spec: canonical canon_name format \u2014 mandatory numeric format token

[gHashTag]

Canonical canon_name spec (per operator instruction 2026-05-02 + trios#445)

Format

IGLA-{LANE}-{format}-h{H}-LR{LR4}-rng{SEED}-{TAG}

Token	Source	Example
IGLA-	mandatory prefix	IGLA-
{LANE}	wave + phase	WAVE-GF-001-PHASE1, BLITZ-T10H, CHAMPION-acc0new, RAILWAY-LEADER
{format}	MANDATORY numeric format token	binary32, binary16, bfloat16, TF32, FP8-E4M3, FP8-E5M2, GF4, GF8, GF16, GF32, Posit8, Posit16, Posit32, INT4, INT8, INT16, UINT8, NF4, MXFP8, MXFP6, MXFP4, AFP, LNS, ...
h{H}	hidden dim	h384, h828
LR{LR4}	LR \u00d7 10000 zero-padded 4-digit	LR0030 (lr=0.0030), LR0573 (lr=0.0573)
rng{SEED}	seed	rng34, rng1597
{TAG}	optional experiment tag	k3 (phi-LR ladder index), step27000, t55803-001

Canonical numeric format catalog (mandatory)

binary16, binary32, binary64, binary128, binary256, decimal32, decimal64, decimal128, FP80, double-double, quad-double, bfloat16, TF32, FP8-E4M3, FP8-E5M2, FP6-E3M2, FP6-E2M3, FP4-E2M1, MXFP8, MXFP6, MXFP4, NF4, AFP, Posit8, Posit16, Posit32, Posit64, LNS, GF4, GF8, GF12, GF16, GF20, GF24, GF32, GF64, INT4, INT8, INT16, INT32, INT64, INT128, UINT4, UINT8, UINT16, UINT32, UINT64, UINT128, Q-format, BCD, IBM-HFP, MBF, VAX-F, VAX-D, VAX-G, VAX-H, Cray-float, minifloat, Unum-I, Unum-II, tapered-fp, block-fp, shared-exp, stochastic-rounding

Mapping from config_json.trainer.format token

config token	canon token
fp16, f16	binary16
fp32, f32	binary32
fp64, f64	binary64
bf16	bfloat16
tf32	TF32
fp8_e4m3	FP8-E4M3
fp8_e5m2	FP8-E5M2
gf16	GF16
posit32	Posit32
... (full list above)

Examples

• IGLA-WAVE-GF-001-PHASE1-binary32-h384-LR0573-rng34-k3 \u2014 phi-LR ladder, fp32 baseline, k=3
• IGLA-WAVE-GF-001-PHASE4-GF16-h384-LR0040-rng34 \u2014 Phase-4 GF16 dtype sweep
• IGLA-CHAMPION-binary32-h828-LR0030-rng43-step27000 \u2014 champion repro
• IGLA-BLITZ-T10H-binary32-h256-LR0020-rng45-t55803-001 \u2014 legacy BLITZ wave (canon was missing format token; future renames should add it)

Rationale

The numeric format is the primary independent variable in the IGLA RACE / WAVE-GF-001 / Universal Numeric-Format Catalog work (zig-golden-float#69). A canon_name without the format token is meaningless because phi-LR ladder convergence, BPB, and Pareto-frontier comparisons are all keyed by format.

Forbidden

• Names without an explicit format token: GF-LR-k3-rng34, SMOKE-PR82-FULLCYCLE-rng34, DIRECT-SMOKE-PR86-rng34, VERIFY-PR88-FRESH-rng99, trios-train-rng43
• Lowercase format tokens: fp32, bf16 \u2014 must be binary32, bfloat16
• Ad-hoc tokens not in the canonical catalog above
• Truncated names (IGLA-rng34)

Enforcement

• CI \u2014 strategy_queue insert trigger validates canon_name regex ^IGLA-[A-Z0-9-]+-({CANON_FMT_RE})-h\d+-LR\d{4}-rng\d+(-[a-z0-9]+)*$
• scarab \u2014 reject claim if canon_name does not match regex (logs and skips)
• Operator \u2014 retroactive rename of any non-canonical rows in bpb_samples + strategy_queue

Retroactive cleanup applied 2026-05-02

Old	New
GF-LR-k3-rng34	IGLA-WAVE-GF-001-PHASE1-binary32-h384-LR0573-rng34-k3
(\u00d7 9 phi-LR rows in strategy_queue 2062..2070)
80 bpb_samples rows renamed
26 SMOKE/VERIFY/DIRECT rows purged

References

• trios#445 \u2014 IGLA RACE 6-account Railway cycle
• trios#446 \u2014 EPIC E2E TTT
• zig-golden-float#69 \u2014 Universal Numeric-Format Catalog (60+ formats)
• trios-trainer-igla#84 \u2014 telemetry restoration

\ud83c\udf3b phi^2 + phi^-2 = 3 \u00b7 TRINITY \u00b7 NEVER STOP

[gHashTag]

✅ OPERATOR RATIFIED — 2026-05-09 (Wave 26)

Operator (Dmitrii Vasilev @gHashTag) has explicitly ratified this spec on 2026-05-09 with the directive:

"да!! все так!! и чтобы по другому не могло быть!"
(English: "yes!! exactly so!! and it must not be possible any other way!")

