spec(igla): own_model.t27 -- the 27-part map of our own model (Closes #10026)

specs/igla/own_model.t27

@@ -0,0 +1,333 @@
+// SPDX-License-Identifier: Apache-2.0
+; specs/igla/own_model.t27 -- the one map of "our own model": IGLA CODER, HSLM in t27, and the
+; harness that has to run them. 27 cells, one T27F reel, the same workflow as the two topic reels
+; (specs/t27f/reel_t27_basics.t27, specs/t27f/reel_gf16_nmse.t27): topic_engine.t27 decides each
+; cell's status from what was measured, topic_reel.t27 checks it, reel.t27 and deck.t27 draw it.
+
+; SOURCES, read on MEASURED_ON:
+;   - gHashTag/igla-coder-gpu at CODER_COMMIT (private): README.md, LEDGER.md,
+;     results/report_pair_10b.md, research/T27_BENCH.md, research/MODEL_ON_FPGA_PLAN.md.
+;   - gHashTag/trinity at TRINITY_COMMIT: docs/experiments/FOUND_EXPERIMENTS_SUMMARY.md (the HSLM
+;     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/t27 at SOURCE_COMMIT: the ML specs named per cell.
+
+; MEASURED on 2026-10-17:
+;   - test: t27c test-report (built from SOURCE_COMMIT, zig 0.16.0) on the Railway lab, 17 ML specs:
+;     cross_entropy_loss 5/5, sgd 4/4, adam 4/4, lamb 4/4 pass, 0 vacuous. BLOCKED, never compiled:
+;     nn/hslm (undeclared 'constants'), hslm/forward_pass ('q_role'), nn/attention
+;     ('PHI_CUBED_INV'), optimizer/adamw ('allocate_array'), and all nine specs/igla/coder/*.t27.
+;     ternary_matvec.t27 (the FPGA matvec, outside the 17) passed 4/4.
+;   - 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
+;     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.
+; 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, 11 verified, 8 drafts, 7 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
+; cell weighted the same; that weighting is a convention, not a measurement.
+
+; 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 and
+; three optimizers pass, the forward pass does not compile, and the backward pass, ternary STE,
+; 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
+
+module own_model;
+
+use t27f::reel::CELL_COUNT;
+use t27f::reel::DEMO_COUNT;
+use t27f::reel::DEMO_CELL;
+use t27f::reel::DEMO_LENS;
+use t27f::topic_engine::STATUS_GAP;
+use t27f::topic_engine::STATUS_DRAFT;
+use t27f::topic_engine::STATUS_VERIFIED;
+use t27f::topic_engine::STATUS_PUBLISHED;
+use t27f::topic_engine::STAGE_PRACTICE;
+use t27f::topic_engine::DOMAIN_ENGINEERING;
+use t27f::topic_engine::CONTENT_WORDS;
+use t27f::topic_engine::ENGINEERING_WORDS;
+use t27f::topic_engine::RESEARCH_WORDS;
+use t27f::topic_engine::SALES_WORDS;
+use t27f::topic_engine::keyword_hits;
+use t27f::topic_engine::route_brief;
+use t27f::topic_engine::step_stage;
+use t27f::topic_reel::count_status;
+use t27f::topic_reel::count_true;
+use t27f::topic_reel::statuses_earned;
+use t27f::topic_reel::gaps_are_named;
+use t27f::topic_reel::slides_fit;
+use t27f::topic_reel::stage_slides_bare;
+
+pub const TOPIC_ID : str = "own-model";
+pub const SOURCE_REPO : str = "gHashTag/t27";
+pub const SOURCE_COMMIT : str = "efe3c9431";
+pub const CODER_REPO : str = "gHashTag/igla-coder-gpu";
+pub const CODER_COMMIT : str = "f60ac0c623e5ac24d74fdcde2dccb2d8e55e6d83";
+pub const TRINITY_COMMIT : str = "ee512359254bcd7438bb6a79c8a4a7f8b554b620";
+pub const TRAINING_COMMIT : str = "3e9648c1817a97c5f9b8c727b0f8e2291928eead";
+pub const FPGA_COMMIT : str = "b2bd5a30cefe32eb7e57e9f2025626f6f0cf4b25";
+/// 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";
+
+/// The brief the engine routes. spec, compiler, fpga are 3 engineering hits; no other list is hit.
