spec(igla): own_model -- ground cells 8, 24 and 25, put the DELIVER evidence to review again (Closes #10184)

specs/igla/own_model.t27

@@ -71,13 +71,35 @@
 ;     0 vacuous (cells 16-18); playra/tern-tc-9m and tern-tc-9m-t27 200, tern-tc-100m 401 (cells 7
 ;     and 8); the board /state 200 (cell 21). A review of that PR is what the engine asks of cells
 ;     7, 16, 17, 18 and 27; REVIEWED turns true only in a later PR that names the approving review.
-;   - 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.
+;   - test on 2026-10-18, t27c with `--backend ml` (#10165), lab checkout at master 8887f17f4:
+;     ml/host_io_model 3/3 pass, 0 vacuous. `t27c compile specs/ml/host_io_model.t27 --backend ml`
+;     builds a binary that exits 0 and leaves a real 24-byte /tmp/t27_tokens.bin (bytes 00000000
+;     01000000 50c40000 ffffffff 07000000 78563412, sha256 48548924cac8...); with that path made a
+;     directory the binary aborts, exit 134. So t27c lowers file I/O (Linux syscalls on a real fd)
+;     and the arena of ml/arena_heap is plain t27 it already lowers: cell 24 has its passing test.
+;     The backend builds that file-I/O program only; the training binary is the next entry.
+;   - test on 2026-10-18, same t27c, lab checkout at master 2d358a708, after the PR that added
+;     ml/train_slice_model (#10186): 5/5 pass, 0 vacuous. `t27c compile
+;     specs/ml/train_slice_model.t27 --backend ml` builds a binary that runs 20 steps of the
+;     nn/hslm_backward training step through `use` (loss 3.2975044 to 0.32346588), writes the trained
+;     trits, biases and both losses to a 379-byte /tmp/t27_train_slice.bin, reads it back byte for
+;     byte and exits 0; with that path made a directory it aborts, exit 134. So t27c lowers a model
+;     spec to a training binary, for one 3-head slice; not for the whole 3-block model (cell 25).
+;   - published on 2026-10-18: playra/tern-tc-100m on Hugging Face, the 100M ternary model of the
+;     igla-coder-gpu ladder (n_layer 12, d_model 768), trained on RunPod; final.pt sha256
+;     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.
+;   - 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).
 ;     GET board-production-2dec.up.railway.app/state returned 200 and a made-up path 404; its
 ;     snapshot is dated 2026-10-05, so the board is live and 12 days stale.
 ;   - review: no PR behind any cell has a review, so no DELIVER cell is VERIFIED. #10164 merged on
-;     2026-10-18 02:43Z without one. The evidence above is put to review in the PR that adds this
-;     line, re-taken on the lab at its own base; REVIEWED still waits for a PR that names its review.
+;     2026-10-18 02:43Z without one, and #10179, which put the evidence to review, was merged at
+;     03:16Z by an auto-merge another session enabled, before the reviewer reached it; #10186
+;     (cell 25) merged at 03:42Z the same way. The PR that adds this line puts the evidence of
+;     cells 7, 8, 16, 17, 18, 25 and 27 to review again; REVIEWED still waits for a later PR that
+;     names the approving review.
 ;   - board on 2026-10-18, cell 26: specs/fpga/ddr3_head_bandwidth.t27 (11/11 tests, 0 vacuous on the
 ;     lab), lowered by t27b verilog around a LiteDRAM core, routed by openXC7 for xc7a200tfbg484-2 and
 ;     flashed on the AX7203 with openFPGALoader -c digilent_hs2. Its UART receipt: 325000 beats
@@ -90,7 +112,7 @@
 ;     from those that do. This is a sequential read of a pattern, no weight on the board and no
 ;     compute from DDR3, so the cell is a draft at 10 percent, a DELIVER step waiting for its review.
 ; 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: 1 published, 16 verified, 6 drafts, 4 gaps.
+; and a DELIVER step for a review. So: 1 published, 17 verified, 8 drafts, 1 gap.
 
