH_FAIL > 0;
+    // Not every .t27 is a module. Zero is allowed: a manifest written before the classifier
+    // crossed the wasm boundary has no notSource, and the gate holds this to that manifest.
+    assert HEALTH_NOT_MODULE < SPEC_COUNT;
     // The qualifier is part of the count, not a footnote to it: each of these is a SUBSET of
     // fail, so a reading of HEALTH_FAIL that ignores them is reading a smaller number wrong.
     // A subset may be the whole: on 2026-09-24 every fail is unparsed, because the specs that

@@ -100,7 +100,7 @@
       "Flow Control - advanced flow control and backpressure Enables end-to-end flow management and congestion prevention",
       "Formula Discovery v1.0 — ULTRA ENGINE Specification",
       "Forward Pass Demo - VSA-based Neural Network Implements transformer-style forward pass using Vector Symbolic Architecture with multi-head attention, residual connections, and autoregressive generation Author: Dmitrii Vasilev",
-      "GENERATED by apps/website/scripts/onboarding-from-spec.mjs (gHashTag/trinity) from specs/catalog/onboarding.t27, sha256 ef7c4c5ad4e37abd2a64300d3409bdf724540905e6daf8f15088dfb712ee3732 Served as https://t27.ai/agents.t27 and https://t27.ai/llms.txt -- the same bytes at both addresses.",
+      "GENERATED by apps/website/scripts/onboarding-from-spec.mjs (gHashTag/trinity) from specs/catalog/onboarding.t27, sha256 9a45a5d4f2d95131b5d9253315b64dcd19f7712f793a62e19c4dcf2a0af543f4 Served as https://t27.ai/agents.t27 and https://t27.ai/llms.txt -- the same bytes at both addresses.",
       "GENERATED by tools/generate-runtime-tables.py from specs/runtimes/*.json; do not edit. Runtime tables (issue #51): for each pinned runtime, the ggml type ids it defines (block size in weights, bytes per block, the ggml_nbytes rule), the architecture names its model mapping turns into a model class and those it refuses as a main model. Sources, per runtime: 1 llama.cpp: ggml-org/llama.cpp@4364bf7232e65c34eca8d9500c5464389662de6b",
       "Generated from FORMULA_TABLE_v06.md and FORMULA_TABLE_v07.md SSOT for Trinity formula discovery",
       "GoldenFloat Cross-Language Conformance All languages must produce identical bits for identical inputs. Reference constant: GF32(φ) = 0x3FCF1BBD",

@@ -26,6 +26,11 @@ import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'
 import { dirname, join } from 'node:path'
 import { fileURLToPath } from 'node:url'
 import { CYRILLIC, SITE, checkSchema, compilerErrors, constsOf, loadCompiler, sha256, verdictOf } from './agents-from-specs.mjs'
+// The one rule for "is this entry a compilation unit?", imported rather than
+// retyped. Every derived count below is a subset of the health states, and the
+// health states are counted over modules only -- a second copy of the test here
+// is exactly how the two denominators drift apart again.
+import { isSourceEntry } from './t27-corpus.mjs'
 import { runSpecTests } from './viewport-from-spec.mjs'
 
 const WASM = 'public/t27/t27_compiler.wasm'
@@ -55,7 +60,8 @@ export const ONBOARDING_REQUIRED = {
   BACKENDS: 'arr', BACKENDS_NOTE: 'str',
   TS_SHARES_THE_JS_VALUE_LAYER: 'bool', JS_TS_DIVERGENCES: 'u16',
   MEASURED_AT: 'str', SPEC_COUNT: 'u16', SPEC_LINES: 'u32',
-  HEALTH_OK: 'u16', HEALTH_WARN: 'u16', HEALTH_FAIL: 'u16', HEALTH_FAIL_NOTE: 'str',
+  HEALTH_OK: 'u16', HEALTH_WARN: 'u16', HEALTH_FAIL: 'u16', HEALTH_NOT_MODULE: 'u16',
+  HEALTH_FAIL_NOTE: 'str', HEALTH_NOT_MODULE_NOTE: 'str',
   HEALTH_FAIL_UNPARSED: 'u16', HEALTH_FAIL_JS_ONLY: 'u16', HEALTH_PARTIAL: 'u16',
   REPO_COUNT: 'u8', WORLD_COUNT: 'u8',
   GAME: 'str', GAME_DOC: 'str', GAME_BOARD: 'str', WIN_CONDITION: 'str',
@@ -102,8 +108,22 @@ export function semanticProblems(f, file) {
     if (!f.COMPILER_EXPORTS.includes(name)) p.push(`${file}: COMPILER_EXPORTS must name ${name}`)
   }
 
