gHashTag · Sep 21, 2026 · Sep 21, 2026
@@ -1,5 +1,5 @@
 ; GENERATED by apps/website/scripts/onboarding-from-spec.mjs (gHashTag/trinity)
-; from specs/catalog/onboarding.t27, sha256 b2a2376e1a81aa892ffd90b0a329532097d5f7569e86d51e2ca39c3192d0de9f
+; from specs/catalog/onboarding.t27, sha256 d70522d6b66222b1e81e13f3be901162916aaa09a0d84de3c1138af878339758
 ; 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.
 
@@ -102,37 +102,44 @@ pub const REQUIRES_RUNNING_OUR_CODE : bool = false;
 
 ; --- The corpus on the day this was written --------------------------------------------------
 pub const MEASURED_AT : str = "2026-09-21";
-pub const SPEC_COUNT : u16 = 1414;
-pub const SPEC_LINES : u32 = 372417;
+pub const SPEC_COUNT : u16 = 1419;
+pub const SPEC_LINES : u32 = 373151;
 ; Health as the manifest reports it. The three 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
 ; emitted C, Rust, Verilog and Zig. The binary now served is built from bootstrap/src, and it
 ; reads what is actually in the file. Nothing in the corpus got worse; it got measured.
-pub const HEALTH_OK : u16 = 976;
-pub const HEALTH_WARN : u16 = 100;
-pub const HEALTH_FAIL : u16 = 338;
+pub const HEALTH_OK : u16 = 1017;
+pub const HEALTH_WARN : u16 = 184;
+pub const HEALTH_FAIL : u16 = 218;
 ; Repositories the manifest was built from, and the worlds the scan found on GitHub.
 pub const REPO_COUNT : u8 = 10;
 pub const WORLD_COUNT : u8 = 5;
-; "fail" is several different things wearing one label: a spec that never parsed (no backend
-; ran, so its clean counters mean nothing ran, not that nothing went wrong), and a spec that
-; parsed and then lost a backend. Do not report the second count as the first, and do not read
-; the total as "338 broken specs" -- measured on 2026-09-21, 145 of the 338 are not t27 at all:
-; 68 are Zig source, 49 are an older `spec Name { }` dialect, 14 are Markdown prose and 23 are
-; fixtures this repository damages on purpose so the compiler has something to reject. Counting
-; them as specs is what the extension does, not what the contents are.
+; "fail" is two different things wearing one label, and the split is measured rather than
+; described: a spec that NEVER PARSED, so no backend ran and its clean counters mean nothing
+; ran rather than nothing went wrong -- and a spec that parsed and then lost a backend.
+; Do not report the second count as the first, and do not read the total as "218 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.
 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";
-; The largest single class: js and ts are declarations-only backends, so a const whose
-; initialiser is an expression rather than a literal has nothing for them to emit and they
-; say so by name. Those files parse, typecheck and emit five other languages.
-; The name of this constant is now half-true and kept anyway: nothing in the corpus fails
-; on js alone or on ts alone, so the class it counts is the same 122 files it always was --
-; renaming it would move a number that did not move. What CHANGED is that the manifest now
-; reports each of those files twice, once per backend, which is why backendFailures reads
-; js=335 ts=335 and not js=335 ts=0.
-pub const HEALTH_FAIL_JS_ONLY : u16 = 122;
+pub const HEALTH_FAIL_UNPARSED : u16 = 211;
+; What is left once the unparsed are set aside: seven specs that parse, typecheck, and emit
+; six of seven languages. That is the honest size of "broken and readable".
+; The largest class used to be a different one. js and ts are declarations-only backends, and
+; a const whose initialiser they had no spelling for failed the WHOLE module -- costing every
+; declaration that was fine to refuse the one that was not, on files the other five backends
+; accepted. They now announce that declaration by name, in the artifact, and emit the rest.
+; The constant below is what that change did to this number, and it is checked against the
+; manifest; the size of the class before the change is not, so it is not printed here.
+pub const HEALTH_FAIL_JS_ONLY : u16 = 0;
+; Announced, not dropped. A spec here emits a real module that says inside itself what it left
+; out, as data -- `__NOT_EMITTED__`, a list a tool can read without parsing comments. That is
+; neither whole nor broken, so it is counted apart and coloured a warning, never a pass.
+pub const HEALTH_PARTIAL : u16 = 71;
 
