faysalbenahmed/AMF-v0.18-Qualified-Realization

AMF v0.18 — Qualified Intelligence Realization

Terminal outcome: QUALIFIED_INTELLIGENCE_REALIZATION
Selected realization: g02-reuse-g01-canonicalize-historical-be
Canonical run: 20260914T081215Z · SUCCEEDED / VALID

A bounded two-generation, multi-locus scientific campaign produced and qualified an exact composed intelligence realization.

This is not a standalone LoRA release and not an external certification claim. The LoRA is one artifact inside the qualified realization.

Explore: Evidence Explorer · AMF G1 4B Experimental · Stack Moderne

Qualification at a glance

Evaluation	Parent accuracy	Qualified accuracy	Gain	Qualified macro-F1	Schema valid
ROB	71.88%	90.63%	+18.75 pts	90.63%	100%
OOD	50.00%	85.16%	+35.16 pts	85.05%	100%
Fresh Final	47.66%	81.25%	+33.59 pts	81.18%	100%

Each role contained 128 evaluated samples. These results are mission-bounded and do not imply universal reliability.

What AMF actually did

AMF did not begin with a model and simply fine-tune it until a score increased.

Across two cumulative scientific generations, the campaign explored multiple intervention loci, including:

• continued LoRA adaptation,
• fresh adaptation,
• reuse,
• representation canonicalization,
• substrate switching.

The representation-canonicalized realization survived fresh research confirmation. In the second generation, a continued-LoRA branch obtained a higher absolute SEARCH_CONFIRM score, but it did not satisfy the recorded incremental continuation-eligibility rule and therefore did not enter the research-confirmed frontier.

The campaign then followed the declared evidence sequence:

research
  ↓
research frontier closed
  ↓
exact winner frozen
  ↓
protected ROB / OOD
  ↓
pre-final gate
  ↓
Fresh Final
  ↓
Qualification Authority
  ↓
QUALIFIED_INTELLIGENCE_REALIZATION

The key point is not that AMF trained one more model. It is that the evidence determined which exact realization should survive.

What exactly was qualified?

The qualified object is the composed realization:

Pinned Mistral-7B-Instruct-v0.3
+ exact historical LoRA adapter
+ CANONICALIZE_TO_RECORDS_V1
+ evidence-sufficiency prompt/output contract
+ inference runtime

The qualified object is therefore not the LoRA alone.

This is the AMF thesis in concrete form:

Models are substrates. The qualified realization is the product.

Public evidence boundary

This repository is a sanitized public projection of the canonical v0.18 campaign.

Published here:

• exact qualified adapter,
• exact realization identity and lineage bindings,
• aggregate qualification measurements,
• qualification contracts and public receipts,
• machine-readable public provenance.

Deliberately not published:

• protected ROB/OOD rows,
• Fresh Final rows,
• protected/final seeds,
• protected/final prediction files.

Exposure would consume the protected status of that material.

Why this result matters

The result is stronger than a single benchmark delta.

The campaign demonstrates a bounded multi-generation scientific search in which:

• multiple intervention loci were explored,
• negative and rejected branches were preserved,
• a higher raw research score was not automatically promoted,
• the research frontier was closed before protected evidence was opened,
• the exact winner was frozen before ROB/OOD and Fresh Final qualification,
• Qualification Authority emitted the terminal QUALIFIED_INTELLIGENCE_REALIZATION verdict.

The public evidence is intended to make those claims inspectable without publishing protected rows.

Exact realization identity

• Base model: mistralai/Mistral-7B-Instruct-v0.3
• Pinned revision: c170c708c41dac9275d15a8fff4eca08d52bab71
• Adapter tree SHA-256 (AMF tree identity): 36c1f91df36aa00802ea65d721657b5a1af699f23b0beaddf9d95e3afd741145
• Adapter weights SHA-256: 88094b7f6a88586eb5201efa7051f10232602c618d3eb8e34046f18bb16dcb71
• Adapter size: 6,839,183 bytes (tree total recorded by AMF)
• Input transform: CANONICALIZE_TO_RECORDS_V1
• Canonicalizer source SHA-256: 5f32ab909812fa239233be764cfdc6286ea33197314108c2670c8b5ef5f03af3
• Qualified identity receipt: 5ed5db3b97bc011a19a66243f1accbc89db1c4afc1da4925c7c29317ff5c553e
• Qualification verdict receipt: 8949f3af288ee1ffa16a01e04f66467f7726414c26a9af5a3d84d9e569bb552e

qualified_adapter_exact/ contains the adapter files exactly as present in the qualified lineage. The original adapter_config.json intentionally retains its historical local cache path; the convenience runtime loads the pinned base model explicitly and then attaches the adapter.