This means:

1. NO exceptions — every row in public.bpb_samples, ssot.bpb_samples, strategy_queue, gardener_runs, experiment_queue MUST have canon_name matching the regex below.
2. NO ad-hoc names — phase1-rng47, trios-train-rng43, cpu_train_gf16_adamw, IGLA-GF16-3-rng47, IGLA-TRAIN-igla_train-rng89, GF-LR-k0-rng34 are ALL retroactively forbidden and must be renamed.
3. Hard-enforced at three layers — DB CHECK constraint, scarab claim-time validator, and CI lint on every INSERT INTO strategy_queue.

---

Canonical regex (single source of truth)

^IGLA-[A-Z0-9-]+-(binary16|binary32|binary64|binary128|binary256|decimal32|decimal64|decimal128|FP80|double-double|quad-double|bfloat16|TF32|FP8-E4M3|FP8-E5M2|FP6-E3M2|FP6-E2M3|FP4-E2M1|MXFP8|MXFP6|MXFP4|NF4|NF8|AFP|Posit8|Posit16|Posit32|Posit64|LNS|GF4|GF8|GF12|GF16|GF20|GF24|GF32|GF64|INT4|INT8|INT16|INT32|INT64|INT128|UINT4|UINT8|UINT16|UINT32|UINT64|UINT128)-h\d+-LR\d{4}-rng(47|89|123|144)(-[a-z0-9]+)*$

Note: rng group is restricted to L-SEED-CANON #600 ruling seeds {47, 89, 123, 144}. Forbidden seeds {42, 43, 44, 45} are explicitly rejected.

---

Wave 26 enforcement plan

Layer	Gate	Owner	Issue
L1 — DB	CHECK (canon_name ~ '<regex>') on public.bpb_samples, ssot.bpb_samples, strategy_queue	Queen + DBA	this issue
L2 — Mirror trigger	trg_mirror_to_ssot validates regex; rejects with RAISE EXCEPTION if canon malformed	Queen	this issue
L3 — Scarab	Reject claim if strategy_queue.canon_name ∉ regex (log + skip + advance to next row)	trainer-igla agent	trios-trainer-igla#83 follow-up
L4 — CI lint	grep-based pre-commit hook in trios-trainer-igla checks all format!("...") canon constructions	trios-trainer-igla agent	new PR
L5 — Operator runbook	scripts/check_canon.sh returns non-zero if any non-canon row exists in any DB	runbook agent	trios-trainer-igla scripts/

---

Drift inventory (Wave 25d, my own fault — to be cleaned)

Source	Pattern	Count	Action
public.bpb_samples	phase1-rng{47,89,123,144}	~329 rows (Phase-1 services)	RENAME to IGLA-PHASE1-binary32-h384-LR0030-rng{seed}
public.bpb_samples	seed=43, h=828, sha=cd91c45 (ALPHA cohort)	81 rows	DELETE (forbidden seed per #600)
ssot.bpb_samples	mirror of above	~410 rows	follow source: rename or delete
ssot.bpb_samples	cpu_train_f32_adamw (matrix_runner.rs)	1+ rows	RENAME to IGLA-MATRIX-binary32-h384-LR0030-rng47
code	format!("IGLA-GF16-{}-rng{}") ngram_train_gf16.rs	source	PATCH to canon
code	format!("IGLA-TRAIN-igla_train-rng{}") igla_train.rs	source	PATCH to canon
code	format!("cpu_train_{format}_{algo}") matrix_runner.rs:268	source	PATCH to canon
MCP	experiment_queue_insert whitelist contains {42,43,44,45}	seed list	UPDATE to {47,89,123,144}

---

Cross-references

• trios#600 — L-SEED-CANON (seed whitelist {47,89,123,144})
• trios-trainer-igla#84 — original phi-LR canon_name bug (CLOSED)
• trios-trainer-igla#83 — scarab canon_name propagation (still OPEN)
• trios#445 — IGLA RACE 6-account Railway cycle
• trios#446 — EPIC E2E TTT Pipeline (Wave 26 dispatch hub)
• zig-golden-float#69 — Universal Numeric-Format Catalog (60+ formats)

---

Severity bump

This issue is being upgraded from spec-only to P0 / blocker because:

1. Wave-25d backfill drifted into forbidden seed=43 cohort
2. Phase-1 services write phase1-rng{seed} — non-canon
3. matrix_runner.rs writes cpu_train_{f}_{a} — non-canon
4. Without DB-level CHECK there is no way to prevent recurrence

🌻 phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP