CoolFace
Modelpublic

faysalbenahmed/AMF-Agent

sourceHugging Faceotherupdated 8d agoView on Hugging Face
0likes
Model Card

AMF Agent — V0.8.0 RC8A

Mission-specific intelligence, qualified through evidence. L'IA vient à vos données, pas l'inverse.

AI Mission Foundry (AMF) is an adaptive system for building, qualifying, selecting, and deploying mission-specific intelligence.

AMF Agent is its first public evidence-backed system release. It is not a fine-tuned Qwen checkpoint, not an adapter, and not a wrapper around a single model. AMF decomposes a mission into capabilities, authorizes execution, evaluates candidate realizations, rejects candidates that fail hard constraints, and executes only the qualified intelligence required for each operation.

No Capability Authorization ⇒ No Execution.

At a glance

ResultAMF Agent RC8A
Publication decisionPUBLICATION_READY
Fresh Final qualified output100.00%
Fresh Final tool safety100.00%
Cost / qualified outcome vs best parent73.55% lower
Qualified throughput vs best parent278.08% higher
Sequential p95 vs best parent45.54% lower
Paired interleaved latencyAMF faster on 96 / 96 rows
Authority executionStrict offline

Work with AMF

Need a qualified, private, deployable AI system for your organization?

AMF turns a mission — plus its quality, safety, latency, cost, memory, privacy, and deployment constraints — into a qualified AI system designed for the target environment.

[Stack Moderne](https://stack-moderne.fr/) Email: contact@stack-moderne.fr Contact: Fayçal Benahmed


Physical result

Fresh Final V8, 96 rows, NVIDIA A100-SXM4-80GB, same-harness comparison against the best final parent.

MetricAMF Agent RC8ABest parent: Qwen3-4B-Instruct-2507
Qualified output100.00%77.08%
Critical OK100.00%82.29%
Tool safety100.00%91.67%
No unauthorized action100.00%94.79%
Model calls / request0.73961.0000
Cost / qualified outcome (GPU-s)0.84593.1981
Qualified throughput / GPU-s1.18220.3127
Sequential p951.9974 s3.6676 s
Resident memory8.0536 GB8.0539 GB

Measured deltas:

  • —73.55% lower cost per qualified outcome
  • —278.08% higher qualified throughput
  • —26.04% fewer model calls per request
  • —45.54% lower sequential p95 latency

Independent interleaved latency authority:

  • —AMF p95: 2.0912 s
  • —parent p95: 3.6821 s
  • —AMF faster on 96 / 96 paired requests
  • —parent faster or equal on 0 / 96 paired requests

These claims are bounded to the included AMF Agent mission contract, protocol, Fresh Final V8 authority, runtime, and hardware evidence. They are not a claim of universal superiority over Qwen or other models outside this evaluated scope.


Substrate competition

AMF did not select Qwen by preference, branding, or static configuration. It emerged as the physical winner of the qualification process.

Three pinned model substrates were physically evaluated under the same capability-qualification harness:

Substrate`TOOL_EXECUTION_ARGUMENTS``DIRECT_RESPONSE`Outcome
Microsoft Phi-4-mini-instructNot qualifiedNot qualifiedRejected
Qwen3-4B-Instruct-2507QualifiedQualifiedSelected
Mistral-7B-Instruct-v0.3Not qualifiedNot qualifiedRejected

Exact pinned revisions:

  • —Phi-4-mini-instruct: cfbefacb99257ffa30c83adab238a50856ac3083
  • —Qwen3-4B-Instruct-2507: cdbee75f17c01a7cc42f958dc650907174af0554
  • —Mistral-7B-Instruct-v0.3: c170c708c41dac9275d15a8fff4eca08d52bab71

“Not qualified” does not mean that a model is generally bad. It means that the tested realization did not satisfy the frozen AMF qualification requirements for that specific capability, mission, interface, and harness.

Qwen3-4B was the only RAW substrate that qualified both required neural capabilities.


What won

text
CR0_ATOMIC_CAPABILITY_PHYSICAL_WINNERS

TOOL_EXECUTION_ARGUMENTS
  -> Qwen/Qwen3-4B-Instruct-2507 RAW
  -> AGENT_EXECUTOR_ARGS_V1

DIRECT_RESPONSE
  -> Qwen/Qwen3-4B-Instruct-2507 RAW
  -> AGENT_DIRECT_RESPONSE_CONTENT_V1

Exact Qwen revision:

text
cdbee75f17c01a7cc42f958dc650907174af0554

The neural model does not generate deterministic Agent fields that AMF already knows. The model is asked only for the atomic, irreducibly uncertain payload; AMF constructs deterministic control and envelope fields itself.


Hard gates first

AMF also evaluated a merged Qwen + LoRA direct-response challenger. It remained semantically qualified, but its deployment footprint violated the frozen memory guardrail.