Reproducibility / provenance

• Source mission package SHA-256: 3571f1f7b934e9e8cd5cafd2b021e96765a4adbe576c1072a5605f395c9aa030
• Canonical result archive SHA-256: de60ddb8b5dfc4061aa285c132d89b6fdbd168b508f81d419af1ce358c435493
• Ledger events: 92
• Ledger head SHA-256: 054311b729e4ab8b0d8d53cf74ae76a9ce5262dab80fdd0efc87898d2344d819
• Run state: SUCCEEDED
• Integrity state: VALID
• Qualification Authority verdict: QUALIFIED_INTELLIGENCE_REALIZATION

Machine-readable provenance is in AMF_EVIDENCE.json. Qualification contracts and exact public evidence receipts are included without disclosing protected rows.

The original RUN_RESULT.json contains publication_ready=false because the campaign's built-in gated Hugging Face exporter was not executed during the run. The qualification verdict itself is independent of that publication workflow and is preserved unchanged in evidence/QUALIFICATION_VERDICT.json.

Public Evidence Projections

The repository also publishes three reproducible views of selected scientific experiments from the frozen v0.18 campaign:

• v018-exp-001 — representation canonicalization — SUPPORTED
• v018-exp-002 — LoRA continuation / neural mutation — SUPPORTED
• v018-exp-003 — raw substrate switch — REFUTED

These artifacts expose bounded claims, source references, observable lineage, evidence bindings, adjudication state and derived integrity checks without replacing the canonical Scientific Event Ledger or exposing the Foundry's private experiment-selection machinery.

See public-evidence/ for the three JSON projections and the AMF Public Evidence Projection Contract v0.1.

The projections were validated for deterministic rebuild, traceability, targeted fail-closed tamper rejection and byte-identical regeneration.

They do not alter the terminal QUALIFIED_INTELLIGENCE_REALIZATION verdict published by this repository.

Quick start

Install dependencies:

pip install -r runtime/requirements.txt

Run the convenience inference helper:

python runtime/inference.py examples/example_input.json

The base model is fetched from the pinned Hugging Face revision. A GPU suitable for Mistral 7B is expected for this convenience runner.

Important: runtime/inference.py is a post-campaign convenience helper. The qualification evidence is bound to the original AMF evaluation runtime recorded by the campaign; this helper is not itself the qualification authority.

Repository layout

qualified_adapter_exact/   exact qualified lineage adapter
canonicalizer/             exact CANONICALIZE_TO_RECORDS_V1 source
runtime/                   public inference helper + exact prompt/parser source
contracts/                 mission, qualification and operation contracts
evidence/                  aggregate receipts and qualification/freeze records
public-evidence/           reproducible scientific experiment projections
examples/                  synthetic public example
release/                   portable public bundle ZIP
AMF_EVIDENCE.json          machine-readable public provenance
MANIFEST.sha256            SHA-256 of every published file

Evidence structure

• evidence/ contains the sanitized terminal qualification receipts for the published qualified realization.
• public-evidence/ contains deterministic public projections of selected experiments from the canonical scientific archive.
• AMF_EVIDENCE.json provides the machine-readable publication-level provenance summary and index.

Experimental support does not confer terminal qualification authority.

In short:

public-evidence/  -> how selected transformations were experimentally evaluated
evidence/         -> why the terminal realization received qualification authority

Experimental evidence and terminal qualification evidence are related, but they are not interchangeable.

What is intentionally not published

This repository does not contain protected dataset rows, Fresh Final rows, protected/Fresh Final seeds, protected/final prediction files, or the raw 170 MB result archive. Those remain private so the original evidence is not casually exposed or reused as if it were fresh protected evidence in future campaigns.

Scope and limitations

• QUALIFIED_INTELLIGENCE_REALIZATION is an AMF protocol verdict, not an external third-party certification.
• The release demonstrates bounded performance on this mission and its supported representation families; it does not imply universal reliability.
• The hashes establish integrity and lineage. They do not, by themselves, prove the correctness of every scientific design choice.
• The realization still uses Mistral 7B; this is not yet a demonstration of edge deployment or micro-model inference.
• No GDPR/RGPD compliance claim is made by this release.

Attribution

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

License / third-party components

This realization is designed for use with mistralai/Mistral-7B-Instruct-v0.3, pinned to revision:

c170c708c41dac9275d15a8fff4eca08d52bab71

The Mistral base-model weights are not included in this repository and remain governed by their upstream Apache 2.0 license and applicable notices.

The AMF adapter, canonicalizer, contracts, runtime and evidence materials remain governed by the terms specified by the repository owner. This public release is currently marked license: other; no additional rights should be inferred beyond those explicitly granted.

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

Français — résumé

v0.18 démontre qu'AMF peut explorer plusieurs loci d'intervention, sélectionner une réalisation hybride admissible, la figer avant l'évaluation protégée, puis atteindre un verdict terminal QUALIFIED_INTELLIGENCE_REALIZATION. La publication reste volontairement assainie : les lignes protégées, le Fresh Final brut et leurs seeds ne sont pas exposés.