spec(hslm): nn/hslm is the model trinity-training trained (Closes #10062)

.trinity/seals/HSLM.json

@@ -1,19 +1,19 @@
 {
-  "gen_hash_c": "sha256:fd94f6962089194da7478e83cd9a015b83a28707b6cb35a9c8e629a96f502fa8",
-  "gen_hash_rust": "sha256:7f60cf6c5c5c2bf0ee07982585e981bb3e84a5444eab90063406f36940d9d76d",
-  "gen_hash_verilog": "sha256:25e87d2ab265533cab0de48e751110380236e26686ba78a711d412eb625cf1a8",
-  "gen_hash_zig": "sha256:769110ea52ada18974a5c561ea7c364bf4c7169c29b55619fb7daf16a2a91451",
+  "gen_hash_c": "sha256:5e6c04320ab12b7f1451bdfd4162ca52a3eb54c1badfb47d84d6fa8411724d19",
+  "gen_hash_rust": "sha256:435b3e5b895f99ea146078d091715ee5a71c82aa13a204090bd4b53939983b26",
+  "gen_hash_verilog": "sha256:cafcce94d7360dcf804d9ef3dc6fe5de0b4de7335c7fcc05a747ee4db0be9800",
+  "gen_hash_zig": "sha256:0b168aaa160d077bfbcb66a19487dce330289422c64da51cec4154fd5ff7799a",
   "module": "HSLM",
   "ring": 12,
-  "sealed_at": "2026-10-17T17:34:44Z",
-  "spec_hash": "sha256:1ffe306efd7cc6b092829cbcfb7645cbadde70fed20b3c3a7e3cfedd3abd27ce",
+  "sealed_at": "2026-10-17T18:02:24Z",
+  "spec_hash": "sha256:df24bc0cb83391c7131fc940940bae4804f95fde9e854ae03065ec8caff85a08",
   "spec_path": "specs/nn/hslm.t27",
   "tests": {
     "failed": 0,
     "failing": [],
     "forced": false,
-    "passed": 26,
-    "total": 26,
+    "passed": 46,
+    "total": 46,
     "vacuous": 0
   }
 }

.trinity/seals/nn_HSLM.json

@@ -7,23 +7,23 @@
     }
   ],
   "config": "gen=c,rust,verilog,zig;tests=zig",
-  "gen_hash_c": "sha256:fd94f6962089194da7478e83cd9a015b83a28707b6cb35a9c8e629a96f502fa8",
-  "gen_hash_rust": "sha256:7f60cf6c5c5c2bf0ee07982585e981bb3e84a5444eab90063406f36940d9d76d",
-  "gen_hash_verilog": "sha256:25e87d2ab265533cab0de48e751110380236e26686ba78a711d412eb625cf1a8",
-  "gen_hash_zig": "sha256:769110ea52ada18974a5c561ea7c364bf4c7169c29b55619fb7daf16a2a91451",
+  "gen_hash_c": "sha256:5e6c04320ab12b7f1451bdfd4162ca52a3eb54c1badfb47d84d6fa8411724d19",
+  "gen_hash_rust": "sha256:435b3e5b895f99ea146078d091715ee5a71c82aa13a204090bd4b53939983b26",
+  "gen_hash_verilog": "sha256:cafcce94d7360dcf804d9ef3dc6fe5de0b4de7335c7fcc05a747ee4db0be9800",
+  "gen_hash_zig": "sha256:0b168aaa160d077bfbcb66a19487dce330289422c64da51cec4154fd5ff7799a",
   "module": "HSLM",
   "ring": 12,
   "seal_schema": 2,
-  "sealed_at": "2026-10-17T17:34:44Z",
+  "sealed_at": "2026-10-17T18:02:24Z",
   "sealed_by": "t27c-bootstrap@0.5.2",
-  "spec_hash": "sha256:1ffe306efd7cc6b092829cbcfb7645cbadde70fed20b3c3a7e3cfedd3abd27ce",
+  "spec_hash": "sha256:df24bc0cb83391c7131fc940940bae4804f95fde9e854ae03065ec8caff85a08",
   "spec_path": "specs/nn/hslm.t27",
   "tests": {
     "failed": 0,
     "failing": [],
     "forced": false,
-    "passed": 26,
-    "total": 26,
+    "passed": 46,
+    "total": 46,
     "vacuous": 0
   },
   "toolchain": {

