🔴 fake_quant.rs broken for ≤8-bit formats (R7 falsified by TIER1-posit8)

[gHashTag]

R7 Falsification Witness — fake_quant.rs broken for ≤8-bit formats

Anchor: φ² + φ⁻² = 3 · PASS-19 · 2026-05-15

Evidence (R5-verified, fresh SQL probe on phd-postgres-ssot @ 2026-05-15 09:58 UTC)

11 of 12 sub-16-bit format/algo combinations produce bpb ≈ log₂(128) = 7.0 — i.e., the model does not learn at all. Vocab size is 128 → log₂(128) = 7.0 is the uniform-distribution baseline.

Format	Algo	Best BPB	Status
posit8	adamw	6.9903	💀 R7 witness (IGLA-TIER1-posit8-h128-LR0.0001-rng1597-adamw)
posit8	muon	4.8115	🔴 partial learn
int8	adamw	6.9926	💀
int8	muon	5.4015	🔴 partial
int4	adamw	6.9703	💀
int4	muon	2.8440	🟡 (muon learns!)
uint8	adamw	7.0000	💀 exact log₂(vocab)
gf32	adamw	6.9610	💀
gf8	adamw	7.0000	💀
gf4	muon	7.0124	💀
mxfp8	muon	6.9780	💀
nf4	adamw/muon	6.9860 / 7.0552	💀

Meanwhile 16-bit formats train cleanly (gf256 adamw → 2.5719, posit16 muon → 2.6251, fp16 adamw → 2.6691).

Hypothesis (R12 Lee/GVSU style)

crates/trios-train/src/fake_quant.rs:449-493 — the dequantization path for bits ≤ 8 either:

1. Doesn't dequant before the residual stream (output stays in integer domain), OR
2. Initializes scale/zero-point with a value that collapses gradients to zero, OR
3. Has a wrong STE (straight-through estimator) — muon happens to bypass it (works for int4), but adamw cannot.

The asymmetry adamw=dead / muon=partial for int4/int8/posit8 is the smoking gun: gradient flow differs between optimizers only when STE is broken.

Falsification protocol (R7)

Cluster reference (the working 14 lanes from same fleet, identical hyperparams except format):

• IGLA-DEEP-CHAMPION-gf256-h384-LR0.0001-rng1597-adamw → 2.5719 (962 rows)
• IGLA-MULTI-SEED-posit16-h128-LR0.0001-rng2584-muon → 2.6251 (58 rows)

All 16-bit lanes learn → infrastructure is fine. Only ≤8-bit lanes fail → fake_quant.rs is broken for low-bit path.

Repro

gh workflow run tier1-explore-acc13.yml --repo gHashTag/trios-railway -f confirm=PHI
# wait 20 min, then:
psql $DATABASE_URL -c "SELECT canon_name, MIN(bpb) FROM ssot.bpb_samples WHERE canon_name LIKE 'IGLA-TIER1-%' GROUP BY canon_name;"

Suggested fix (action items)

• Add unit test in crates/trios-train/tests/fake_quant_test.rs that asserts dequant(quant(x)) round-trips within eps for all formats: int4, int8, uint8, posit8, gf4, gf8, gf32, nf4, mxfp8.
• Add property test: bpb after 100 steps on tiny shakespeare with vocab=128 must be < 6.5 for all formats. Fail CI if any format ≥ 6.9.
• Audit bits ≤ 8 branch of fake_quant.rs:449-493 — likely a missing .mul(scale).add(zero_point) after the int cast.
• After fix, dispatch new TIER1-LOWBIT wave on Railway ACC1+ACC3 to verify all ≤8-bit formats reach bpb < 3.0.

Refs

• R7 witness lane: IGLA-TIER1-posit8-h128-LR0.0001-rng1597-adamw @ 64 rows @ bpb=6.9903
• Throne 👑 TRINITY HIVE — Queen's Registry & ONE SHOT Dispatch #264 PASS-19: 👑 TRINITY HIVE — Queen's Registry & ONE SHOT Dispatch #264 (comment)
• Workflow run: https://github.com/gHashTag/trios-railway/actions/runs/25906740217

---

φ² + φ⁻² = 3 · TRINITY · PASS-19 · DEFENSE 2026-06-15