🧭 SoT drift: GF format constants disagree across trios-trainer-igla / trios / zig-golden-float

Severity: P1 — data-integrity risk (Wave-3..7 BPB measurements attribute to "format X" with three different meanings of X)
R5 classification: methodology-drift

Summary

The canonical bit layouts for the GoldenFloat family live in gHashTag/zig-golden-float/specs/*.tri and docs/spec-gf16.md. Only GF16 and GF8 are sealed (production + experimental). Everything else is either a proposal in docs/research/gf16_family_phi_layout.md §3.1, or has no spec at all.

Downstream consumers drifted:

GF16

Conflict: bias=15 vs 31 → roundtrip through the local src/gf16.rs vs phi_numbers::gf16 may scale values by 2^16.

GF8

GF32 — three different layouts

Three files, three different GF32. There is no specs/gf32.tri to arbitrate.

GF12 / GF20 / GF24 / GF64 — no SoT at all

Mentioned in:

• Family proposals §3.1 (GF12 proposal 1:5:6,bias=15, GF24 proposal 1:7:16,bias=63)
• fake_quant.rs assigns mantissa=7/12/14/39 for GF12/20/24/64
• phi_numbers/gf64.rs exists with its own layout

No specs/gf{12,20,24,64}.tri file exists.

Wave-3..7 runs recorded BPB against these format labels. The label is ambiguous.

Requested actions

1. Declare the canonical layout authoritatively. For each of GF4/8/12/16/20/24/32/64: either seal a .tri in zig-golden-float/specs/ or mark the format RETIRED.
2. Remove or conform the trios-trainer-igla local copies:
   • Delete src/gf16.rs in favour of FFI into crates/trios-golden-float::GF16
   • Delete src/phi_numbers/gf8.rs / gf32.rs / gf64.rs (or conform EXP_BIAS to the sealed spec)
3. Add a cross-repo conformance CI: generate vectors.json from specs/*.tri and run Rust/Zig/Python/C++/Go implementations against it on every PR.
4. Patch fake_quant.rs (see trios#509-B) to call the canonical phi_numbers::gf*::from_f32/to_f32 instead of an ad-hoc mantissa mask.
5. Freeze claims in phd_format_registry for GF12/20/24/32/64 until their SoT is sealed — cells in phd_format_algo_matrix with these format rows should carry a sot_status='unsealed' flag.

Cites

• PR fix(gf16): reconcile production codec to standard (1+M/512), reclaim wasted mantissa bit #94 (43f817b1) introduced src/fake_quant.rs
• specs/gf16.tri
• specs/gf8.tri
• docs/spec-gf16.md
• docs/research/gf16_family_phi_layout.md §3.1
• trios-trainer-igla/src/gf16.rs l.16
• trios-trainer-igla/src/phi_numbers/gf8.rs l.22
• trios-trainer-igla/src/phi_numbers/gf32.rs l.18-22

phi^2 + phi^-2 = 3 · TRINITY · ONE SOURCE OF TRUTH