gHashTag · gHashTag · Jul 8, 2026 · Jul 8, 2026 · Jul 8, 2026 · Jul 8, 2026
@@ -0,0 +1,41 @@
+# DECONTAM.md — IGLA CODER v0.1
+
+**Version:** 0.1
+**Method:** Lee 2022 k=50 substring bidirectional check.
+**Reference:** Lee et al., "Deduplicating Training Data Makes Language Models Better" (arXiv:2107.06499).
+
+## Contract (mandatory when the dataset is used to train IGLA CODER/RACE)
+
+For every training pair `(spec, gen)` and every held-out evaluation program `eval_i`:
+
+1. **Forward direction:** no 50-character substring of `spec` or `gen` appears verbatim in any `eval_i` (case-sensitive, whitespace-normalised).
+2. **Reverse direction:** no 50-character substring of any `eval_i` appears verbatim in any `spec` or `gen`.
+
+**Failure of either direction** => the pair is quarantined (moved to `dataset/igla-coder/v0.1/quarantine/`), never enters `pairs/`.
+
+## v0.1 status: SKIPPED — trivially clean
+
+Held-out evaluation set at v0.1 is **empty**. The directory `held-out-eval/` exists but contains no programs. Therefore:
+
+- Bidirectional decontam is **trivially clean** (there is nothing to compare against).
+- `manifest.decontam_bidirectional = false` (honest: check was not meaningfully performed).
+- Per-pair `metadata.json` marks `decontam_status = "Skipped"`.
+
+**This is a known v0.1 gap and is intentional.** Populating the held-out set before running decontam properly is a v0.2 responsibility.
+
+## v0.2 plan (out of scope this Wave)
+
+1. Author 12-16 held-out programs (matching topic distribution of pairs but name-disjoint).
+2. Store under `held-out-eval/<eval_id>/spec.tri` (and `gen.<lang>` where applicable).
+3. Add `decontam_check.py` (deterministic, stdlib-only, k=50).
+4. Re-run: assert every pair passes both directions. Fail-loud on any violation.
+5. Publish `decontam_report_v02.md` alongside MANIFEST v0.2.
+
+## Bibliographic honesty
+
+- Lee 2022 (arXiv:2107.06499) establishes the k=50 threshold for near-duplicate detection in LM training corpora. Applying the same threshold bidirectionally to eval-vs-train is a stricter variant used by BIG-bench maintainers and OpenAI's HumanEval-style protocols.
+- We do NOT claim originality on this method. We ship a placeholder because v0.1 is a **seed**, not a training corpus.
+
+## Verdict
+
+`DECONTAM v0.1 = SKIPPED (justified). NOT SAFE TO TRAIN on this dataset as-is; awaiting held-out eval.`
@@ -0,0 +1,145 @@
+{
+  "schema_version": 1,
+  "dataset_version": "0.1",
+  "built_at_iso": "2026-07-08T07:31:18.026643+00:00",
+  "pair_count": 8,
+  "decontam_bidirectional": false,
+  "held_out_eval_defined": false,
+  "note": "v0.1 seed: 8 spec-only pairs. Codegen pending TRAIN-BOX (rust/zig/ts). Decontam trivially clean because held-out-eval is empty at v0.1; v0.2 must populate held-out-eval FIRST.",
+  "law_compliance": {
+    "L1": "traceable via Closes #new-N in each ring PR body",
+    "L3": "ASCII+English verified per-file below",
+    "L4": ">=3 invariants + >=8 tests + >=2 bench per spec (visual audit)",
+    "L6": "no new numeric formats introduced; live catalog count = 83 verified via grep -c '// CATALOG:' specs/numeric/formats_catalog.t27"
+  },
+  "pairs": [
+    {
+      "pair_id": "0001-experience-format",
+      "spec_path_in_t27": "specs/organism/experience.tri",
+      "spec_local_path": "pairs/0001-experience-format/spec.tri",
+      "spec_sha256": "ea43af76d3c1214721cd40dbc624062cc17707525c55c824612e3941b3c3c7ef",
+      "spec_bytes": 4096,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "None",
+      "ring_id": "ring-105-001",
+      "license": "Apache-2.0",
+      "decontam_status": "Skipped",
+      "codegen_status": "spec_only",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    },
+    {
+      "pair_id": "0002-mozg-state-machine",
+      "spec_path_in_t27": "specs/organism/mozg.tri",
+      "spec_local_path": "pairs/0002-mozg-state-machine/spec.tri",
+      "spec_sha256": "ace8235ead1fded0cdd3a2bad33fefcfddfa4547bff66bfd5c6c11d62f392dc6",
+      "spec_bytes": 3473,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "Zig",
+      "ring_id": "ring-105-002",