+pub const BRIEF : str = "Our own model, written in t27: IGLA CODER, HSLM and the harness that runs them. One map per spec, compiler step and fpga receipt, with every gap named.";
+/// The domain route_brief gives that brief (see the_domain_is_routed_not_chosen).
+pub const DOMAIN : u8 = 2;
+
+pub const LENS_TITLES : [3]str = ["IGLA CODER", "HSLM in t27", "harness"];
+
+pub const TITLES : [27]str = [
+    "Corpus, 11.7B tokens of code",
+    "BPE 32K tokenizer with FIM",
+    "Size ladder 13M to 100M",
+    "t27 bench, 466 items",
+    "MultiPL-E pass@1 harness",
+    "IGLA CODER specs in t27",
+    "tern-tc-9m on Hugging Face",
+    "A public 100M model",
+    "Scale-up to 0.5-1.5B",
+    "HSLM experiment ledger",
+    "HSLM Zig trainer",
+    "The one checkpoint left",
+    "Loss: cross entropy",
+    "Optimizers: sgd, adam, lamb",
+    "Forward pass",
+    "Backward pass",
+    "Ternary straight-through",
+    "Data loader and checkpoint I/O",
+    "tc engine, C and PyTorch parity",
+    "FPGA: 8 tokens on the AX7203",
+    "Live IGLA board",
+    "Ternary matvec in t27",
+    "ML specs that compile",
+    "Heap and file I/O in t27c",
+    "t27c --backend ml",
+    "Full model on FPGA, DDR3",
+    "This map"
+];
+
+/// Where each cell's evidence lives; "" is a gap.
+pub const SOURCES : [27]str = [
+    "igla-coder-gpu LEDGER.md (11.7B-token 32K corpus)",
+    "igla-coder-gpu README.md data.py row",
+    "igla-coder-gpu LEDGER.md ladder",
+    "igla-coder-gpu research/T27_BENCH.md",
+    "igla-coder-gpu results/report_pair_10b.md",
+    "t27 specs/igla/coder/*.t27: 0 of 9 compile",
+    "huggingface.co/playra/tern-tc-9m",
+    "",
+    "",
+    "trinity docs/experiments/FOUND_EXPERIMENTS_SUMMARY.md",
+    "trinity-training src/hslm, 64 Zig files",
+    "trinity data/wave9/worker-2/hslm_step_10000.bin",
+    "t27 specs/ml/loss/cross_entropy_loss.t27",
+    "t27 specs/ml/optimizer/{sgd,adam,lamb}.t27",
+    "t27 specs/nn/hslm.t27 and specs/hslm/forward_pass.t27",
+    "t27 specs/nn/hslm.t27 backward placeholders",
+    "",
+    "",
+    "igla-coder-gpu research/MODEL_ON_FPGA_PLAN.md",
+    "trinity-fpga b2bd5a30, 7682304/7682304 receipts",
+    "board-production-2dec.up.railway.app/state",
+    "t27 specs/port/trinity/fpga/openxc7-synth/ternary_matvec.t27",
+    "t27c test-report on 17 ML specs: 4 compile",
+    "",
+    "",
+    "",
+    "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",
+    "", "", "", "", "", "", "",
+    "a ternary straight-through estimator in any .t27 spec: none exists",
+    "a t27 reader for token streams and MSLH checkpoints: none exists",
+    "", "", "", "", "",
+    "heap allocation and file I/O lowered by t27c: neither exists",
+    "a t27c backend that lowers a model spec to a training binary: none exists",
+    "embeddings, head and weights in DDR3 on the board: they stay on the host",
+    ""
+];
+
+/// 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, false, false, true, true, true, true, true, false, false, false, true];
+pub const TEST_PASS : [27]bool = [false, false, false, false, false, false, false, false, false, false, false, false, true, true, false, false, false, false, false, false, false, true, false, false, 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];
+
+/// 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, 3, 3, 3, 3, 0, 0, 4, 4, 4, 4, 4, 3, 3, 0, 0, 4, 4, 5, 4, 3, 0, 0, 0, 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, 10, 50, 0, 0, 100, 100, 20, 80, 65, 35, 5, 0, 0, 70, 15, 60, 40, 23, 0, 0, 0, 50];
+
+/// The t27c test-report counts the PRACTICE cells quote.
