igla cell 9: tern-tc-500m grounded, then reviewed (Closes #10420)

.trinity/seals/igla_own_model.json

@@ -1,5 +1,5 @@
 {
-  "built_by": "t27c-bootstrap@0.5.2+4ab200073",
+  "built_by": "t27c-bootstrap@0.5.2+fd332dfd5",
   "closure": [
     {
       "spec": "specs/t27f/reel.t27",
@@ -15,16 +15,16 @@
     }
   ],
   "config": "gen=c,rust,verilog,zig;tests=zig",
-  "gen_hash_c": "sha256:7749a3a3eee34f915a1dd07db7b752032481c5abc90a670209c993416aa1773f",
-  "gen_hash_rust": "sha256:88adf9c918ec50be845725e9cd197b18fb3701004e43e23b225aca912631146b",
-  "gen_hash_verilog": "sha256:3081084523bf60707644f1a5c073b5d3a5571319408889af08a7b1973ac3c65f",
-  "gen_hash_zig": "sha256:613a50a9a748b999c289d93274de064d0f30d425a7ec256e00b56fc3f5959811",
+  "gen_hash_c": "sha256:34f442903c5b74b1ddc28b310d9782b0f61c45098e4ff931771176583341057d",
+  "gen_hash_rust": "sha256:a0e037fe6d5dd9b82586a2bf2a80c27ff50f6c9ba7b9babe60e9b52349efaa5c",
+  "gen_hash_verilog": "sha256:64cba74ad08af9252262494d75e1ec94d598f81d5fa4b06bc3bff797d5ca5ad8",
+  "gen_hash_zig": "sha256:a180869d634e33aeca7c64cde565c2115d43ded528f7613a9ce9ad084ab787be",
   "module": "own_model",
   "ring": 12,
   "seal_schema": 2,
-  "sealed_at": "2026-10-18T09:41:47Z",
+  "sealed_at": "2026-10-18T13:33:41Z",
   "sealed_by": "t27c-bootstrap@0.5.2",
-  "spec_hash": "sha256:f7471c56c57c69f08cc607d930968459f470372c17ed6eb75368b6c82a6245c2",
+  "spec_hash": "sha256:2c3daec9e62ce9123e19cad921e57af75beb25a1870c234230ae5505a7d7ac79",
   "spec_path": "specs/igla/own_model.t27",
   "tests": {
     "failed": 0,

.trinity/seals/igla_own_model_board.json

@@ -1,9 +1,9 @@
 {
-  "built_by": "t27c-bootstrap@0.5.2+4ab200073",
+  "built_by": "t27c-bootstrap@0.5.2+fd332dfd5",
   "closure": [
     {
       "spec": "specs/igla/own_model.t27",
-      "spec_hash": "sha256:f7471c56c57c69f08cc607d930968459f470372c17ed6eb75368b6c82a6245c2"
+      "spec_hash": "sha256:2c3daec9e62ce9123e19cad921e57af75beb25a1870c234230ae5505a7d7ac79"
     },
     {
       "spec": "specs/t27f/reel.t27",
@@ -19,16 +19,16 @@
     }
   ],
   "config": "gen=c,rust,verilog,zig;tests=zig",
-  "gen_hash_c": "sha256:9c6f443ba6f6d8214d1e979c90105be99ecba919716e547d0e896d89ed0b75e1",
-  "gen_hash_rust": "sha256:94f42c2726ffcc3015a5d346b36565f343ce57600604b4931c7430c693cb1158",
-  "gen_hash_verilog": "sha256:cb284309b32ad27988df63ab1e7696c2ae0b549978afb80a6f83b8c8d1a8972e",
-  "gen_hash_zig": "sha256:0969cbc653c06569242e3c9a58023b001c5c1a143c276a5e917e12d79924b189",
+  "gen_hash_c": "sha256:a0125cda52bdb9a87de3b170cc991ad71fc627caf8c60e4bc41a2f4a9a3a86e8",
+  "gen_hash_rust": "sha256:dc61033d126c3f954f8b6ae3f33c33f773edd8f5eda07530cf3576e56d39dd73",
+  "gen_hash_verilog": "sha256:0d56c85728676e996f03f9636d7492c02c2901cd034348565c28a657aa86731f",
+  "gen_hash_zig": "sha256:c742951d611cb908f642a7262478d7219717d0b3b3284c123aebd0d4cd63bff5",
   "module": "own_model_board",
   "ring": 12,
   "seal_schema": 2,
-  "sealed_at": "2026-10-18T09:41:48Z",
+  "sealed_at": "2026-10-18T13:33:45Z",
   "sealed_by": "t27c-bootstrap@0.5.2",
-  "spec_hash": "sha256:ba4e797ce2ea44fe5ea6265fb4c54bbbc8117cb914de18f9ff90bf0890109cbb",
+  "spec_hash": "sha256:69e4e67d00d0a17aed8d5e9ed45d7c745c970d069388db37f7835ff39179adae",
   "spec_path": "specs/igla/own_model_board.t27",
   "tests": {
     "failed": 0,