 ; --- The game --------------------------------------------------------------------------------
 pub const GAME : str = "Queen Hive";
@@ -184,10 +191,17 @@ pub const ASKS_TO_ACT_ALONE : bool = false;
 test health_accounts_for_every_spec {
     assert HEALTH_OK + HEALTH_WARN + HEALTH_FAIL == SPEC_COUNT;
     assert HEALTH_FAIL > 0;
-    // The qualifier is part of the count, not a footnote to it: js-only failures are a subset
-    // of fail, so a reading of HEALTH_FAIL that ignores them is reading a smaller number wrong.
+    // 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.
     assert HEALTH_FAIL_JS_ONLY < HEALTH_FAIL;
-    assert HEALTH_FAIL_JS_ONLY > 0;
+    assert HEALTH_FAIL_UNPARSED < HEALTH_FAIL;
+    // The subset that is worth stating on its own: 209 of the 216 never reached a backend at
+    // all. An earlier version of this test demanded HEALTH_FAIL_JS_ONLY > 0, which was true
+    // the day it was written and is a lie the moment the class it counts is fixed. A test
+    // that forbids its own subject from reaching zero is a test against the work.
+    assert HEALTH_FAIL_UNPARSED > 0;
+    // A partial artifact is a warning, never a pass, so it cannot outnumber the warnings.
+    assert HEALTH_PARTIAL <= HEALTH_WARN;
 }
 
 test the_cycle_is_five_steps_in_order {

@@ -1,5 +1,5 @@
 ; GENERATED by apps/website/scripts/onboarding-from-spec.mjs (gHashTag/trinity)
-; from specs/catalog/onboarding.t27, sha256 b2a2376e1a81aa892ffd90b0a329532097d5f7569e86d51e2ca39c3192d0de9f
+; from specs/catalog/onboarding.t27, sha256 d70522d6b66222b1e81e13f3be901162916aaa09a0d84de3c1138af878339758
 ; 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.
 
@@ -102,37 +102,44 @@ pub const REQUIRES_RUNNING_OUR_CODE : bool = false;
 
 ; --- The corpus on the day this was written --------------------------------------------------
 pub const MEASURED_AT : str = "2026-09-21";
-pub const SPEC_COUNT : u16 = 1414;
-pub const SPEC_LINES : u32 = 372417;
+pub const SPEC_COUNT : u16 = 1419;
+pub const SPEC_LINES : u32 = 373151;
 ; Health as the manifest reports it. The three 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
 ; emitted C, Rust, Verilog and Zig. The binary now served is built from bootstrap/src, and it
 ; reads what is actually in the file. Nothing in the corpus got worse; it got measured.
-pub const HEALTH_OK : u16 = 976;
-pub const HEALTH_WARN : u16 = 100;
-pub const HEALTH_FAIL : u16 = 338;
+pub const HEALTH_OK : u16 = 1017;
+pub const HEALTH_WARN : u16 = 184;
+pub const HEALTH_FAIL : u16 = 218;
 ; Repositories the manifest was built from, and the worlds the scan found on GitHub.
 pub const REPO_COUNT : u8 = 10;
 pub const WORLD_COUNT : u8 = 5;
-; "fail" is several different things wearing one label: a spec that never parsed (no backend
-; ran, so its clean counters mean nothing ran, not that nothing went wrong), and a spec that
-; parsed and then lost a backend. Do not report the second count as the first, and do not read
-; the total as "338 broken specs" -- measured on 2026-09-21, 145 of the 338 are not t27 at all:
-; 68 are Zig source, 49 are an older `spec Name { }` dialect, 14 are Markdown prose and 23 are
-; fixtures this repository damages on purpose so the compiler has something to reject. Counting
-; them as specs is what the extension does, not what the contents are.
+; "fail" is two different things wearing one label, and the split is measured rather than
+; described: a spec that NEVER PARSED, so no backend ran and its clean counters mean nothing
+; ran rather than nothing went wrong -- and a spec that parsed and then lost a backend.
+; Do not report the second count as the first, and do not read the total as "218 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.
 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";
-; The largest single class: js and ts are declarations-only backends, so a const whose
-; initialiser is an expression rather than a literal has nothing for them to emit and they
-; say so by name. Those files parse, typecheck and emit five other languages.
-; The name of this constant is now half-true and kept anyway: nothing in the corpus fails
-; on js alone or on ts alone, so the class it counts is the same 122 files it always was --
-; renaming it would move a number that did not move. What CHANGED is that the manifest now
-; reports each of those files twice, once per backend, which is why backendFailures reads
-; js=335 ts=335 and not js=335 ts=0.
-pub const HEALTH_FAIL_JS_ONLY : u16 = 122;
+pub const HEALTH_FAIL_UNPARSED : u16 = 211;
+; What is left once the unparsed are set aside: seven specs that parse, typecheck, and emit
+; six of seven languages. That is the honest size of "broken and readable".
+; The largest class used to be a different one. js and ts are declarations-only backends, and
+; a const whose initialiser they had no spelling for failed the WHOLE module -- costing every
+; declaration that was fine to refuse the one that was not, on files the other five backends
+; accepted. They now announce that declaration by name, in the artifact, and emit the rest.
+; The constant below is what that change did to this number, and it is checked against the
+; manifest; the size of the class before the change is not, so it is not printed here.
+pub const HEALTH_FAIL_JS_ONLY : u16 = 0;
+; Announced, not dropped. A spec here emits a real module that says inside itself what it left
+; out, as data -- `__NOT_EMITTED__`, a list a tool can read without parsing comments. That is
+; neither whole nor broken, so it is counted apart and coloured a warning, never a pass.
+pub const HEALTH_PARTIAL : u16 = 71;
 
