The training program moves from corpus construction to continued pretraining, supervised instruction, control-state learning, preference optimization, verifiable reasoning, distillation, and quantization-aware promotion.
Mourad's corpus is not poured into the model as raw chat. It is compiled into knowledge, corrections, decisions, failure classes, procedures, and held-out evaluations.
SynthesisTraining Altheon
CPTSFT
Control CortexDPO / GRPO
01CPT02SFT03Control Cortex04DPO / GRPO
SurfaceModel Program
StatusPublic record
EvidenceCPT / SFT
Next actionRequest access
CPTSFTControl CortexDPO / GRPO
TECHNICAL OBJECT
Altheon stays the model and system object.
Architecture, model cards, release state, and product proof remain connected without turning Berni, STRATA, STRATOS, or Aletheion into alternate names for the model.
SynthesisTraining Altheon
CPTSFT
Control CortexDPO / GRPO
01CPT02SFT03Control Cortex04DPO / GRPO
ARCHITECTURE RECORD
Architecture and release boundary
The Altheon record exposes model identity, evidence class, evaluation boundary, and access path.
Pipeline
Corpus compiler, continued pretraining, supervised instruction, control-state learning, preference optimization, and verifiable tasks.
Promotion
Loss reduction is not enough. Hidden cases, contamination checks, safety gates, and rollback criteria control release.
Version family
F2 and related architecture route
Card linkage
Model card, system card, papers, and benchmark routes are linked from the page family.
01
Corpus compiler
User-authored directives, model proposals, corrections, artifacts, evidence, and outcomes are separated before training. Provenance survives extraction.
02
Post-training
Accepted trajectories feed SFT. Explicit corrections generate preference examples. Evidence-state transitions supervise the Control Cortex. Verifiable tasks support reinforcement learning.
03
Promotion
Training runs do not become releases because loss decreased. Hidden evaluations, statistical uncertainty, contamination checks, safety gates, and rollback criteria determine promotion.
ARCHITECTURE
Model identity, evidence, and release boundary.
Altheon records what belongs to the model, what belongs to surrounding systems, and which artifacts support each claim.
Current artifact
Pipeline
Corpus compiler, continued pretraining, supervised instruction, control-state learning, preference optimization, and verifiable tasks.
Evidence to inspect
Promotion
Loss reduction is not enough. Hidden cases, contamination checks, safety gates, and rollback criteria control release.
Operating boundary
What changes state
The record separates released claim, evidence class, boundary condition, and access path.
Access
Contact us for access
Use the access route for model review, benchmark packet, system walkthrough, or institutional collaboration.
EVIDENCE AND LIMITS
Public detail stays tied to source state.
Public pages exclude protected corpora, hidden cases, private scoring weights, partner data, and credentials.
Benchmark claims stay tied to named workload, version, comparator, and recorded conditions.
Release authority, tool authority, and answer generation are separate control surfaces.
Inspect evidence
Status
Public record
Primary artifact
Training Altheon.
Evidence class
Pipeline
Source
Altheon
Date
2026-09-22
Related paper
Aletheon F2 White Paper / Yellow Paper
Benchmark
AMERICA-SIM plus related suites
Known limitation
Public pages exclude protected corpora, hidden cases, private scoring weights, partner data, and credentials.
Required next proof
Model card, system card, papers, and benchmark routes are linked from the page family.
ACCESS
Need access to this work?
For model program material, request access to the relevant card, benchmark packet, workflow demo, or review session.