-  if (f.HEALTH_OK + f.HEALTH_WARN + f.HEALTH_FAIL !== f.SPEC_COUNT) {
-    p.push(`${file}: health ${f.HEALTH_OK}+${f.HEALTH_WARN}+${f.HEALTH_FAIL} does not add up to SPEC_COUNT ${f.SPEC_COUNT}`)
+  // Four groups now, not three. The health states are counted over modules, and
+  // HEALTH_NOT_MODULE is the rest of the `.t27` files -- Markdown documents,
+  // TRI-27 assembly, fixtures damaged on purpose. Dropping the fourth term
+  // rather than adding it would have let this invariant pass while the document
+  // claimed 1454 specs and accounted for 1333.
+  const counted = f.HEALTH_OK + f.HEALTH_WARN + f.HEALTH_FAIL + f.HEALTH_NOT_MODULE
+  if (counted !== f.SPEC_COUNT) {
+    p.push(`${file}: health ${f.HEALTH_OK}+${f.HEALTH_WARN}+${f.HEALTH_FAIL}+${f.HEALTH_NOT_MODULE} = ${counted} does not add up to SPEC_COUNT ${f.SPEC_COUNT}`)
+  }
+  // Both are subsets of the failures, and a subset larger than its set is the
+  // shape of a denominator mismatch rather than a new fact.
+  if (f.HEALTH_FAIL_UNPARSED > f.HEALTH_FAIL) {
+    p.push(`${file}: HEALTH_FAIL_UNPARSED ${f.HEALTH_FAIL_UNPARSED} exceeds HEALTH_FAIL ${f.HEALTH_FAIL}; one of them is counting non-modules`)
+  }
+  if (f.HEALTH_FAIL_JS_ONLY > f.HEALTH_FAIL) {
+    p.push(`${file}: HEALTH_FAIL_JS_ONLY ${f.HEALTH_FAIL_JS_ONLY} exceeds HEALTH_FAIL ${f.HEALTH_FAIL}`)
   }
   if (!/^\d{4}-\d{2}-\d{2}$/.test(f.MEASURED_AT)) p.push(`${file}: MEASURED_AT must be an ISO date; a snapshot without its day is not a snapshot`)
 