specs/igla/own_model.t27

@@ -22,9 +22,15 @@
 ;   - re-test on 2026-10-18, same t27c, after the PR that rewrote both HSLM specs (#7778):
 ;     nn/hslm 26/26 and hslm/forward_pass 14/14 pass, 0 vacuous, so 6 of the 17 compile. The
 ;     spec's backward pass is now real, each gradient checked against a finite difference.
-;     The spec is not the trained model: 6 layers at ctx 81, FFN 243-972-243 with GELU and no
+;     That spec was not the trained model: 6 layers at ctx 81, FFN 243-972-243 with GELU and no
 ;     embedding or vocab head, where trinity-training ran 3 blocks, its best run at ctx 27, and an
 ;     FFN 243-729-243 with ReLU, biases and a TWN scale.
+;   - re-test on 2026-10-18, same t27c, after the PR that rewrote nn/hslm to the trained model
+;     (#10062): nn/hslm 46/46 pass, 0 vacuous. It now has the trained shape (3 blocks, ctx 27,
+;     phi position encoding and phi-RoPE, a ReLU FFN 243-729-243 with biases and TWN alphas, the
+;     vocab-729 head and cross-entropy), its backward reaches the embedding, and every gradient is
+;     checked against a central difference; a mutant that drops one alpha fails 10 of the 46.
+;     Not in it: the per-block VSA path, f32, and the trainer's STE update (shadow weights).
 ;   - live: GET huggingface.co/api/models/playra/tern-tc-9m and .../tern-tc-9m-t27 returned 200;
 ;     .../tern-tc-100m returned 401, the same as a made-up name, so no 100M is public.
 ;     GET board-production-2dec.up.railway.app/state returned 200 and a made-up path 404; its
@@ -39,9 +45,9 @@
 
 ; THE ANSWER this map gives: HSLM was trained before, in Zig only (64 files); its best runs
 ; (PPL 2.96 at ctx 27) were wiped by EXP-002 and survive only in the ledger. In t27 the loss, three
-; optimizers and the HSLM spec's forward and backward passes pass, but that spec is not the trained
-; model (see MEASURED); ternary STE, data loader, checkpoint I/O, heap and file I/O and
-; `t27c --backend ml` do not exist yet.
+; optimizers and the HSLM spec's forward and backward passes pass, and that spec is now the trained
+; model but for its VSA path (see MEASURED); ternary STE with shadow weights, data loader,
+; checkpoint I/O, heap and file I/O and `t27c --backend ml` do not exist yet.
 
 ; Not claimed: no model quality beyond the numbers in SOURCES; no GIF or MP4, stills only.
 ; phi^2 + 1/phi^2 = 3 | TRINITY
@@ -178,7 +184,7 @@ pub const STATUS : [27]u8 = [4, 4, 4, 3, 3, 3, 3, 0, 0, 4, 4, 4, 4, 4, 4, 3, 0,
 pub const LABELS : [27]str = ["1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", "17", "18", "19", "20", "21", "22", "23", "24", "25", "26", "27"];
 
 /// Percent of each part that exists, estimated from its source (see the header). Never a status.
-pub const PCT : [27]u8 = [60, 70, 80, 80, 60, 10, 50, 0, 0, 100, 100, 20, 80, 65, 60, 40, 0, 0, 70, 15, 60, 40, 35, 0, 0, 0, 50];
+pub const PCT : [27]u8 = [60, 70, 80, 80, 60, 10, 50, 0, 0, 100, 100, 20, 80, 65, 85, 70, 0, 0, 70, 15, 60, 40, 35, 0, 0, 0, 50];
 
 /// The t27c test-report counts the PRACTICE cells quote.
 pub const ML_SPECS_MEASURED : i32 = 17;
@@ -204,9 +210,9 @@ pub const ROWS : [20]str = [
     "6 of 17 ML specs compile on the lab",
     "0 of 9 IGLA CODER specs compile",
     "1 published, 12 verified, 7 drafts",
-    "7 gaps, about 42 percent overall",
-    "Match the spec to the trained HSLM",
-    "Ternary STE, embedding and head",
+    "7 gaps, about 44 percent overall",
+    "HSLM spec: add the VSA path",
+    "Ternary STE with shadow weights",
     "Loader, checkpoint I/O, --backend ml",
     ""
 ];
@@ -313,10 +319,10 @@ test the_compile_counts_are_the_measured_ones {
 