CandidateResident bytesMemory regressionPre-Final deployment gate
CR0 RAW/RAW8,056,627,200~0%PASS
Merged direct-response16,102,238,208+99.86%FAIL

Frozen guardrail: maximum +35% resident-memory regression.

The merged candidate was rejected before Fresh Final.

Hard gates first. Optimization second.

This is a core AMF property: a candidate does not become deployable merely because it is accurate or fast. Known deployment constraints are authority gates, not post-hoc observations.


AI Mission Foundry

AMF is adaptive at the system-construction level.

text
Mission
  ↓
Capability decomposition
  ↓
Candidate neural / deterministic realizations
  ↓
Capability qualification
  ↓
Hard deployment gates
  ↓
Physical selection
  ↓
SYSTEM_FROZEN
  ↓
Fresh Final authority
  ↓
Independent verification
  ↓
Qualified deployable system

The model is a substrate. The capability is the unit of qualification.

Routing selects capabilities. Qualification selects realizations. Runtime executes only the cheapest qualified intelligence needed.

Atomic neural execution

AMF follows a simple principle:

Models generate only the irreducibly uncertain payload. AMF deterministically constructs everything already known.

For AMF Agent RC8A, that means:

text
Request
   ↓
Deterministic policy / authorization
   ↓
Capability router
   ├── deterministic refusal / early exit
   ├── TOOL_EXECUTION_ARGUMENTS → atomic neural payload
   └── DIRECT_RESPONSE          → atomic neural payload
   ↓
Deterministic Agent envelope
   ↓
Qualified outcome

This reduces unnecessary neural work instead of asking a general-purpose model to regenerate known structure on every request.


Offline authority

RC8A stages exact pinned model snapshots before authority. During the physical campaign:

text
authority_network = FORBIDDEN
model_source_mode = OFFLINE_PINNED_SNAPSHOT

No OpenAI, Anthropic, Gemini, or external inference API is required by the qualified runtime.

This makes private/on-premise deployment possible when the target infrastructure, security controls, licensing, and operational requirements allow it.

Private/offline deployment is an architectural capability, not an automatic legal or regulatory compliance claim.

Publication authority

text
decision             = PUBLICATION_READY
mission_qualified    = true
system_qualified     = true
efficiency           = true
public_value         = true
publication_ready    = true
independent_verify   = PASS
artifact_integrity   = PASS

AMF artifact SHA256:

text
2af41068f1e1540dbee39d12ab6f6d18e9b75386698d9ca9c9b5d2c5eb7ea513

Fresh Final V8 SHA256:

text
ddb426048606fef2dd607a5d0f73e37e990c91af685305faa756dec8f5b50f4d

Physical evidence archive SHA256:

text
fc9926392759084ea6600f30009c86508ad37d64b66fe676ffaedeca425528b4

Artifact boundary

This repository publishes the qualified system specification and physical evidence.

It does not redistribute Qwen base weights and does not claim that AMF trained the winning Qwen weights.

text
mode                  = PINNED_RECONSTRUCTION
base_weights_embedded = false
adapters_embedded     = false

The winning neural substrate remains Qwen/Qwen3-4B-Instruct-2507 at the exact pinned revision above.

AMF Agent is therefore a qualified system release, not a newly trained standalone AMF neural checkpoint.


Evidence

  • —system/ — exact publication-ready system specification and identity receipts
  • —evidence/selected/ — Final qualification, independent verification, champion selection, runtime, and deployment receipts
  • —evidence/raw/ — complete exported RC8A physical evidence archive and SHA256
  • —release/RELEASE_SUMMARY.json — compact public release summary

See:

  • —`QUALIFICATION.md`
  • —`ARCHITECTURE.md`
  • —`REPRODUCE_EVIDENCE.md`

Why AMF exists

AMF is designed for a different question than:

“Which model should we call?”

Its question is:

“Given this mission and these constraints, what is the smallest qualified intelligence that deserves to execute?”

That makes AMF suitable for building expert, qualified, lightweight, deployable AI systems rather than treating a general-purpose model as the entire application.


AMF thesis

Models are substrates. Capabilities are qualified components.
Models generate only the irreducibly uncertain payload. AMF deterministically constructs everything already known.
Hard gates first. Optimization second.
No Capability Authorization ⇒ No Execution.

Attribution and license

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

This repository publishes a qualified system specification and its evidence package. It does not redistribute Qwen base-model weights and does not grant an AMF source-code license unless a separate LICENSE file is added.

The winning Qwen substrate and all third-party components remain governed by their respective licenses and terms. See `NOTICE.md` for the exact release boundary.


Build a qualified AI system for your mission

[Stack Moderne](https://stack-moderne.fr/) Email: contact@stack-moderne.fr Contact: Fayçal Benahmed

AI Mission Foundry — AMF Mission-specific intelligence, qualified through evidence.