🎓

Fleet AI — Training

Workshop Mode · fine-tune the local model on everything it's learned
← Back to Fleet

Python mastery engine Qwen3.8-27B

Verified micro-improvement loop. Unsloth/QLoRA is the preferred Qwen3.8 stack; DPO learns from executable chosen/rejected pairs. Nothing is promoted unless it beats the golden checkpoint without regression.

golden executable score
chat corrections awaiting review
training stack
Golden adapter
ModelQwen/Qwen3.8-27B
Precision4-bit QLoRA
Protected passes
Latest self-improvement
Learning mode
Team learning
Owner corrections from chat are stored as pending evidence first. Team Mode can review them before they become training pairs.

Fleet Operator Bridge assistant-ready

One evidence-grounded control snapshot for you, Fleet AI, Forge and external assistants. It exposes the current target, source grounding, teacher panel, cartridge verification debt, mastery evidence, stall alerts and next action without making another AI reconstruct state from dozens of files.

training state
teachers responding
Fast Proof passed
ready for 27B core challenge
cartridges awaiting proof
attention flags
Current target
Evidence grounding
Progress strategy
Latest proof lane
Recommended next action
This bridge is read-only. It helps an operator AI diagnose the live system quickly; it does not award mastery or change training state.

Capability Autopilot 54-domain knowledge

Fleet's metacognitive control layer. It routes each task into the installed Supergenius domains, uses the smallest relevant knowledge slice, and diagnoses what is actually missing before asking for training.

/54 knowledge installed
usable/deep domains
route-only domains
thin domains
depth backlog
open capability needs
Operating ruleKnow → retrieve → diagnose → acquire smallest missing capability → verify → continue
Self-diagnosisKnowledge · skill · tool/code/runtime · orchestration · evaluation · model training
MasteryKnowledge installation is not mastery. Exams, practicals, residency and retention stay separate.
Current status
Authoritative acquisitionidle
Fleet can use installed domain knowledge immediately while independently identifying residual gaps.

Capability Fabric live App body

Fleet's dynamically discovered tool body. New Game Studio, Imaging, Reality, Audio, Browser, Coding and future capabilities become available to planning without model retraining.

live App tools
game/world
imaging/visual
3D/reality
audio/voice
research/learning
compiled workflows
proven workflows
Operating ruleDiscover capability → select smallest useful toolchain → act inside permissions → verify outcome → retain learning
Missing capabilityInspect existing modules/open source → prototype reversible App tool → verify → register automatically
AuthorityCapability discovery never bypasses credentials, spend, GPU, production mutation, destructive or security gates.
Largest capability families
Live inventory is rebuilt from actual exposed Fleet tools, so this stays current as the App grows.

Shared Causal World checking

One reusable Fleet actor/NPC runtime across RoutedLab and Flickbuster, with separate authoritative domain truth. Dialogue and 3D embodiment remain proposal/presentation layers only.

257Fleet tools
cross-domain gate
Flickbuster actors
RoutedLab worlds
Truth boundaryAuthoritative domain state → bounded perception/belief → GOAP/procedures → validated actions → presentation-safe 3D projection
Memorybounded · content-addressed · deduplicated · no Wasabi per-tick scratch
3D contractpresentation only · hidden truth excluded
Training residencycausal-world-engineering · trained/ingested · fresh unseen exam queued · mastery not yet awarded
Fleet routes NPC/world work through the live causal-world contracts before Game Studio embodiment.

Forge Engineering V42

Fleet's evidence-grounded network/architecture documentation and diagram compiler. Forge turns requirements, observed evidence and proof into consistent engineering deliverables without inventing missing facts.

platform
artifact families
generic jobs
acceptance
Lifecycle
Truth contract
Generic projection profiles
Source Authority Fabric
Structured vendor guidancesentence-level constraints · source/hash receipts · version fail-closed · contradiction blocking
Known issues / field experience
Source drift
Vendor API adapters
Latest artifact QA
Integration debt
Fleet can route HLD/LLD/Visio/change/test/implementation/runbook/matrix work into Forge through Capability Fabric.

Network Sensorium read-only evidence

Fleet's network-engineering instrument layer for Cisco IOS XE, FortiGate/FortiOS, Azure and generic network evidence. It extracts line-linked interfaces/routes/neighbors, keeps observed/inferred/unknown state separate, and plans the minimum evidence needed before a change.

evidence snapshots
4network contexts
0automatic mutations
Safety contractRead-only · no invented commands · no production mutation
Capabilities
Last snapshot
Network knowledge is now paired with an evidence sensorium. Live-device integrations can be added later without changing this truth contract.

Cognitive Reactor Matrix cartridges

Instant capability activation first. Fleet compiles verified knowledge into a runtime cartridge, plugs it into the skill/competence graph immediately, then Reality Gym proves transfer before any mastery claim or weight training.

active cartridges
compiled cartridges
competence nodes
ModeMatrix cartridge · instant runtime capability
Next stageAutonomous verified practice → changed transfer → retention
Weight trainingOnly when a residual model-behavior gap remains after runtime skill + practice
Last event
Compiled cartridges are reversible externalized capability. They do not falsely claim that Fleet's base-model weights changed.

Training dataset

Built automatically from every teacher-escalated answer + verified skill. It grows as you teach Fleet AI.

unique examples
ready to train
200recommended min

GPU status

The shared Vast controller. Fleet only ever wakes it for training — never automatically.

State
Reserved by
Purpose
Hold remaining
Rate

General knowledge LoRA workshop

Separate from the Qwen3.8 Python mastery engine above. This legacy/general workshop fine-tunes the current local assistant dataset and remains manual only.

Safety: if a customer production build holds the GPU, this shows busy and refuses — never a takeover. While you're training, customer TTS uses its OpenAI/Kokoro fallback. Idle-stop is paused while your hold is active. Trained adapters always archive to Wasabi (never only on the ephemeral box).

Cognitive Clone Lab AI ZIP prototype

Capture transferable capability from a teacher AI into an installable Cognitive Clone Capsule: knowledge, causal models, procedures, failure patterns, decision rules, tool policies, counterexamples, preference pairs, tests and verification rules. This does not copy proprietary weights.

clone sessions
AI ZIP capsules
cognitive genomes
compiled genes
gene tournament rounds
Pipelineinterrogate → capture → compress → ZIP → Matrix install → changed-scenario proof → optional adapter distillation
Teacher sourcesOpenAI · Anthropic Claude · Gemini · DeepSeek · open-weight teachers
Weight policyNever claim proprietary provider weights were copied; only open/owned weights may be imported directly
Latest capsule
Balanced gene champions
Verified natural selectionNo core-lane champions yet
Shadow Champion Laneidle
Prototype backend is active. Provider-wide extraction is deliberately not automatic because it can consume substantial API credits; captured capsules are independently verified before mastery.

Matrix Cognitive Delta Compiler ready

Rapid capability import without replacing Fleet's Qwen3.8-27B core. Measure only missing competence, import structured deltas from OpenAI / Anthropic / Gemini / DeepSeek, preserve provenance, quarantine weak claims, compile an instant Matrix cartridge, then prove it in changed practice.

teacher APIs ready
delta sessions
latest state
Provider calls are not automatic. Creating and compiling sessions is local/free. External teacher calls require an explicit owner-authorized action because they may consume API credits.
Checking provider bridge…

Recent training runs

No runs yet.