+      "license": "Apache-2.0",
+      "decontam_status": "Skipped",
+      "codegen_status": "codegen_pending_train_box",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    },
+    {
+      "pair_id": "0003-dna-schema",
+      "spec_path_in_t27": "specs/organism/dna.tri",
+      "spec_local_path": "pairs/0003-dna-schema/spec.tri",
+      "spec_sha256": "e17879a9431831d03d101417be6eb36748e07112de95cedadac917742abcad37",
+      "spec_bytes": 3399,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "Zig",
+      "ring_id": "ring-105-002",
+      "license": "Apache-2.0",
+      "decontam_status": "Skipped",
+      "codegen_status": "codegen_pending_train_box",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    },
+    {
+      "pair_id": "0004-git-orchestrator",
+      "spec_path_in_t27": "specs/git/orchestrator.tri",
+      "spec_local_path": "pairs/0004-git-orchestrator/spec.tri",
+      "spec_sha256": "0eebd8a9045fd16b1c3bd46fc1518d5c82dcf51cdc579006222c91fcfeb4e244",
+      "spec_bytes": 4263,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "Zig",
+      "ring_id": "ring-105-003",
+      "license": "Apache-2.0",
+      "decontam_status": "Skipped",
+      "codegen_status": "codegen_pending_train_box",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    },
+    {
+      "pair_id": "0005-mcp-tool-registry",
+      "spec_path_in_t27": "specs/mcp/tool_registry.tri",
+      "spec_local_path": "pairs/0005-mcp-tool-registry/spec.tri",
+      "spec_sha256": "d00d00154073aa13a376b762702a36c50c4b5c2ea83ac1cd6afdfa5696c2872b",
+      "spec_bytes": 3959,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "Rust",
+      "ring_id": "ring-105-004",
+      "license": "Apache-2.0",
+      "decontam_status": "Skipped",
+      "codegen_status": "codegen_stub_pending_wave2",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    },
+    {
+      "pair_id": "0006-scene-schema",
+      "spec_path_in_t27": "specs/scenes/scene_schema.tri",
+      "spec_local_path": "pairs/0006-scene-schema/spec.tri",
+      "spec_sha256": "88029f3274e76ae9112fe05dd672904f761d8e686c273e06719e8f6ffe1712d1",
+      "spec_bytes": 3843,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "TypeScript",
+      "ring_id": "ring-105-005",
+      "license": "Apache-2.0-anonymised",
+      "decontam_status": "Skipped",
+      "codegen_status": "codegen_deferred_wave3",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    },
+    {
+      "pair_id": "0007-ring-runtime",
+      "spec_path_in_t27": "specs/organism/ring_runtime.tri",
+      "spec_local_path": "pairs/0007-ring-runtime/spec.tri",
+      "spec_sha256": "ba5421b4fb88095c3edb629415d12fb2ed6bc1a7392934f6426799d430ae2979",
+      "spec_bytes": 3848,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "Rust",
+      "ring_id": "ring-105-006",
+      "license": "Apache-2.0-anonymised",
+      "decontam_status": "Skipped",
+      "codegen_status": "codegen_stub_pending_wave2",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    },
+    {
+      "pair_id": "0008-igla-manifest",
+      "spec_path_in_t27": "specs/dataset/igla_coder_manifest.tri",
+      "spec_local_path": "pairs/0008-igla-manifest/spec.tri",
+      "spec_sha256": "a09c423ceda7d99360a66e60c661feb39a35ddc53453b6aeb788bf325894219e",
+      "spec_bytes": 4366,
+      "gen_path_in_t27": null,
+      "gen_sha256": null,
+      "target_lang": "None",
+      "ring_id": "ring-105-007",
+      "license": "Apache-2.0",
+      "decontam_status": "Skipped",
+      "codegen_status": "spec_only",
+      "l3_ascii_pass": true,
+      "l3_non_ascii_bytes": 0
+    }
+  ]
+}
@@ -0,0 +1,63 @@
+# IGLA CODER Dataset v0.1 — Seed
+
+**Status:** SEED (not a training corpus).
+**Pair count:** 8.
+**Total spec bytes:** ~31 KB.
+**Estimated token count:** ~7-10K (bpe token equivalent, rough estimate; measured at v0.2 with a tokenizer spec).
+
+## Purpose
+
+First public seed of the IGLA CODER training corpus: parallel `(spec, gen)` pairs where `spec` is a `.tri` T27 specification and `gen` is the corresponding generated code (Zig / Rust / Verilog / TypeScript — the latter deferred to Wave-3).