 ; 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
@@ -100,9 +122,10 @@
 ; (PPL 2.96 at ctx 27) were wiped by EXP-002 and survive only in the ledger. In t27 the loss, four
 ; optimizers and the HSLM spec's forward and backward passes pass, and that spec is now the trained
 ; model, its VSA path included (see MEASURED); the ternary STE, the training step and the data
-; and checkpoint layout now pass too, on a model slice and on bytes; the STE over the whole model,
-; heap and file I/O exist as pure t27 contracts (ml/arena_heap, ml/file_io_contract) that t27c does not
-; lower, and `t27c --backend ml` does not exist.
+; 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.
 
 ; Not claimed: no model quality beyond the numbers in SOURCES; no GIF or MP4, stills only.
 ; phi^2 + 1/phi^2 = 3 | TRINITY
@@ -192,7 +215,7 @@ pub const SOURCES : [27]str = [
     "igla-coder-gpu results/report_pair_10b.md",
     "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",
     "",
     "trinity docs/experiments/FOUND_EXPERIMENTS_SUMMARY.md",
     "trinity-training src/hslm, 64 Zig files",
@@ -208,39 +231,39 @@ pub const SOURCES : [27]str = [
     "board-production-2dec.up.railway.app/state",
     "t27 specs/port/trinity/fpga/openxc7-synth/ternary_matvec.t27",
     "t27c test-report on 17 ML specs: 17 compile",
-    "",
-    "",
+    "t27 specs/ml/host_io_model.t27 3/3 via t27c --backend ml (#10165); arena in ml/arena_heap.t27",
+    "t27 specs/ml/train_slice_model.t27 5/5; its --backend ml binary trains one slice 20 steps (#10186)",
     "t27 specs/fpga/ddr3_head_bandwidth.t27: head read 2769 MB/s on the AX7203; no weights on it",
     "t27 specs/igla/own_model.t27"
 ];
 
 /// What would ground each gap; "" for a cell that has its source.
 pub const MISSING : [27]str = [
     "", "", "", "", "", "", "",
-    "a public 100M checkpoint: playra/tern-tc-100m answers 401, like a made-up name",
+    "",
     "any run past 100M parameters: none exists",
     "", "", "", "", "", "", "",
     "",
     "",
     "", "", "", "", "",
-    "heap allocation and file I/O lowered by t27c: neither exists; ml/arena_heap.t27 and ml/file_io_contract.t27 are the contract it must meet",
-    "a t27c backend that lowers a model spec to a training binary: none exists",
+    "",
+    "",
     "",
     ""
 ];
 
 /// The measurements per cell (see the header for how each was taken).
-pub const GROUNDED : [27]bool = [true, true, true, true, true, true, true, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, 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, false, false, false, false];
+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 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false];
-pub const LIVE : [27]bool = [false, false, false, false, false, false, true, false, false, false, false, false, false, false, false, false, false, false, false, false, true, false, false, false, false, false, false];
+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];
 
 /// 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, 3, 0, 0, 4, 4, 4, 4, 4, 4, 3, 3, 3, 4, 4, 5, 4, 4, 0, 0, 3, 3];
+pub const STATUS : [27]u8 = [4, 4, 4, 4, 4, 4, 3, 3, 0, 4, 4, 4, 4, 4, 4, 3, 3, 3, 4, 4, 5, 4, 4, 4, 3, 3, 3];
 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, 0, 0, 100, 100, 20, 80, 65, 95, 75, 55, 40, 70, 15, 60, 40, 100, 0, 0, 10, 50];
+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];
 
 /// The t27c test-report counts the PRACTICE cells quote.
 pub const ML_SPECS_MEASURED : i32 = 17;
@@ -265,11 +288,11 @@ 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",
-    "1 published, 16 verified, 6 drafts",
-    "4 gaps, about 54 percent overall",
+    "1 published, 17 verified, 8 drafts",
+    "1 gap, about 61 percent overall",
     "Batched f32 HSLM trainer",
     "Shadow weights for the whole model",
-    "File I/O, heap, --backend ml",
+    "A whole-model training binary",
     ""
 ];
 