gen/json/igla/own_model_board.json

@@ -1 +1 @@
-{"job":"own_model","stage":"own_model","msg":"3 published, 22 verified, 1 drafts, 1 gaps; 61%","progress":0.61,"metrics":{"lens_pct":[65,70,48],"cells":[{"status":4,"pct":60},{"status":4,"pct":70},{"status":4,"pct":80},{"status":4,"pct":80},{"status":4,"pct":60},{"status":4,"pct":100},{"status":5,"pct":50},{"status":5,"pct":90},{"status":0,"pct":0},{"status":4,"pct":100},{"status":4,"pct":100},{"status":4,"pct":20},{"status":4,"pct":80},{"status":4,"pct":65},{"status":4,"pct":95},{"status":4,"pct":75},{"status":4,"pct":55},{"status":4,"pct":40},{"status":4,"pct":70},{"status":4,"pct":15},{"status":5,"pct":60},{"status":4,"pct":40},{"status":4,"pct":100},{"status":4,"pct":60},{"status":4,"pct":30},{"status":3,"pct":10},{"status":4,"pct":50}]}}
+{"job":"own_model","stage":"own_model","msg":"4 published, 22 verified, 1 drafts, 0 gaps; 62%","progress":0.62,"metrics":{"lens_pct":[68,70,48],"cells":[{"status":4,"pct":60},{"status":4,"pct":70},{"status":4,"pct":80},{"status":4,"pct":80},{"status":4,"pct":60},{"status":4,"pct":100},{"status":5,"pct":50},{"status":5,"pct":90},{"status":5,"pct":30},{"status":4,"pct":100},{"status":4,"pct":100},{"status":4,"pct":20},{"status":4,"pct":80},{"status":4,"pct":65},{"status":4,"pct":95},{"status":4,"pct":75},{"status":4,"pct":55},{"status":4,"pct":40},{"status":4,"pct":70},{"status":4,"pct":15},{"status":5,"pct":60},{"status":4,"pct":40},{"status":4,"pct":100},{"status":4,"pct":60},{"status":4,"pct":30},{"status":3,"pct":10},{"status":4,"pct":50}]}}

specs/igla/own_model.t27

@@ -11,6 +11,7 @@
 ;     experiment ledger) and data/wave9/worker-2/hslm_step_10000.bin (the one HSLM checkpoint left).
 ;   - gHashTag/trinity-training at TRAINING_COMMIT: src/hslm, 64 Zig files, the HSLM trainer.
 ;   - gHashTag/trinity-fpga at FPGA_COMMIT: 8 tokens on the AX7203, 7682304/7682304 receipts.
+;   - gHashTag/igla-coder-gpu at CARD_500M_COMMIT: specs/hf/model_card_500m.t27 (cell 9).
 ;   - gHashTag/t27 at SOURCE_COMMIT: the ML specs named per cell.
 