 test percent_per_lens_is_a_plain_mean {
     assert mean_pct(0, 9) == 45;
-    assert mean_pct(9, 9) == 51;
+    assert mean_pct(9, 9) == 57;
     assert mean_pct(18, 9) == 30;
-    // the evidence slide's "about 42 percent overall"
-    assert mean_pct(0, CELL_COUNT) == 42;
+    // the evidence slide's "about 44 percent overall"
+    assert mean_pct(0, CELL_COUNT) == 44;
 }
 
 test the_camera_is_the_demo_camera {

specs/igla/own_model_board.t27

@@ -48,20 +48,20 @@ pub const LENS_CELLS : i32 = 9;
 
 pub const JOB : str = "own_model";
 pub const STAGE : str = "own_model";
-pub const MSG : str = "1 published, 12 verified, 7 drafts, 7 gaps; 42%";
+pub const MSG : str = "1 published, 12 verified, 7 drafts, 7 gaps; 44%";
 /// What MSG says once its numbers are taken out.
 pub const MSG_WORDS : str = " published,  verified,  drafts,  gaps; %";
-pub const OVERALL_PCT : u8 = 42;
-pub const PROGRESS : f64 = 0.42;
+pub const OVERALL_PCT : u8 = 44;
+pub const PROGRESS : f64 = 0.44;
 
 pub const KEY_LENS : str = "lens_pct";
 pub const KEY_CELLS : str = "cells";
 pub const KEY_STATUS : str = "status";
 pub const KEY_PCT : str = "pct";
 
-pub const LENS_PCT : [3]u8 = [45, 51, 30];
+pub const LENS_PCT : [3]u8 = [45, 57, 30];
 pub const CELL_STATUS : [27]u8 = [4, 4, 4, 3, 3, 3, 3, 0, 0, 4, 4, 4, 4, 4, 4, 3, 0, 0, 4, 4, 5, 4, 3, 0, 0, 0, 3];
-pub const CELL_PCT : [27]u8 = [60, 70, 80, 80, 60, 10, 50, 0, 0, 100, 100, 20, 80, 65, 60, 40, 0, 0, 70, 15, 60, 40, 35, 0, 0, 0, 50];
+pub const CELL_PCT : [27]u8 = [60, 70, 80, 80, 60, 10, 50, 0, 0, 100, 100, 20, 80, 65, 85, 70, 0, 0, 70, 15, 60, 40, 35, 0, 0, 0, 50];
 
 /// The size of the compact metrics object, in bytes, computed from the keys and values above.
 pub const METRICS_BYTES : i32 = 622;
@@ -273,19 +273,19 @@ test the_message_is_the_tally {
     const n : i32 = @intCast(MSG.len);
     assert n <= MSG_MAX;
     // one number wrong, a word changed, or a number added: each refused
-    assert msg_matches("1 published, 11 verified, 7 drafts, 7 gaps; 42%", STATUS, overall) == false;
-    assert msg_matches("1 published, 12 verified, 7 drafts, 7 holes; 42%", STATUS, overall) == false;
-    assert msg_matches("1 published, 12 verified, 7 drafts, 7 gaps; 42% of 100", STATUS, overall) == false;
-    assert msg_matches(MSG, STATUS, 43) == false;
+    assert msg_matches("1 published, 11 verified, 7 drafts, 7 gaps; 44%", STATUS, overall) == false;
+    assert msg_matches("1 published, 12 verified, 7 drafts, 7 holes; 44%", STATUS, overall) == false;
+    assert msg_matches("1 published, 12 verified, 7 drafts, 7 gaps; 44% of 100", STATUS, overall) == false;
+    assert msg_matches(MSG, STATUS, 45) == false;
     assert nth_number(MSG, 1) == 12;
-    assert nth_number(MSG, 4) == 42;
+    assert nth_number(MSG, 4) == 44;
     assert nth_number(MSG, 5) == -1;
 }
 
 test progress_is_the_overall_percent_over_a_hundred {
     assert progress_matches(PROGRESS, mean_pct(0, CELL_COUNT));
     // a bar one point off, or past the end of the board's 0..1, is refused
-    assert progress_matches(0.43, mean_pct(0, CELL_COUNT)) == false;
+    assert progress_matches(0.45, mean_pct(0, CELL_COUNT)) == false;
     assert progress_matches(1.39, 139) == false;
     assert progress_matches(-0.01, -1) == false;
     assert JOB.len > 0;