AMF-G1-4B-Experimental

AMF G1 — 4B Experimental

First learned neural artifact fabricated by AI Mission Foundry (AMF).

Created and developed by Fayçal Benahmed
Stack Moderne — France

G1 is not published as a qualified production model.
It is published as evidence of the AMF fabrication process that produced it.

Status

What is G1?

G1 is the first learned neural model artifact produced during the early AMF model-fabrication experiments.

Its exact upstream substrate is:

Qwen/Qwen3-4B-Instruct-2507
revision: cdbee75f17c01a7cc42f958dc650907174af0554

AMF applied:

supervised LoRA training
rank = 16
learning rate = 4e-5
steps = 72
seed = 190901
followed by a safe merge

The result is a materialized derived neural checkpoint.

Important experimental limitation

G1 was produced before the intended AMF autonomous optimization and REPAIR loop was correctly implemented.

RC1 effectively executed:

MEASURE
  -> IDENTIFY CAPABILITY GAP
  -> FABRICATE G1
  -> REMEASURE
  -> FAIL
  -> STOP

The intended Foundry loop is:

MEASURE
  -> DIAGNOSE
  -> FABRICATE / COMPOSE
  -> REMEASURE
  -> ANALYZE RESIDUAL FAILURE
  -> CHOOSE REPAIR LOCUS
  -> REPAIR
  -> REMEASURE
  -> repeat until PASS or justified refusal

Therefore G1 should be interpreted as:

the first fabricated candidate from an incomplete optimization loop,
not the final output of a completed AMF search.

Its measured failure is part of the research record.

Qualification result

The sealed G1 DEV measurement produced:

quality_state: FAIL_OBSERVED
qualified_output: 0.3472
latency_p95_s: 4.4042

Therefore:

MODEL_CAPABILITY_QUALIFIED = NO

AMF does not convert a failed qualification into a marketing PASS.

Known observed limitation

An original smoke test produced an exact-surface fidelity error:

evidence: Sarah Klein
output:   sarah klien

This observation contributed to the later design of evidence-grounded diagnosis and REPAIR.

Local deployment derivative

The canonical BF16 model was transported to a consumer laptop and executed locally under CPU-only inference.

A GGUF Q5_K_M derivative was then created.

Measured Q5 deployment:

parameters:             4.02B
model size:             2.69 GiB
CPU:                    AMD Ryzen 5 8540U
threads:                6
prompt throughput:      11.73 +/- 0.35 tokens/s
generation throughput:  5.68 +/- 1.17 tokens/s
peak RSS:               approximately 3.04 GiB
swap:                   0

Local execution:

G1_Q5_LOCAL_EXECUTION = PASS

Local behavior gate

A separate deployment-oriented micro-evaluation tested:

SUPPORTED_FACTS
MISSING_EVIDENCE
CONTRADICTION
STRICT_SCHEMA
EXACT_EVIDENCE_FIDELITY

Each probe was repeated three times.

Result:

exact passes:            15 / 15
exact pass rate:         1.000
exact evidence fidelity: PASS
local behavior gate:     PASS

This local gate is not the sealed AMF DEV/ROB/OOD qualification suite and does not change G1's qualification status.

The local gate also used a newly defined protocol rather than an exact replay of the historical A100 smoke prompts.

Artifact identity

Canonical G1 transport archive SHA256:

413ef141957d3dc4e3ec431a02bc37f4e02b0649c8d0cb61883de385ab86cb76

Internal G1 artifact SHA256:

9ee32a40e2aec3fe329fcef7f4c7fbe11444781c28fa14660d735df5ef7cebfa

GGUF Q5_K_M SHA256:

9cb29f9b6c1fffc6d8bf749c15f25b050c2c85569f5abfc7a888e099ea5e56a2

Training dataset receipt SHA256:

666462534451178582d29eb2214697bc9dbc1598bcf27b843b438f64ee2f0182

The immutable RC1 dataset receipt records:

split = train
final_access = NONE

This receipt establishes the identity of the historical training dataset. The historical receipt records final_access = NONE, consistent with the campaign rule that final data does not participate in search, training or fabrication. It does not, by itself, establish the dataset's origin, authorship, license, collection method, consent status, personal-data status, or suitability for a particular legal or production context.

No additional provenance or compliance claim is made by this release. See DATA_PROVENANCE_NOTICE.md for the complete public claim boundary and the versioned remediation path.

Why publish G1?

G1 is not presented as a state-of-the-art model.

It records a concrete transition in AMF research:

existing intelligence
  -> capability measurement
  -> identified gap
  -> learned fabrication
  -> materialized model
  -> measured failure
  -> local deployment transformation
  -> evidence for future REPAIR

Its failure is deliberately preserved.

AMF Agent is a separate public artifact demonstrating a system that passed its defined qualification gates.

G1 demonstrates that AMF can also fabricate learned neural material.

AMF

AI Mission Foundry

Mission + Constraints
        |
        v
       AMF
        |
        v
Qualified Intelligence Realization
    or qualified refusal

AMF treats models as raw material, capabilities as qualified components, and systems as architectures engineered for a mission.

Attribution

AMF — AI Mission Foundry
Created and developed by Fayçal Benahmed
Stack Moderne — France
Independent research and engineering project
https://stack-moderne.fr/

License

G1 is derived from Qwen/Qwen3-4B-Instruct-2507, which is distributed under the Apache License 2.0.

Third-party software and components remain governed by their respective licenses.