# Mout CPU Research v1 — final report

## Executive conclusion

Research v1 is complete. It produced reproducible format, quantization, native
GGML, parity, transport, and same-DLL benchmarking infrastructure, but no
custom dense or mixed format met the simultaneous quality, size, and AVX2
runtime bar. This is a valid architecture result, not a project failure.

## Master decision table

Projected rows use the AAQ-policy dense-frontier convention; they are not
interchangeable with standalone GGUF artifact bytes.

| format | ATTENTION_Q BPW | whole BPW / MiB | quality metric | dev error | Q4 gap closed | runtime evidence | Q4-relative speed | complexity | final status / reason |
|---|---:|---|---|---:|---:|---|---|---|---|
| Q4 production | 4.5000 | standalone 469,081,312 bytes | frozen Q4 reference | 14,421.4029 | 100% | Release tg128 54.633 t/s | reference | LOW | PRODUCTION_REFERENCE |
| Type45 / AAQ | 2.5000 | projected 4.9240 / 471.60 MiB; actual 494,512,352 bytes | compact base | 135,875.1167 | 0% | Release tg128 58.280 t/s | >Q4 decode in retained benchmark | MEDIUM | RETAIN_RESEARCH_BASE; high-coverage quality insufficient |
| R2 | 2.7367 | projected 4.947167 / 496.26 MiB | eval-v3 40/100; Q4-valid 31/35 | — | 66.3438% local recovery | tg128 51.943 t/s | 95.1% | HIGH | FROZEN; complexity and one-task gate miss |
| R3 | — | — | granular refinement | — | weak incremental evidence | no justified replacement | — | HIGH | REJECTED |
| Family A | 3.5000 | projected 5.0219 / 478.61 MiB | packed dev | 50,987.2917 | 69.8931% | exact native type 46; n128 30.411 ns | 34.37% block speed | MEDIUM | FROZEN_RUNTIME_MISMATCH |
| A4-G128-SYM | 4.1250 | projected 5.0830 / 482.98 MiB | C-packed replay | 22,290.1296 | 93.5212% | exact private AVX2; n128 21.103 ns, n1024 164.795 ns | 42.83% / 43.85% | MEDIUM | FROZEN_RUNTIME_MISMATCH |

## Chronological evidence and retained artifacts

| stage | authoritative path | role | SHA256 | retention |
|---|---|---|---|---|
| v0.8 | experiments/mout-v2-r2-full-model-integration-v08/MOUT_R2_FULL_MODEL_INTEGRATION_V08_REPORT.md | canonical R2 integration/eval | 345F00E3C7AC703EB797F45388A06BC8811028C10517F48ACA3B963F36EC2529 | KEEP |
| v0.9 | experiments/mout-v2-r2-runtime-quality-repair-v09/MOUT_R2_RUNTIME_QUALITY_REPAIR_V09_REPORT.md | runtime economics | 4192D187144AB7796774E79FEADA7481C4B6E45CE407D31C258F516B7803A428 | KEEP |
| v0.10 | experiments/mout-v2-r2-closure-next-format-v010/MOUT_R2_CLOSURE_NEXT_FORMAT_V010_REPORT.md | R2/R3 closure | B18556A045AFB60D3819D55C82A1588B13C71CB49FFDD1868964C594B051CD63 | KEEP |
| v0.11 | experiments/mout-v011-output-sensitivity-gate/MOUT_V011_OUTPUT_SENSITIVITY_REPORT.md | output-metric alignment | report preserved | KEEP |
| v0.12 | experiments/mout-v012-base-quant-architecture-decision/MOUT_V012_BASE_QUANT_ARCHITECTURE_DECISION.md | dense frontier | report preserved | KEEP |
| v0.13–17 | experiments/mout-v013... through mout-v017... reports | Family-A specification, kernel, integration, repair | reports preserved | KEEP |
| v0.18 | experiments/mout-v018-nibble-dense-format-screen/MOUT_V018_NIBBLE_DENSE_FORMAT_SCREEN.md | nibble screen | report + metrics hash in manifest | KEEP |
| v0.19 | experiments/mout-v019-a4-g128-format-kernel-gate/MOUT_V019_A4_G128_FORMAT_KERNEL_GATE.md | packed reference/kernel gate | report + source hash in manifest | KEEP |