@@ -322,7 +345,7 @@ test every_step_is_a_cell {
 
 test a_gap_is_named_not_filled {
     assert gaps_are_named(SOURCES, MISSING, GROUNDED);
-    assert count_true(GROUNDED) == 23;
+    assert count_true(GROUNDED) == 26;
     var c : i32 = 0;
     while (c < CELL_COUNT) {
         const i : usize = @intCast(c);
@@ -336,15 +359,16 @@ 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 "1 published, 16 verified, 6 drafts" and "4 gaps"
+    // the evidence slide's "1 published, 17 verified, 8 drafts" and "1 gap"
     assert count_status(STATUS, STATUS_PUBLISHED) == 1;
-    assert count_status(STATUS, STATUS_VERIFIED) == 16;
-    assert count_status(STATUS, STATUS_DRAFT) == 6;
-    assert count_status(STATUS, STATUS_GAP) == 4;
+    assert count_status(STATUS, STATUS_VERIFIED) == 17;
+    assert count_status(STATUS, STATUS_DRAFT) == 8;
+    assert count_status(STATUS, STATUS_GAP) == 1;
     assert count_true(REVIEWED) == 0;
     // every passing test is a PRACTICE step: the t27 bench, MultiPL-E and coder specs (cells 4 to
-    // 6), loss, optimizers, forward pass, matvec, and the ML specs that compile (cell 23)
-    assert count_true(TEST_PASS) == 8;
+    // 6), loss, optimizers, forward pass, matvec, the ML specs that compile (cell 23) and the
+    // file I/O t27c lowers (cell 24)
+    assert count_true(TEST_PASS) == 9;
     var c : i32 = 0;
     while (c < CELL_COUNT) {
         const i : usize = @intCast(c);
@@ -356,6 +380,14 @@ test a_status_is_earned_not_chosen {
     // the 9M model is live, but a live page does not stand in for a review
     assert LIVE[6];
     assert STATUS[6] == STATUS_DRAFT;
+    // and the same holds for the public 100M model (step 8)
+    assert LIVE[7];
+    assert STATUS[7] == STATUS_DRAFT;
+    assert STATUS[23] == STATUS_VERIFIED;
+    // the slice training binary (step 25) is grounded, but a DELIVER step waits for a review
+    assert GROUNDED[24];
+    assert step_stage(24) == STAGE_DELIVER;
+    assert STATUS[24] == STATUS_DRAFT;
     // the board is live and was sourced, so it is the one published cell
     assert STATUS[20] == STATUS_PUBLISHED;
 }
@@ -386,11 +418,11 @@ test the_compile_counts_are_the_measured_ones {
 }
 
 test percent_per_lens_is_a_plain_mean {
-    assert mean_pct(0, 9) == 55;
+    assert mean_pct(0, 9) == 65;
     assert mean_pct(9, 9) == 70;
-    assert mean_pct(18, 9) == 38;
-    // the evidence slide's "about 54 percent overall"
-    assert mean_pct(0, CELL_COUNT) == 54;
+    assert mean_pct(18, 9) == 48;
+    // the evidence slide's "about 61 percent overall"
+    assert mean_pct(0, CELL_COUNT) == 61;
 }
 
 test the_training_step_cells_are_drafts_until_reviewed {

specs/igla/own_model_board.t27

@@ -48,23 +48,23 @@ pub const LENS_CELLS : i32 = 9;
 
 pub const JOB : str = "own_model";
 pub const STAGE : str = "own_model";
-pub const MSG : str = "1 published, 16 verified, 6 drafts, 4 gaps; 54%";
+pub const MSG : str = "1 published, 17 verified, 8 drafts, 1 gaps; 61%";
 /// What MSG says once its numbers are taken out.
 pub const MSG_WORDS : str = " published,  verified,  drafts,  gaps; %";
-pub const OVERALL_PCT : u8 = 54;
-pub const PROGRESS : f64 = 0.54;
+pub const OVERALL_PCT : u8 = 61;
+pub const PROGRESS : f64 = 0.61;
 