 ; MEASURED on 2026-10-17:
@@ -90,8 +91,25 @@
 ;     620c80b642ff87f06f945f6dbedf122e69fd524e93b7b20d1fd9bfcfd3d08300, both run summaries pinned
 ;     by sha256 in its model_card.t27 (14/14 on the lab). Validation bits per byte 0.4722 against
 ;     0.4217 for the fp twin of the same recipe: the ternary model is the worse of the two.
+;   - trained on 2026-10-18, cell 9: tern-tc-500m, a 512,881,920-parameter ternary model of the
+;     same family (n_layer 24, d_model 1280, GQA 20 to 4 heads, d_ff 4096), on the owner's yes to
+;     0.5B parameters x 1.2B tokens. RunPod pod pkv4k8xhs8jo2b (RTX PRO 6000 Blackwell, EU-RO-1):
+;     2288 steps, 1,199,570,944 tokens, 5.235 h, 69.4k tokens/s at 48 percent MFU, 13.65 USD from the
+;     balance (LEDGER.md). Validation bits per byte 0.5307 on 2M held-out tokens per language, worse
+;     than tern-tc-100m's 0.4722 after 10B tokens: more parameters did not make up for 8.3 times
+;     fewer tokens. final.pt is 2,051,594,699 bytes, sha256
+;     e6925b7c30e3a6c680fdb3591feeae7a1a99a19fe194405daf51ddbb1d8c15bc. Every number is in
+;     specs/hf/model_card_500m.t27 of igla-coder-gpu at CARD_500M_COMMIT, 12/12 tests and 0 vacuous on
+;     the lab; a copy with the byte overhead and the bits per byte each one off fails 2 of the 12. Live
+;     since 2026-10-18: playra/tern-tc-500m at Hugging Face commit HF_500M_COMMIT, whose final.pt
+;     carries LFS sha256 e6925b7c...1d8c15bc and whose eight other files match the release machine
+;     byte for byte. Its evidence went to review in PR REVIEW_500M_PR: review REVIEW_500M_ID is the
+;     reviewer bee on head 88f47f32d, whose model verdict was approve; the bee posted it as
+;     NEEDS_PERSON, not APPROVE, because that head changes seal and gen files, and the PR merged over
+;     that hold under the owner's standing instruction, as #10275 did. Live and reviewed, so the cell
+;     is published at 30 percent: one run at the low end of 0.5-1.5B, under-trained.
 ;   - live: GET huggingface.co/api/models/playra/tern-tc-9m, .../tern-tc-9m-t27 and
-;     .../tern-tc-100m returned 200 (tern-tc-100m answered 401 before 2026-10-18).
+;     .../tern-tc-100m and .../tern-tc-500m returned 200 (tern-tc-100m answered 401 before 2026-10-18).
 ;     GET board-production-2dec.up.railway.app/state returned 200 and a made-up path 404; its
 ;     live jobs include own_model since 2026-10-18 06:33 UTC, the event t27c gen-json makes from
 ;     specs/igla/own_model_board.t27 (it was stale since 2026-10-05 before that).
@@ -113,7 +131,7 @@
 ;     compute from DDR3, so the cell is a draft at 10 percent, a DELIVER step waiting for its review.
 ;     Cell 26 has no review: its PR, #10248, merged without one, so it stays a draft.
 ; The engine asks an engineering LEARN step for a source link, a PRACTICE step for a passing test
-; and a DELIVER step for a review. So: 3 published, 22 verified, 1 draft, 1 gap.
+; and a DELIVER step for a review. So: 4 published, 22 verified, 1 draft, 0 gaps.
 
 ; PERCENT is a second, separate number per cell: an estimate of how much of that part exists, with
 ; its basis in SOURCES. It never feeds a status. The lens and overall figures are plain means, every
@@ -126,7 +144,7 @@
 ; and checkpoint layout now pass too, on a model slice and on bytes; heap and file I/O are t27
 ; (ml/arena_heap, ml/file_io_contract, ml/host_io_model) and `t27c --backend ml` lowers them to a
 ; binary that writes a real file, and to a training binary for one head slice
-; (ml/train_slice_model); the STE and a training binary over the whole model are not done. A 100M ternary model is public; nothing past 100M parameters is.
+; (ml/train_slice_model); the STE and a training binary over the whole model are not done. A 100M ternary model is public; a 0.5B ternary model is public too (1.2B tokens, weaker than the 100M one) and not yet reviewed.
 
 ; Not claimed: no model quality beyond the numbers in SOURCES; no GIF or MP4, stills only.
 ; phi^2 + 1/phi^2 = 3 | TRINITY
@@ -166,6 +184,10 @@ pub const CODER_COMMIT : str = "f60ac0c623e5ac24d74fdcde2dccb2d8e55e6d83";
 pub const TRINITY_COMMIT : str = "ee512359254bcd7438bb6a79c8a4a7f8b554b620";
 pub const TRAINING_COMMIT : str = "3e9648c1817a97c5f9b8c727b0f8e2291928eead";
 pub const FPGA_COMMIT : str = "b2bd5a30cefe32eb7e57e9f2025626f6f0cf4b25";
+/// igla-coder-gpu branch hf/card-tern-tc-500m: the commit whose card pins the 0.5B run (cell 9).
+pub const CARD_500M_COMMIT : str = "510a22c7520df01e9c768b972ed79ae58eaadead";
+/// playra/tern-tc-500m on Hugging Face: the commit that put the 0.5B release live (cell 9).
+pub const HF_500M_COMMIT : str = "7f203b66d53459f8edcfbc5c39c0e0060914875f";
 /// The realtime harness: the IGLA board's state endpoint, which trinity compiles in the browser.
 pub const LIVE_STATE_URL : str = "https://board-production-2dec.up.railway.app/state";
 pub const MEASURED_ON : str = "2026-10-17";
@@ -217,7 +239,7 @@ pub const SOURCES : [27]str = [
     "t27 specs/igla/coder/*.t27: 9 of 9 compile",
     "huggingface.co/playra/tern-tc-9m",
     "huggingface.co/playra/tern-tc-100m, final.pt sha256 620c80b6...d08300",
-    "",
+    "igla-coder-gpu specs/hf/model_card_500m.t27: 0.5B ternary, 1.2B tokens, bpb 0.5307, final.pt sha256 e6925b7c...8c15bc",
     "trinity docs/experiments/FOUND_EXPERIMENTS_SUMMARY.md",
     "trinity-training src/hslm, 64 Zig files",
     "trinity data/wave9/worker-2/hslm_step_10000.bin",
@@ -242,7 +264,7 @@ pub const SOURCES : [27]str = [
 pub const MISSING : [27]str = [
     "", "", "", "", "", "", "",
     "",
-    "any run past 100M parameters: none exists",
+    "",
     "", "", "", "", "", "", "",
     "",
     "",
@@ -254,21 +276,26 @@ pub const MISSING : [27]str = [
 ];
 