## Patch inventory and repository decision

- Type 43 (MOUT_W2_128): historical diagnostic implementation; B/C, retain only
  on the research branch/tag.
- Type 44 (MOUT_W2_G32_C4): historical pre-TP implementation; B/C, research only.
- Type 45 (MOUT_W2_G32_C4_TP): retained diagnostic/research infrastructure and
  AAQ reference; E.
- Type 46 (MOUT_A3_G32 / Family A): B, exact but non-winning implementation.
  Recommended strategy: preserve the exact v1 research branch/tag, then remove
  type 46 and Family-A code in a later cleanup commit of the production-facing
  branch; do not erase artifacts.
- R2 model/graph/CPU routing in llama-model/llama-graph and CPU x86: B/E,
  preserve branch/tag and parity harnesses; cleanup candidate after tag.
- x86 tile2 and v0.19 production bridges: C/D, private diagnostic exports only;
  remove in the later cleanup commit after reproducibility tag.
- quant/dequant/type traits in ggml-common.h, ggml-quants.c, ggml-cpu/quants.c,
  ggml.c, ggml.h: mixed A/B/E; retain until tag, then split retained Type45
  infrastructure from abandoned types.

## Reproducibility smoke

A clean Release ggml-cpu rebuild passed. The retained scheduled Family-A
normal-MUL_MAT parity executable passed all RHS 1/2/32/128 and threads 1/2/4/6
cases (41,728 exact outputs). Q4/type45, Family-A, and A4 exact ladders remain
authoritatively recorded in v0.16 and v0.19; the million/random suites were
not unnecessarily repeated in this freeze stage.

## Scientific scorecard

No custom format beats or matches Q4 on the full target. Useful technology was
still produced: calibrated AAQ, low-bit GGML types, R2 mixed routing, frozen
format specs, byte-exact scalar/AVX2 validation, native integration, same-DLL
benchmarking, and corrected frozen-eval transport.

The quality bottleneck is aggressive high-coverage low-bit ATTENTION_Q.
The runtime bottleneck is not merely bit decode: Family-A predecoded dot was
16.220 ns versus Q4 10.452 ns, and A4 predecoded was 18.112 ns versus Q4
9.039 ns, because Q8_0 retains four activation-scale regions and exact
integer/FP accumulation costs. Mixed sidecars add routing and architecture
complexity without a v1-quality/runtime win.

## Research v2 hypotheses (not implementations)

| hypothesis | upside | cost | risk |
|---|---|---|---|
| training-time architecture co-design for CPU-friendly quantization | HIGH | HIGH | HIGH |
| activation-format co-design rather than fixed Q8_0 compatibility | HIGH | HIGH | HIGH |
| ISA-specific format for VNNI/AVX-512/AMX-class hardware | MEDIUM-HIGH | HIGH | MEDIUM-HIGH |

MOUT_CPU_RESEARCH_V1_STATUS: COMPLETE

CUSTOM_FORMAT_PRODUCTION_WINNER:
NONE

BEST_QUALITY_SIZE_RESEARCH_FORMAT:
A4-G128-SYM

BEST_A4_PACKED_Q4_GAP_CLOSED:
93.5212%

BEST_A4_ATTENTION_Q_BPW:
4.1250

BEST_A4_PRODUCTION_Q4_RELATIVE_SPEED_N128:
42.83%

BEST_A4_PRODUCTION_Q4_RELATIVE_SPEED_N1024:
43.85%

FAMILYA_STATUS:
FROZEN_RUNTIME_MISMATCH

A4_G128_STATUS:
FROZEN_RUNTIME_MISMATCH

Q4_REFERENCE_STATUS:
PRODUCTION_REFERENCE

NEW_DENSE_FORMAT_SEARCH:
FROZEN

RESEARCH_V1_REPRODUCIBLE:
YES

SOURCE_CLEANUP_READY:
YES

PRIVATE_REPO_READY:
YES

RECOMMENDED_NEXT_PROGRAM:
RESEARCH_V2

NEXT_STEP:
Create a private repository from this frozen evidence, tag the exact research branch, then perform only the separately planned cleanup commit.