+pub const ML_SPECS_MEASURED : i32 = 17;
+pub const ML_SPECS_COMPILING : i32 = 4;
+pub const CODER_SPECS : i32 = 9;
+pub const CODER_SPECS_COMPILING : i32 = 0;
+
+/// Slide text per demo scene, checked by topic_reel.t27 slides_fit (index scene*4+row).
+pub const HEAD : [5]str = [
+    "Own model: 27 parts",
+    "Steps 1-9: coder",
+    "HSLM existed",
+    "Evidence",
+    "What is left"
+];
+pub const ROWS : [20]str = [
+    "", "", "", "",
+    "", "", "", "",
+    "Step 11 of 27: the HSLM trainer",
+    "64 Zig files; best PPL 2.96 at ctx 27",
+    "Those checkpoints were wiped",
+    "In t27: loss and 3 optimizers pass",
+    "4 of 17 ML specs compile on the lab",
+    "0 of 9 IGLA CODER specs compile",
+    "1 published, 11 verified, 8 drafts",
+    "7 gaps, about 39 percent overall",
+    "Make nn/hslm.t27 compile",
+    "Backward pass and ternary STE",
+    "Loader, checkpoint I/O, --backend ml",
+    ""
+];
+
+pub fn step_number(cell: i32) -> i32 {
+    return cell + 1;
+}
+
+/// part of whole, in whole percent, truncated.
+pub fn share_pct(part: i32, whole: i32) -> i32 {
+    return @divTrunc(part * 100, whole);
+}
+
+/// Mean PCT of cells [from, from+n), truncated: every cell weighs the same.
+pub fn mean_pct(from: i32, n: i32) -> i32 {
+    var sum : i32 = 0;
+    var c : i32 = from;
+    while (c < from + n) {
+        const p : i32 = @intCast(PCT[@intCast(c)]);
+        sum = sum + p;
+        c = c + 1;
+    }
+    return @divTrunc(sum, n);
+}
+
+test the_domain_is_routed_not_chosen {
+    assert DOMAIN == DOMAIN_ENGINEERING;
+    assert keyword_hits(BRIEF, BRIEF.len, ENGINEERING_WORDS, ENGINEERING_WORDS.len) == 3;
+    assert keyword_hits(BRIEF, BRIEF.len, CONTENT_WORDS, CONTENT_WORDS.len) == 0;
+    assert keyword_hits(BRIEF, BRIEF.len, RESEARCH_WORDS, RESEARCH_WORDS.len) == 0;
+    assert keyword_hits(BRIEF, BRIEF.len, SALES_WORDS, SALES_WORDS.len) == 0;
+    assert route_brief(BRIEF, BRIEF.len) == DOMAIN;
+}
+
+test every_step_is_a_cell {
+    assert CELL_COUNT == 27;
+    var c : i32 = 0;
+    while (c < CELL_COUNT) {
+        const i : usize = @intCast(c);
+        assert TITLES[i].len > 0;
+        var want : i32 = 1;
+        if (step_number(c) >= 10) {
+            want = 2;
+        }
+        const len : i32 = @intCast(LABELS[i].len);
+        assert len == want;
+        c = c + 1;
+    }
+    assert LENS_TITLES[2].len > 0;
+}
+
+test a_gap_is_named_not_filled {
+    assert gaps_are_named(SOURCES, MISSING, GROUNDED);
+    assert count_true(GROUNDED) == 20;
+    var c : i32 = 0;
+    while (c < CELL_COUNT) {
+        const i : usize = @intCast(c);
+        // a gap is drawn as a gap, and only a gap is; a gap is a part that does not exist: 0 percent