+
+## What is here (v0.1)
+
+Eight `.tri` specs, one per pilot ring of the 2026-07-08 ecosystem Wave (`ring-105-001..007`):
+
+| Pair ID | Spec | Target lang | Codegen status | Ring |
+|---|---|---|---|---|
+| 0001-experience-format | organism/experience.tri | None (spec-only) | n/a | 105-001 |
+| 0002-mozg-state-machine | organism/mozg.tri | Zig | TRAIN-BOX pending | 105-002 |
+| 0003-dna-schema | organism/dna.tri | Zig | TRAIN-BOX pending | 105-002 |
+| 0004-git-orchestrator | git/orchestrator.tri | Zig | TRAIN-BOX pending | 105-003 |
+| 0005-mcp-tool-registry | mcp/tool_registry.tri | Rust | STUB (Wave-2) | 105-004 |
+| 0006-scene-schema | scenes/scene_schema.tri | TypeScript | DEFERRED (Wave-3) | 105-005 |
+| 0007-ring-runtime | organism/ring_runtime.tri | Rust | STUB (Wave-2) | 105-006 |
+| 0008-igla-manifest | dataset/igla_coder_manifest.tri | None (spec-only) | n/a | 105-007 |
+
+## What is NOT here (honest gaps)
+
+- **No generated code.** All pairs are spec-only or spec-with-codegen-stub. Wave-2 (rust codegen) and Wave-3 (typescript codegen + real train-box regeneration) fill this in.
+- **No held-out evaluation set.** `held-out-eval/` is empty. See DECONTAM.md.
+- **No tokenizer.** Byte-level fallback assumed; explicit tokenizer spec is Wave-4.
+- **No BPB baseline.** Impossible without a real train run.
+- **~7-10K tokens is 3+ orders below phi-1-small floor.** IGLA CODER v0.1 CANNOT train a working model. This is a seed, not a corpus. See report/FINAL_REPORT.md §W-18.
+
+## What v0.2..v0.5 must add (roadmap)
+
+- v0.2: populate held-out-eval, rerun decontam bidirectionally, publish decontam_report_v02.md.
+- v0.3: apply WP-10 deterministic alpha-rename augmentation to reach ~20K train tokens.
+- v0.4: add tokenizer spec `specs/tokenizer/tri_tokenizer.tri`.
+- v0.5: reach >=200K train tokens (deficit <=262x per WP-10 analog); actual train-box run authorised.
+
+## Files
+
+- `MANIFEST.json` — canonical index (schema in `specs/dataset/igla_coder_manifest.tri`)
+- `DECONTAM.md` — decontamination policy and v0.1 status
+- `pairs/<pair_id>/spec.tri` — the .tri spec
+- `pairs/<pair_id>/metadata.json` — per-pair metadata (sha256, ring_id, target_lang, license)
+- `held-out-eval/` — v0.1: empty by design
+
+## License
+
+Each pair inherits its source repo's license, recorded per-pair in `metadata.json`.
+
+- ring-105-001..004 + 007: Apache-2.0 (from gHashTag/t27)
+- ring-105-005, 105-006: Apache-2.0 (anonymised abstractions of PRIVATE 999-multibots-* repos; NO secrets, NO PII, NO wallet, NO tokens in this dataset)
+
+## SHA-256 reproducibility
+
+All `spec_sha256` values in MANIFEST.json are computed with Python stdlib `hashlib.sha256(open(path,"rb").read())` at build time. Re-running `compute_manifest.py` in this workspace must yield identical hex digests.
+
+## Not for training as-is
+
+**Do not attempt to train IGLA CODER on this dataset.** It is a seed for architecture and pipeline validation, not for capability. Real training gates on Wave-5.
@@ -0,0 +1,16 @@
+{
+  "pair_id": "0001-experience-format",
+  "spec_path_in_t27": "specs/organism/experience.tri",
+  "spec_local_path": "pairs/0001-experience-format/spec.tri",
+  "spec_sha256": "ea43af76d3c1214721cd40dbc624062cc17707525c55c824612e3941b3c3c7ef",
+  "spec_bytes": 4096,
+  "gen_path_in_t27": null,
+  "gen_sha256": null,
+  "target_lang": "None",
+  "ring_id": "ring-105-001",
+  "license": "Apache-2.0",
+  "decontam_status": "Skipped",
+  "codegen_status": "spec_only",
+  "l3_ascii_pass": true,
+  "l3_non_ascii_bytes": 0
+}