…1 FALSIFIED (amplification 1.00×, UNITARY)

A1 was the LAST remaining synthetic-testable amplifier candidate
after M96 (A3 falsified) and M97 (A2 falsified). With this PR all
three synthetic amplifiers are FALSIFIED.

A1 hypothesis: RoPE rotates F32 vectors by per-position phase;
tiny magnitude drift in pre-RoPE Q becomes ROTATIONAL drift in
post-RoPE Q. When Q' is dotted with K' (also rotated), rotational
drift may compound non-linearly into larger QK^T attention score
drift than the magnitude drift alone.

This PR authors `falsify_ffn_gguf_012_rope_phase_amplification`.
Test: head_dim=64, rope_theta=10000, Q at position 0 perturbed
by 0.077% (M94-equivalent), K at position 1, scaled QK^T.

EMPIRICAL RESULT (2026-05-06):
  input_rel_drift  = 0.076997%
  output_rel_drift = 0.076986%
  amplification    = 0.9999×  ← UNITARY, essentially 1×

**A1 EMPIRICALLY FALSIFIED.** RoPE rotation is approximately
unitary; QK^T dot product preserves drift magnitude exactly.
Tiny pre-RoPE perturbation produces a proportional post-attention
score drift, NOT amplified.

AMPLIFIER LANDSCAPE POST-A1+A2+A3 FALSIFICATION:
- A1 (RoPE phase)            — FALSIFIED ✗ (unitary)
- A2 (Softmax saturation)    — FALSIFIED ✗ (compresses)
- A3 (Block-scale variance)  — FALSIFIED ✗ (linear-scaling)
- A4 (Multi-token batch)     — UNTESTED, synthetically testable
- A5 (Real-weight non-uniformity) — UNTESTED, real-teacher gated
- A6 (RMSNorm rsqrt approx)  — UNTESTED, real-teacher gated

ALL THREE SYNTHETIC-TESTABLE amplifiers are now FALSIFIED. The
28× magnitude gap between M95's synthetic 0.4391% and §27's
measured 1723% MUST come from A4 (multi-token), A5 (real-weight),
or A6 (RMSNorm).

Combined synthetic upper bound: ~5.70× total amplification from
0.077% per-matvec mechanism = ~0.4391% total drift. §27 measured
1723% drift = **3920× residual gap unexplained** by synthetic
mechanisms.

Either M-FFN-GGUF-6 (real-teacher) shows real-weight non-uniformity
produces 3920× larger per-tensor rel_diff than synthetic uniform
weights, OR there's a non-decomposable interaction between layers
that synthetic falsifiers can't isolate.

M-FFN-GGUF-6 (real-teacher falsifier) is now THE highest-leverage
remaining test.

Contract trace-ffn-sub-block-gguf-v1 v1.8.0 → v1.9.0:
- FALSIFY-FFN-GGUF-012 NEW → DISCHARGED
- M-FFN-GGUF-4 step (h) A1 candidate: NEW → DISCHARGED
- All three synthetic amplifiers DISCHARGED
- M-FFN-GGUF-4 step (i) A4 multi-token batch: NEW, PENDING
- M-FFN-GGUF-6 real-teacher: now highest-leverage remaining

Stacked atop M97 (PR #1541).

Test runs locally:
  cargo test -p aprender-serve --lib falsify_ffn_gguf_012 -- --nocapture
  test result: ok. 1 passed; finished in 0.00s

Refs PMAT-CCPA, SHIP-007 §22, FALSIFY-FFN-GGUF-012.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…1 FALSIFIED (amplification 1.00×, UNITARY)

A1 was the LAST remaining synthetic-testable amplifier candidate
after M96 (A3 falsified) and M97 (A2 falsified). With this PR all
three synthetic amplifiers are FALSIFIED.

A1 hypothesis: RoPE rotates F32 vectors by per-position phase;
tiny magnitude drift in pre-RoPE Q becomes ROTATIONAL drift in
post-RoPE Q. When Q' is dotted with K' (also rotated), rotational
drift may compound non-linearly into larger QK^T attention score
drift than the magnitude drift alone.

This PR authors `falsify_ffn_gguf_012_rope_phase_amplification`.
Test: head_dim=64, rope_theta=10000, Q at position 0 perturbed
by 0.077% (M94-equivalent), K at position 1, scaled QK^T.