+        assert (STATUS[i] == STATUS_GAP) == (GROUNDED[i] == false);
+        assert (GROUNDED[i] == false) == (PCT[i] == 0);
+        assert PCT[i] <= 100;
+        c = c + 1;
+    }
+}
+
+test a_status_is_earned_not_chosen {
+    assert statuses_earned(STATUS, DOMAIN, GROUNDED, TEST_PASS, REVIEWED, LIVE);
+    // the evidence slide's "1 published, 11 verified, 8 drafts" and "7 gaps"
+    assert count_status(STATUS, STATUS_PUBLISHED) == 1;
+    assert count_status(STATUS, STATUS_VERIFIED) == 11;
+    assert count_status(STATUS, STATUS_DRAFT) == 8;
+    assert count_status(STATUS, STATUS_GAP) == 7;
+    assert count_true(REVIEWED) == 0;
+    // every passing test is a PRACTICE step: loss, optimizers, matvec
+    assert count_true(TEST_PASS) == 3;
+    var c : i32 = 0;
+    while (c < CELL_COUNT) {
+        const i : usize = @intCast(c);
+        if (TEST_PASS[i]) {
+            assert step_stage(@intCast(c)) == STAGE_PRACTICE;
+        }
+        c = c + 1;
+    }
+    // the 9M model is live, but a live page does not stand in for a review
+    assert LIVE[6];
+    assert STATUS[6] == STATUS_DRAFT;
+    // the board is live and was sourced, so it is the one published cell
+    assert STATUS[20] == STATUS_PUBLISHED;
+}
+
+test the_compile_counts_are_the_measured_ones {
+    // the IGLA CODER specs cell: grounded, none compiles, so it is a draft at 10 percent
+    assert CODER_SPECS_COMPILING == 0;
+    assert STATUS[5] == STATUS_DRAFT;
+    assert TEST_PASS[5] == false;
+    // the compile cell's percent is the measured share, truncated
+    const share : i32 = share_pct(ML_SPECS_COMPILING, ML_SPECS_MEASURED);
+    assert share == 23;
+    const p22 : i32 = @intCast(PCT[22]);
+    assert p22 == share;
+    assert share_pct(CODER_SPECS_COMPILING, CODER_SPECS) == 0;
+    assert CODER_SPECS <= ML_SPECS_MEASURED;
+}
+
+test percent_per_lens_is_a_plain_mean {
+    assert mean_pct(0, 9) == 45;
+    assert mean_pct(9, 9) == 45;
+    assert mean_pct(18, 9) == 28;
+    // the evidence slide's "about 39 percent overall"
+    assert mean_pct(0, CELL_COUNT) == 39;
+}
+
+test the_camera_is_the_demo_camera {
+    assert DEMO_COUNT == 5;
+    // scene 1 shows lens 0, the row HEAD[1] calls "Steps 1-9"
+    assert DEMO_LENS[1] == 0;
+    // scenes 2 and 3 focus cell 10, which the focus slide calls "Step 11": the HSLM trainer
+    assert DEMO_CELL[2] == 10;
+    assert DEMO_CELL[3] == 10;
+    assert step_number(DEMO_CELL[2]) == 11;
+    assert STATUS[10] == STATUS_VERIFIED;
+}
+
+test every_slide_fits_every_preset {
+    assert slides_fit(HEAD, ROWS);
+    assert stage_slides_bare(ROWS);
+}
+
+invariant twenty_seven_parts {
+    assert CELL_COUNT == 27;
+}