 ; --- The game --------------------------------------------------------------------------------
 pub const GAME : str = "Queen Hive";
@@ -184,10 +191,17 @@ pub const ASKS_TO_ACT_ALONE : bool = false;
 test health_accounts_for_every_spec {
     assert HEALTH_OK + HEALTH_WARN + HEALTH_FAIL == SPEC_COUNT;
     assert HEALTH_FAIL > 0;
-    // The qualifier is part of the count, not a footnote to it: js-only failures are a subset
-    // of fail, so a reading of HEALTH_FAIL that ignores them is reading a smaller number wrong.
+    // 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.
     assert HEALTH_FAIL_JS_ONLY < HEALTH_FAIL;
-    assert HEALTH_FAIL_JS_ONLY > 0;
+    assert HEALTH_FAIL_UNPARSED < HEALTH_FAIL;
+    // The subset that is worth stating on its own: 209 of the 216 never reached a backend at
+    // all. An earlier version of this test demanded HEALTH_FAIL_JS_ONLY > 0, which was true
+    // the day it was written and is a lie the moment the class it counts is fixed. A test
+    // that forbids its own subject from reaching zero is a test against the work.
+    assert HEALTH_FAIL_UNPARSED > 0;
+    // A partial artifact is a warning, never a pass, so it cannot outnumber the warnings.
+    assert HEALTH_PARTIAL <= HEALTH_WARN;
 }
 
 test the_cycle_is_five_steps_in_order {

@@ -79,13 +79,13 @@ module ForwardPass {
     // Single-head attention: query = bind(Q, K), score similarity with V
     // Returns best-matching role vector as value
     // Complexity: O(n) where n = ROLE_DIM (VSA bind + similarity)
-    pub fn singleHeadAttention(positioned: []RoleVector, q_role: &RoleVector, k_role: &RoleVector, v_role: &RoleVector) -> RoleVector;
+    pub fn singleHeadAttention(positioned: []RoleVector, q_role: &RoleVector, k_role: &RoleVector, v_role: &RoleVector) -> RoleVector { }
 
     // multiHeadAttention(context: []HyperVector, roles: []RoleVector, num_heads: usize) -> ForwardOutput
     // Multi-head attention: 3 independent heads merged via bundle3
     // Computes separate Q/K/V attention heads, merges via bundle3
     // Complexity: O(n * h) where n = ROLE_DIM, h = HEAD_COUNT
-    pub fn multiHeadAttention(context: []HyperVector, roles: []RoleVector, num_heads: usize) -> ForwardOutput;
+    pub fn multiHeadAttention(context: []HyperVector, roles: []RoleVector, num_heads: usize) -> ForwardOutput { }
 
     // ========================================================================
     // 6. Forward Pass Functions
@@ -95,13 +95,13 @@ module ForwardPass {
     // v2.29 single-head forward pass
     // Uses Q/K/V attention to compute value = bind(Q/K/V)
     // Complexity: O(n) where n = ROLE_DIM (VSA operations)
-    pub fn forwardPass(context: []HyperVector, roles: []RoleVector, target: &HyperVector) -> ForwardOutput;
+    pub fn forwardPass(context: []HyperVector, roles: []RoleVector, target: &HyperVector) -> ForwardOutput { }
 
     // forwardPassMultiHead(context: []HyperVector, roles: []RoleVector) -> ForwardOutput
     // v2.30 multi-head forward pass (3 heads merged)
     // Uses independent Q/K/V attention heads, merges via bundle3
     // Complexity: O(n * h) where n = ROLE_DIM, h = HEAD_COUNT
-    pub fn forwardPassMultiHead(context: []HyperVector, roles: []RoleVector) -> ForwardOutput;
+    pub fn forwardPassMultiHead(context: []HyperVector, roles: []RoleVector) -> ForwardOutput { }
 
     // ========================================================================
     // 7. Training Functions
@@ -112,13 +112,13 @@ module ForwardPass {
     // Uses bind-based targeted correction
     // Returns loss (1 - similarity) for this sample
     // Complexity: O(n * h) where n = ROLE_DIM, h = 2 (FF roles)
-    pub fn resonatorTrainStep(context: []HyperVector, target: &HyperVector, roles: []RoleVector, dim: usize) -> f64;
+    pub fn resonatorTrainStep(context: []HyperVector, target: &HyperVector, roles: []RoleVector, dim: usize) -> f64 { }
 