 /// The measurements per cell (see the header for how each was taken).
-pub const GROUNDED : [27]bool = [true, true, true, true, true, true, true, true, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true];
+pub const GROUNDED : [27]bool = [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true];
 pub const TEST_PASS : [27]bool = [false, false, false, true, true, true, false, false, false, false, false, false, true, true, true, false, false, false, false, false, false, true, true, true, false, false, false];
-pub const REVIEWED : [27]bool = [false, false, false, false, false, false, true, true, false, false, false, false, false, false, false, true, true, true, false, false, false, false, false, false, true, false, true];
+pub const REVIEWED : [27]bool = [false, false, false, false, false, false, true, true, true, false, false, false, false, false, false, true, true, true, false, false, false, false, false, false, true, false, true];
 /// The approving review behind REVIEWED: PR and review id on gHashTag/t27 (5561941493: t27-bees
 /// APPROVE of 014a34d21, glm-4.7-flash and glm-4.5-flash independently, 2026-10-18 07:28 UTC).
 pub const REVIEW_PR : i32 = 10275;
 pub const REVIEW_ID : str = "5561941493";
-pub const LIVE : [27]bool = [false, false, false, false, false, false, true, true, false, false, false, false, false, false, false, false, false, false, false, false, true, false, false, false, false, false, false];
+/// The review behind REVIEWED[8], cell 9: PR and review id on gHashTag/t27 (5563181513: t27-bees on
+/// 88f47f32d, 2026-10-18 13:29 UTC, model verdict approve, posted as NEEDS_PERSON because the head
+/// changes .trinity/seals and gen/json files; merged over that hold on the owner's instruction).
+pub const REVIEW_500M_PR : i32 = 10428;
+pub const REVIEW_500M_ID : str = "5563181513";
+pub const LIVE : [27]bool = [false, false, false, false, false, false, true, true, true, false, false, false, false, false, false, false, false, false, false, false, true, false, false, false, false, false, false];
 
 /// What a host passes the viewer, per cell: the earned status code and the step number as a label.
-pub const STATUS : [27]u8 = [4, 4, 4, 4, 4, 4, 5, 5, 0, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 4, 4, 4, 4, 3, 4];
+pub const STATUS : [27]u8 = [4, 4, 4, 4, 4, 4, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 4, 4, 4, 4, 3, 4];
 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, 100, 50, 90, 0, 100, 100, 20, 80, 65, 95, 75, 55, 40, 70, 15, 60, 40, 100, 60, 30, 10, 50];
+pub const PCT : [27]u8 = [60, 70, 80, 80, 60, 100, 50, 90, 30, 100, 100, 20, 80, 65, 95, 75, 55, 40, 70, 15, 60, 40, 100, 60, 30, 10, 50];
 