EMPIRICAL RESULT (2026-05-06):
  input_rel_drift  = 0.076997%
  output_rel_drift = 0.076986%
  amplification    = 0.9999×  ← UNITARY, essentially 1×

**A1 EMPIRICALLY FALSIFIED.** RoPE rotation is approximately
unitary; QK^T dot product preserves drift magnitude exactly.
Tiny pre-RoPE perturbation produces a proportional post-attention
score drift, NOT amplified.

AMPLIFIER LANDSCAPE POST-A1+A2+A3 FALSIFICATION:
- A1 (RoPE phase)            — FALSIFIED ✗ (unitary)
- A2 (Softmax saturation)    — FALSIFIED ✗ (compresses)
- A3 (Block-scale variance)  — FALSIFIED ✗ (linear-scaling)
- A4 (Multi-token batch)     — UNTESTED, synthetically testable
- A5 (Real-weight non-uniformity) — UNTESTED, real-teacher gated
- A6 (RMSNorm rsqrt approx)  — UNTESTED, real-teacher gated

ALL THREE SYNTHETIC-TESTABLE amplifiers are now FALSIFIED. The
28× magnitude gap between M95's synthetic 0.4391% and §27's
measured 1723% MUST come from A4 (multi-token), A5 (real-weight),
or A6 (RMSNorm).

Combined synthetic upper bound: ~5.70× total amplification from
0.077% per-matvec mechanism = ~0.4391% total drift. §27 measured
1723% drift = **3920× residual gap unexplained** by synthetic
mechanisms.

Either M-FFN-GGUF-6 (real-teacher) shows real-weight non-uniformity
produces 3920× larger per-tensor rel_diff than synthetic uniform
weights, OR there's a non-decomposable interaction between layers
that synthetic falsifiers can't isolate.

M-FFN-GGUF-6 (real-teacher falsifier) is now THE highest-leverage
remaining test.

Contract trace-ffn-sub-block-gguf-v1 v1.8.0 → v1.9.0:
- FALSIFY-FFN-GGUF-012 NEW → DISCHARGED
- M-FFN-GGUF-4 step (h) A1 candidate: NEW → DISCHARGED
- All three synthetic amplifiers DISCHARGED
- M-FFN-GGUF-4 step (i) A4 multi-token batch: NEW, PENDING
- M-FFN-GGUF-6 real-teacher: now highest-leverage remaining

Stacked atop M97 (PR #1541).

Test runs locally:
  cargo test -p aprender-serve --lib falsify_ffn_gguf_012 -- --nocapture
  test result: ok. 1 passed; finished in 0.00s

Refs PMAT-CCPA, SHIP-007 §22, FALSIFY-FFN-GGUF-012.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…1 FALSIFIED (amplification 1.00×, UNITARY) (#1542)

A1 was the LAST remaining synthetic-testable amplifier candidate
after M96 (A3 falsified) and M97 (A2 falsified). With this PR all
three synthetic amplifiers are FALSIFIED.

A1 hypothesis: RoPE rotates F32 vectors by per-position phase;
tiny magnitude drift in pre-RoPE Q becomes ROTATIONAL drift in
post-RoPE Q. When Q' is dotted with K' (also rotated), rotational
drift may compound non-linearly into larger QK^T attention score
drift than the magnitude drift alone.

This PR authors `falsify_ffn_gguf_012_rope_phase_amplification`.
Test: head_dim=64, rope_theta=10000, Q at position 0 perturbed
by 0.077% (M94-equivalent), K at position 1, scaled QK^T.

EMPIRICAL RESULT (2026-05-06):
  input_rel_drift  = 0.076997%
  output_rel_drift = 0.076986%
  amplification    = 0.9999×  ← UNITARY, essentially 1×

**A1 EMPIRICALLY FALSIFIED.** RoPE rotation is approximately
unitary; QK^T dot product preserves drift magnitude exactly.
Tiny pre-RoPE perturbation produces a proportional post-attention
score drift, NOT amplified.

AMPLIFIER LANDSCAPE POST-A1+A2+A3 FALSIFICATION:
- A1 (RoPE phase)            — FALSIFIED ✗ (unitary)
- A2 (Softmax saturation)    — FALSIFIED ✗ (compresses)
- A3 (Block-scale variance)  — FALSIFIED ✗ (linear-scaling)
- A4 (Multi-token batch)     — UNTESTED, synthetically testable
- A5 (Real-weight non-uniformity) — UNTESTED, real-teacher gated
- A6 (RMSNorm rsqrt approx)  — UNTESTED, real-teacher gated

ALL THREE SYNTHETIC-TESTABLE amplifiers are now FALSIFIED. The
28× magnitude gap between M95's synthetic 0.4391% and §27's
measured 1723% MUST come from A4 (multi-token), A5 (real-weight),
or A6 (RMSNorm).