     // refineDirectRole(corpus: []const u8, dim: usize, initial_role: &RoleVector, num_passes: usize) -> RoleVector
     // Iteratively refine direct role using error measurement
     // Measures error, computes sparse correction, blends with sparsified annealing
     // Complexity: O(p * n * m) where p = num_passes, n = dim, m = CORPUS_SIZE
-    pub fn refineDirectRole(corpus: []const u8, dim: usize, initial_role: &RoleVector, num_passes: usize) -> RoleVector;
+    pub fn refineDirectRole(corpus: []const u8, dim: usize, initial_role: &RoleVector, num_passes: usize) -> RoleVector { }
 
     // ========================================================================
     // 8. Context Functions
@@ -128,7 +128,7 @@ module ForwardPass {
     // Permute and bundle 8 context vectors into 1 summary HV
     // Uses positional permutation then sequential bundling
     // Complexity: O(n) where n = ROLE_DIM (permute + bundle operations)
-    pub fn summarizeContext(context: []HyperVector) -> HyperVector;
+    pub fn summarizeContext(context: []HyperVector) -> HyperVector { }
 
     // ========================================================================
     // 9. Direct Role Functions
@@ -138,13 +138,13 @@ module ForwardPass {
     // Pre-compute ideal direct role from corpus
     // Uses charToHV encoding and bundle operations
     // Complexity: O(c * n) where c = context_size, n = dim
-    pub fn computeDirectRole(corpus: []const u8, dim: usize, offsets: []const usize, context_size: usize) -> RoleVector;
+    pub fn computeDirectRole(corpus: []const u8, dim: usize, offsets: []const usize, context_size: usize) -> RoleVector { }
 
     // directDecode(context: []HyperVector, role: &RoleVector, dim: usize) -> []u8
     // Decode output HV to character sequence
     // Uses learned role to predict next character
     // Complexity: O(n) where n = output length
-    pub fn directDecode(context: []HyperVector, role: &RoleVector, dim: usize) -> []u8;
+    pub fn directDecode(context: []HyperVector, role: &RoleVector, dim: usize) -> []u8 { }
 
     // ========================================================================
     // 10. Generation Functions
@@ -154,7 +154,7 @@ module ForwardPass {
     // Autoregressive generation using pre-computed direct role
     // Shifts context, predicts tokens using directDecode
     // Complexity: O(t * n) where t = max_tokens, n = ROLE_DIM
-    pub fn generateWithDirectRole(initial_context: []HyperVector, role: &RoleVector, output_buf: []u8, max_tokens: usize) -> usize;
+    pub fn generateWithDirectRole(initial_context: []HyperVector, role: &RoleVector, output_buf: []u8, max_tokens: usize) -> usize { }
 
     // ========================================================================
     // 11. Hebbian Functions
@@ -164,13 +164,13 @@ module ForwardPass {
     // Build character-pair association matrix from corpus
     // Returns counts[a][b] = frequency of char b following char a
     // Complexity: O(c * m) where c = corpus.len, m = HEBBIAN_CHARS^2
-    pub fn buildHebbianCounts(corpus: []const u8) -> [HEBBIAN_CHARS][HEBBIAN_CHARS]u16;
+    pub fn buildHebbianCounts(corpus: []const u8) -> [HEBBIAN_CHARS][HEBBIAN_CHARS]u16 { }
 
     // hebbianLookup(counts: *const [HEBBIAN_CHARS][HEBBIAN_CHARS]u16, char_idx: usize, dim: usize) -> HyperVector
     // Look up character successor using Hebbian matrix
     // Bundles successors weighted by count, proportional to frequency
     // Complexity: O(n * m) where n = HEBBIAN_CHARS, m = dim
-    pub fn hebbianLookup(counts: *const [HEBBIAN_CHARS][HEBBIAN_CHARS]u16, char_idx: usize, dim: usize) -> HyperVector;
+    pub fn hebbianLookup(counts: *const [HEBBIAN_CHARS][HEBBIAN_CHARS]u16, char_idx: usize, dim: usize) -> HyperVector { }
 
     // ========================================================================
     // TDD - Tests

@@ -313,6 +313,13 @@ test bench_proxy_compute_pass_at_1_none_match
     when p = compute_pass_at_1(template, problems)
     then p == 0.0
 
+test run_full_baseline_inner_singleton
+    given templates = ["adder"]
+    given problems = verilog_eval_problems()
+    when scores = run_full_baseline_inner(templates, problems, 0)
+    then len(scores) == 1 && scores[0] >= 0.0 && scores[0] <= 1.0
+
+
 // ============================================================================
 // 5. Invariants
 // ============================================================================