 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 = [55, 70, 38];
-pub const CELL_STATUS : [27]u8 = [4, 4, 4, 4, 4, 4, 3, 0, 0, 4, 4, 4, 4, 4, 4, 3, 3, 3, 4, 4, 5, 4, 4, 0, 0, 3, 3];
-pub const CELL_PCT : [27]u8 = [60, 70, 80, 80, 60, 100, 50, 0, 0, 100, 100, 20, 80, 65, 95, 75, 55, 40, 70, 15, 60, 40, 100, 0, 0, 10, 50];
+pub const LENS_PCT : [3]u8 = [65, 70, 48];
+pub const CELL_STATUS : [27]u8 = [4, 4, 4, 4, 4, 4, 3, 3, 0, 4, 4, 4, 4, 4, 4, 3, 3, 3, 4, 4, 5, 4, 4, 4, 3, 3, 3];
+pub const CELL_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];
 
 /// The size of the compact metrics object, in bytes, computed from the keys and values above.
-pub const METRICS_BYTES : i32 = 627;
+pub const METRICS_BYTES : i32 = 630;
 
 pub fn is_digit(ch: u8) -> bool {
     return ch >= '0' and ch <= '9';
@@ -245,10 +245,10 @@ test a_gap_cell_is_never_above_zero_percent {
         assert CELL_PCT[i] <= 100;
         c = c + 1;
     }
-    // the public 100M model, cell 7, is a gap: drawn at 10 percent it is refused
-    assert CELL_STATUS[7] == STATUS_GAP;
+    // the run past 100M parameters, cell 8, is a gap: drawn at 10 percent it is refused
+    assert CELL_STATUS[8] == STATUS_GAP;
     var wrong : [27]u8 = CELL_PCT;
-    wrong[7] = 10;
+    wrong[8] = 10;
     assert gaps_at_zero(CELL_STATUS, wrong) == false;
 }
 
@@ -261,9 +261,9 @@ test the_tally_sums_to_twenty_seven {
     assert count_status(CELL_STATUS, STATUS_OPEN) == 0;
     assert published + verified + drafts + gaps == CELL_COUNT;
     assert published == 1;
-    assert verified == 16;
-    assert drafts == 6;
-    assert gaps == 4;
+    assert verified == 17;
+    assert drafts == 8;
+    assert gaps == 1;
 }
 
 test the_message_is_the_tally {
@@ -273,12 +273,12 @@ 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, 15 verified, 6 drafts, 4 gaps; 54%", STATUS, overall) == false;
-    assert msg_matches("1 published, 16 verified, 6 drafts, 4 holes; 54%", STATUS, overall) == false;
-    assert msg_matches("1 published, 16 verified, 6 drafts, 4 gaps; 54% of 100", STATUS, overall) == false;
-    assert msg_matches(MSG, STATUS, 48) == false;
-    assert nth_number(MSG, 1) == 16;
-    assert nth_number(MSG, 4) == 54;
+    assert msg_matches("1 published, 17 verified, 7 drafts, 1 gaps; 61%", STATUS, overall) == false;
+    assert msg_matches("1 published, 17 verified, 8 drafts, 1 holes; 61%", STATUS, overall) == false;
+    assert msg_matches("1 published, 17 verified, 8 drafts, 1 gaps; 61% of 100", STATUS, overall) == false;
+    assert msg_matches(MSG, STATUS, 60) == false;
+    assert nth_number(MSG, 2) == 8;
+    assert nth_number(MSG, 4) == 61;
     assert nth_number(MSG, 5) == -1;
 }