…fication 0.26× FALSIFIED, but std-ratio measurement is 50× more sensitive

M96/M97/M98 falsified A1, A2, A3 (per-tensor synthetic amplifiers).
This PR closes the synthetic-amplifier landscape by testing A4
(multi-token batch dimension).

Authors `falsify_ffn_gguf_013_multi_token_batch_amplification`.
Test: 7-token batch (B=7); 5 chained matvecs (256×256 each) PER
TOKEN with RMSNorm between layers. Reports per-token rel_diff AND
batch-std-ratio (mimicking §27 measurement).

EMPIRICAL RESULT (2026-05-07):
  per-token rel_diffs:
    token[0]: 0.439143%
    token[1]: 0.246297%
    token[2]: 0.020914%
    token[3]: 0.028250%
    token[4]: 0.023573%
    token[5]: 0.024674%
    token[6]: 0.020246%
  mean per-token rel_diff: 0.114728%
  variance_across_tokens:  21.69×

  Batch-dimension std (mimics §27 measurement):
    Path A mean std (across batch): 0.033416
    Path B mean std (across batch): 0.032228
    std-ratio deviation from 1.0:   3.69%

  multi_token_amplification = 0.2613×  ← COMPRESSES vs single-token

A4 SYNTHETIC AMPLIFICATION FALSIFIED (0.26× < 1×).

**HOWEVER, A SECONDARY FINDING THAT WAS NOT PREDICTED**: the §27-
comparable measurement (std across batch) shows 3.69% deviation
from 1.0 between Path A and Path B — that is **50× the per-tensor
0.077% baseline**. The std-ratio MEASUREMENT amplifies M94 mechanism
by ~50× over per-tensor rel_diff.

REFINED §27 MAGNITUDE EXPLANATION:

  M94 mechanism × M95 compounding × M99 batch-std-amplification
  = 0.077% × 5.70× × 50× ≈ 22% drift (synthetic upper bound)

§27 measured = 1723% drift = ~78× the synthetic upper bound. A
78× residual gap is still unexplained, but is **DRAMATICALLY closer
to feasible than the prior 3920× gap** (M98 closing).

The std-ratio finding is LOAD-BEARING for the SHIP-007 §22 fix scope
analysis: the choice between Option-A (PROMOTE GGUF-PATH semantics
into APR forward) and Option-B (PROMOTE APR-PATH semantics into
GGUF forward) hinges on whether the std-ratio measurement is a
real signal of layer-level divergence or an artifact of batch-
dimension noise. M99 confirms it's real signal.

POSSIBLE EXPLANATION FOR REMAINING 78× GAP:
- A5 (Real-weight non-uniformity): synthetic uniform weights may
  produce 5-10× smaller rel_diff than real Qwen weights
- A6 (RMSNorm rsqrt): real RMSNorm interacts with per-token drift
  via 1/sqrt(σ²) non-linearly
- Cumulative-layer interaction: §27 is layer-3 (3 layers deep);
  M99 was 5 chained matvecs OF THE SAME WEIGHT (different layers
  have different weight distributions)

AMPLIFIER LANDSCAPE POST-A1+A2+A3+A4 FALSIFICATION:
- A1 (RoPE phase)            — FALSIFIED ✗ (1.00×)
- A2 (Softmax saturation)    — FALSIFIED ✗ (0.01×)
- A3 (Block-scale variance)  — FALSIFIED ✗ (1.00×)
- A4 (Multi-token batch)     — FALSIFIED ✗ (0.26× per-token,
                                50× std-ratio sensitivity)
- A5 (Real-weight non-uniformity) — UNTESTED, real-teacher gated
- A6 (RMSNorm rsqrt approx)  — UNTESTED, real-teacher gated

ALL SYNTHETIC amplifier candidates exhausted. M-FFN-GGUF-6
(real-teacher) is now THE ONLY remaining test for the 78× residual.

Contract trace-ffn-sub-block-gguf-v1 v1.9.0 → v1.10.0:
- FALSIFY-FFN-GGUF-013 NEW → DISCHARGED
- M-FFN-GGUF-4 step (i) A4 candidate: NEW → DISCHARGED
- All four synthetic amplifiers DISCHARGED
- M-FFN-GGUF-6 (real-teacher): now THE ONLY remaining synthetic-falsifier

Stacked atop M98 (PR #1542). Will rebase on main after #1542 merges.

Test runs locally:
  cargo test -p aprender-serve --lib falsify_ffn_gguf_013 -- --nocapture
  test result: ok. 1 passed; finished in 0.06s

Refs PMAT-CCPA, SHIP-007 §22, FALSIFY-FFN-GGUF-013.