 /// The t27c test-report counts the PRACTICE cells quote.
 pub const ML_SPECS_MEASURED : i32 = 17;
@@ -293,8 +320,8 @@ pub const ROWS : [20]str = [
     "In t27: loss, optimizers, HSLM pass",
     "17 of 17 ML specs compile on the lab",
     "9 of 9 IGLA CODER specs compile",
-    "3 published, 22 verified, 1 draft",
-    "1 gap, about 61 percent overall",
+    "4 published, 22 verified, 1 draft",
+    "0 gaps, about 62 percent overall",
     "Batched f32 HSLM trainer",
     "Shadow weights for the whole model",
     "A whole-model training binary",
@@ -350,7 +377,7 @@ test every_step_is_a_cell {
 
 test a_gap_is_named_not_filled {
     assert gaps_are_named(SOURCES, MISSING, GROUNDED);
-    assert count_true(GROUNDED) == 26;
+    assert count_true(GROUNDED) == 27;
     var c : i32 = 0;
     while (c < CELL_COUNT) {
         const i : usize = @intCast(c);
@@ -364,12 +391,12 @@ test a_gap_is_named_not_filled {
 
 test a_status_is_earned_not_chosen {
     assert statuses_earned(STATUS, DOMAIN, GROUNDED, TEST_PASS, REVIEWED, LIVE);
-    // the evidence slide's "3 published, 22 verified, 1 draft" and "1 gap"
-    assert count_status(STATUS, STATUS_PUBLISHED) == 3;
+    // the evidence slide's "4 published, 22 verified, 1 draft" and "0 gaps"
+    assert count_status(STATUS, STATUS_PUBLISHED) == 4;
     assert count_status(STATUS, STATUS_VERIFIED) == 22;
     assert count_status(STATUS, STATUS_DRAFT) == 1;
-    assert count_status(STATUS, STATUS_GAP) == 1;
-    assert count_true(REVIEWED) == 7;
+    assert count_status(STATUS, STATUS_GAP) == 0;
+    assert count_true(REVIEWED) == 8;
     // every passing test is a PRACTICE step: the t27 bench, MultiPL-E and coder specs (cells 4 to
     // 6), loss, optimizers, forward pass, matvec, the ML specs that compile (cell 23) and the
     // file I/O t27c lowers (cell 24)
@@ -397,15 +424,18 @@ test a_status_is_earned_not_chosen {
     assert LIVE[24] == false;
     assert step_stage(24) == STAGE_DELIVER;
     assert STATUS[24] == STATUS_VERIFIED;
-    // the board is live and was sourced: the third published cell
+    // the board is live and was sourced: a third published cell
     assert STATUS[20] == STATUS_PUBLISHED;
 }
 
 test a_review_is_named_not_assumed {
-    // REVIEWED is true exactly at the grounded DELIVER cells PR REVIEW_PR put to review: every
-    // grounded DELIVER cell except step 26, whose evidence came in a PR with no review
+    // REVIEWED is true exactly at the grounded DELIVER cells put to review: the seven of PR
+    // REVIEW_PR, step 9 from PR REVIEW_500M_PR, and not step 26, whose evidence came in a PR with
+    // no review
     assert REVIEW_PR == 10275;
     assert REVIEW_ID.len > 0;
+    assert REVIEW_500M_PR == 10428;
+    assert REVIEW_500M_ID.len > 0;
     var c : i32 = 0;
     while (c < CELL_COUNT) {
         const i : usize = @intCast(c);
@@ -416,10 +446,15 @@ test a_review_is_named_not_assumed {
         assert REVIEWED[i] == want;
         c = c + 1;
     }
-    // the gap, step 9, is a DELIVER step with nothing to review
+    // step 9, the 0.5B run, is grounded by its model card (a full commit id) and live on Hugging
+    // Face (a full commit id), and reviewed in PR REVIEW_500M_PR: live and reviewed is published
+    assert CARD_500M_COMMIT.len == 40;
+    assert HF_500M_COMMIT.len == 40;
     assert step_stage(8) == STAGE_DELIVER;
-    assert GROUNDED[8] == false;
-    assert REVIEWED[8] == false;
+    assert GROUNDED[8];
+    assert REVIEWED[8];
+    assert LIVE[8];
+    assert STATUS[8] == STATUS_PUBLISHED;
     // step 26 is grounded but not reviewed, so it stays a draft
     assert step_stage(25) == STAGE_DELIVER;
     assert GROUNDED[25];
@@ -453,11 +488,11 @@ test the_compile_counts_are_the_measured_ones {
 }
 
 test percent_per_lens_is_a_plain_mean {
-    assert mean_pct(0, 9) == 65;
+    assert mean_pct(0, 9) == 68;
     assert mean_pct(9, 9) == 70;
     assert mean_pct(18, 9) == 48;
-    // the evidence slide's "about 61 percent overall"
-    assert mean_pct(0, CELL_COUNT) == 61;
+    // the evidence slide's "about 62 percent overall"
+    assert mean_pct(0, CELL_COUNT) == 62;
 }
 
 test the_training_step_cells_are_verified_once_reviewed {