@@ -165,6 +185,12 @@ export function measuredCounts(manifest) {
     HEALTH_OK: manifest.health.ok,
     HEALTH_WARN: manifest.health.warn,
     HEALTH_FAIL: manifest.health.fail,
+    // The `.t27` files that are not modules: Markdown documents that happen to
+    // carry the extension, TRI-27 assembly, fixtures damaged on purpose. They
+    // used to be counted as broken specs -- 81 of the 218 reds were files the
+    // parser was never meant to accept -- so they now leave the health states
+    // and are reported as their own group instead of vanishing.
+    HEALTH_NOT_MODULE: manifest.notSource?.total ?? 0,
     REPO_COUNT: manifest.repos.length,
     WORLD_COUNT: manifest.discovery.worlds.length,
     // The share of health=fail that is the declarations-only backends declining to emit, on
@@ -177,16 +203,20 @@ export function measuredCounts(manifest) {
     // matches nothing, and the constant would have been "corrected" to zero by a generator
     // that was measuring the wrong set. Compare against DECLARATIONS_ONLY instead, so the
     // next backend of this kind is a one-word edit there and not a wrong number here.
-    HEALTH_FAIL_JS_ONLY: manifest.specs.filter((s) => sameSet(s.failedBackends ?? [], DECLARATIONS_ONLY)).length,
+    HEALTH_FAIL_JS_ONLY: manifest.specs.filter((s) => isSourceEntry(s) && sameSet(s.failedBackends ?? [], DECLARATIONS_ONLY)).length,
     // The half of health=fail that HEALTH_FAIL_NOTE promises the manifest separates, actually
     // separated. Zero AST nodes means the parser produced nothing, so no backend ran: its
     // clean counters say nothing ran, not that nothing went wrong. The remainder parsed and
     // then lost a backend, which is a different fault and a different fix.
-    HEALTH_FAIL_UNPARSED: manifest.specs.filter((s) => s.health === 'fail' && !s.nodes).length,
+    //
+    // Module-only, like HEALTH_FAIL itself. Over every entry this counts 228
+    // against a HEALTH_FAIL of 156 -- a subset larger than its set, which is
+    // the arithmetic of a Markdown document being called an unparsed spec.
+    HEALTH_FAIL_UNPARSED: manifest.specs.filter((s) => isSourceEntry(s) && s.health === 'fail' && !s.nodes).length,
     // A spec whose artifact exists and says, in itself, what it could not print. Counted here
     // for the same reason as the rest: a number this document states about the corpus is
     // measured from the corpus or it is not published.
-    HEALTH_PARTIAL: manifest.specs.filter((s) => (s.partialBackends ?? []).length > 0).length,
+    HEALTH_PARTIAL: manifest.specs.filter((s) => isSourceEntry(s) && (s.partialBackends ?? []).length > 0).length,
   }
 }
 

@@ -421,7 +421,7 @@ else console.log(`  warning: no tutorial specs found -- page will open on the fi
 entries.length = 0
 entries.push(...tutorial, ...rest)
 
-const { totalLines, categories, tags, health, backendFailures, totals } = corpusAggregates(entries)
+const { totalLines, categories, tags, health, notSource, backendFailures, totals } = corpusAggregates(entries)
 
 // ---------------------------------------------------------------------------
 // Which vendored specs could someone else actually reproduce?
@@ -509,6 +509,7 @@ writeFileSync(join(OUT_DIR, 'manifest.json'), JSON.stringify({
   duplicates: duplicatePairs.sort((x, y) => x.path.localeCompare(y.path)),
   tags,
   health,
+  notSource,
   backendFailures,
   featured: FEATURED,
   totals,
@@ -530,7 +531,11 @@ writeFileSync(join(OUT_DIR, 'manifest.json'), JSON.stringify({
 
 console.log(`sync-t27-specs: ${entries.length} specs, ${Object.keys(categories).length} categories`)
 console.log(`  sources: ${sources.map((s) => s.repo).join(', ')}  (${duplicates} duplicate files skipped by content hash)`)
-console.log(`  health: ${health.ok} ok · ${health.warn} warn · ${health.fail} fail`)
+// Over modules, and the line says so. These three used to cover every vendored
+// file, which meant 81 of the failures were Markdown documents and damaged
+// fixtures being counted as broken specs -- a number that could not go down.
+console.log(`  health: ${health.ok} ok · ${health.warn} warn · ${health.fail} fail   (of ${health.ok + health.warn + health.fail} modules)`)
+console.log(`  not a module: ${notSource.total}  ${Object.entries(notSource.byKind).sort((a, b) => b[1] - a[1]).map(([k, n]) => `${k} ${n}`).join(' · ')}`)
 if (Object.keys(backendFailures).length) console.log(`  backend failures: ${JSON.stringify(backendFailures)}`)
 console.log(`  t27 @ ${shortSha}${dirty ? ' (DIRTY -- snapshot includes uncommitted spec/compiler changes)' : ''}`)
 console.log(`  wasm ${(wasmBytes / 1024).toFixed(0)} KB -> public/t27/t27_compiler.wasm  (${wasmFrom})`)