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

…fication 0.26× FALSIFIED, but std-ratio measurement is 50× more sensitive (#1543)

M96/M97/M98 falsified A1, A2, A3 (per-tensor synthetic amplifiers).
This PR closes the synthetic-amplifier landscape by testing A4
(multi-token batch dimension).

Authors `falsify_ffn_gguf_013_multi_token_batch_amplification`.
Test: 7-token batch (B=7); 5 chained matvecs (256×256 each) PER
TOKEN with RMSNorm between layers. Reports per-token rel_diff AND
batch-std-ratio (mimicking §27 measurement).

EMPIRICAL RESULT (2026-05-07):
  per-token rel_diffs:
    token[0]: 0.439143%
    token[1]: 0.246297%
    token[2]: 0.020914%
    token[3]: 0.028250%
    token[4]: 0.023573%
    token[5]: 0.024674%
    token[6]: 0.020246%
  mean per-token rel_diff: 0.114728%
  variance_across_tokens:  21.69×

  Batch-dimension std (mimics §27 measurement):
    Path A mean std (across batch): 0.033416
    Path B mean std (across batch): 0.032228
    std-ratio deviation from 1.0:   3.69%

  multi_token_amplification = 0.2613×  ← COMPRESSES vs single-token

A4 SYNTHETIC AMPLIFICATION FALSIFIED (0.26× < 1×).

**HOWEVER, A SECONDARY FINDING THAT WAS NOT PREDICTED**: the §27-
comparable measurement (std across batch) shows 3.69% deviation
from 1.0 between Path A and Path B — that is **50× the per-tensor
0.077% baseline**. The std-ratio MEASUREMENT amplifies M94 mechanism
by ~50× over per-tensor rel_diff.

REFINED §27 MAGNITUDE EXPLANATION:

  M94 mechanism × M95 compounding × M99 batch-std-amplification
  = 0.077% × 5.70× × 50× ≈ 22% drift (synthetic upper bound)

§27 measured = 1723% drift = ~78× the synthetic upper bound. A
78× residual gap is still unexplained, but is **DRAMATICALLY closer
to feasible than the prior 3920× gap** (M98 closing).

The std-ratio finding is LOAD-BEARING for the SHIP-007 §22 fix scope
analysis: the choice between Option-A (PROMOTE GGUF-PATH semantics
into APR forward) and Option-B (PROMOTE APR-PATH semantics into
GGUF forward) hinges on whether the std-ratio measurement is a
real signal of layer-level divergence or an artifact of batch-
dimension noise. M99 confirms it's real signal.

POSSIBLE EXPLANATION FOR REMAINING 78× GAP:
- A5 (Real-weight non-uniformity): synthetic uniform weights may
  produce 5-10× smaller rel_diff than real Qwen weights
- A6 (RMSNorm rsqrt): real RMSNorm interacts with per-token drift
  via 1/sqrt(σ²) non-linearly
- Cumulative-layer interaction: §27 is layer-3 (3 layers deep);
  M99 was 5 chained matvecs OF THE SAME WEIGHT (different layers
  have different weight distributions)

AMPLIFIER LANDSCAPE POST-A1+A2+A3+A4 FALSIFICATION:
- A1 (RoPE phase)            — FALSIFIED ✗ (1.00×)
- A2 (Softmax saturation)    — FALSIFIED ✗ (0.01×)
- A3 (Block-scale variance)  — FALSIFIED ✗ (1.00×)
- A4 (Multi-token batch)     — FALSIFIED ✗ (0.26× per-token,
                                50× std-ratio sensitivity)
- A5 (Real-weight non-uniformity) — UNTESTED, real-teacher gated
- A6 (RMSNorm rsqrt approx)  — UNTESTED, real-teacher gated

ALL SYNTHETIC amplifier candidates exhausted. M-FFN-GGUF-6
(real-teacher) is now THE ONLY remaining test for the 78× residual.

Contract trace-ffn-sub-block-gguf-v1 v1.9.0 → v1.10.0:
- FALSIFY-FFN-GGUF-013 NEW → DISCHARGED
- M-FFN-GGUF-4 step (i) A4 candidate: NEW → DISCHARGED
- All four synthetic amplifiers DISCHARGED
- M-FFN-GGUF-6 (real-teacher): now THE ONLY remaining synthetic-falsifier

Stacked atop M98 (PR #1542). Will rebase on main after #1542 merges.

Test runs locally:
  cargo test -p aprender-serve --lib falsify_ffn_gguf_013 -- --nocapture
  test result: ok. 1 passed; finished in 0.06s

Refs PMAT-CCPA, SHIP-007 §22, FALSIFY-FFN-GGUF-013.

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>