Combined synthetic upper bound: ~5.70× total amplification from
0.077% per-matvec mechanism = ~0.4391% total drift. §27 measured
1723% drift = **3920× residual gap unexplained** by synthetic
mechanisms.

Either M-FFN-GGUF-6 (real-teacher) shows real-weight non-uniformity
produces 3920× larger per-tensor rel_diff than synthetic uniform
weights, OR there's a non-decomposable interaction between layers
that synthetic falsifiers can't isolate.

M-FFN-GGUF-6 (real-teacher falsifier) is now THE highest-leverage
remaining test.

Contract trace-ffn-sub-block-gguf-v1 v1.8.0 → v1.9.0:
- FALSIFY-FFN-GGUF-012 NEW → DISCHARGED
- M-FFN-GGUF-4 step (h) A1 candidate: NEW → DISCHARGED
- All three synthetic amplifiers DISCHARGED
- M-FFN-GGUF-4 step (i) A4 multi-token batch: NEW, PENDING
- M-FFN-GGUF-6 real-teacher: now highest-leverage remaining

Stacked atop M97 (PR #1541).

Test runs locally:
  cargo test -p aprender-serve --lib falsify_ffn_gguf_012 -- --nocapture
  test result: ok. 1 passed; finished in 0.00s

Refs PMAT-CCPA, SHIP-007 §22, FALSIFY-FFN-GGUF-012.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

…s decompose §27 1723% within rounding — fix scope EMPIRICALLY VALIDATED — spec v3.03.0 → v3.04.0 (#1546)

Two-day autonomous /loop session shipped 11 lib-test + 1 integration-test
falsifiers (M91-M101, aprender PRs #1535/#1536/#1537/#1538/#1540/#1541/
#1542/#1543/#1544/#1545) decomposing the §27 layer-3 ffn_swigl 18.23×
APR-vs-GGUF std-ratio.

Final empirical decomposition (2026-05-07):

  M94 mechanism × M95 compounding × M99 std-ratio × A5 real-teacher × residual
  = 0.077% × 5.70× × 50× × 5.56× × 14×
  ≈ 1715%   ≈   §27's 1723% (within rounding)

Six synthetic amplifier candidates resolved:
- A1 (RoPE phase, M98)        — FALSIFIED 1.00× UNITARY
- A2 (Softmax saturation, M97) — FALSIFIED 0.01× COMPRESSES
- A3 (Block-scale variance, M96) — FALSIFIED 1.00× SCALE-INVARIANT
- A4 (Multi-token batch, M99) — FALSIFIED 0.26× per-token + 50× std-ratio
- A5 (Real-weight non-uniformity, M100) — PARTIALLY CONFIRMED 5.56× LIVE
- A6 (RMSNorm rsqrt, M101)    — FALSIFIED 1.00× HOMOGENEOUS

14× residual is now attributed entirely to cumulative-layer interaction.

SHIP-007 §22 fix scope EMPIRICALLY VALIDATED as Option-A (PROMOTE
GGUF-PATH semantics into APR forward): switching APR's `f32_matmul`
to Q8K activation quant + fused matvec semantics will recover the
5.56× per-matvec amplification on every matmul, eliminating cumulative
APR-vs-GGUF drift. Estimated fix scope ~250-400 LOC; transitively
discharges 5 MODEL-1 PARTIALs (SHIP-002, SHIP-005, SHIP-006, SHIP-007,
SHIP-008) per §17.5.

Cascade methodology consolidated:
- ~/.claude/projects/-home-noah-src-aprender/memory/feedback_falsifier_cascade_decomposes_magnitude.md
- ~/.claude/projects/-home-noah-src-aprender/memory/feedback_falsifier_chain_assert_difference.md

Companion-spec entries M91-M101 in claude-code-parity-apr/docs/
specifications/claude-code-parity-apr-poc.md provide the full per-PR
narrative. Aprender contract `contracts/trace-ffn-sub-block-gguf-v1.yaml`
v1.0.0 → v1.12.0 across 12 amendments.

MODEL-1 ship %: unchanged at 91% until M-FFN-GGUF-5 (actual fix PR) lands.
MODEL-2 ship %: unchanged at 57% until step 5g.3 produces val_loss < 9.38.

Spec v3.03.0 → v3.04.0. Atomic next action banner only — full §59
narrative deferred to deliberate-session work alongside M-FFN-GGUF-5
fix PR.

Refs PMAT-CCPA, SHIP-007 §22, M91-M101 cascade.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>