Fenrua-Labs/FML-Mosaic-527B

FML Mosaic 527B

Status: Fenrua-authored architecture design only. This repository does not contain model weights, a checkpoint, a tokenizer, an inference runtime, or an inference endpoint.

FML Mosaic 527B is Fenrua Labs Pty Ltd's planned crown architecture for a future standalone text-generation family and knowledgeable AI assistant. The 527B identifier is a design-scale target, not a realised parameter count, capability claim, or hardware-fit claim. No trained FML tensor exists in this repository or in the current public release, so it is not a live assistant.

Current execution scope

The local program is currently limited to preparing its first fail-closed gates. There is no approved named-file corpus, sealed corpus receipt, current human authorization, or bound machine inventory. Consequently, native 196,608-token genesis, model/tokenizer binding, FML .fndl segments, a Hugging Face training-vault target or transfer, training, evaluation, a realised tensor census, and release remain blocked. This public repository contains no credential, corpus payload, tokenizer artifact, model material, or shortcut around those local approvals.

What is present

This repository is a public, maturity-labelled design record. It publishes only Fenrua-authored status, licensing, and release-boundary documentation for the future FML Mosaic 527B program.

FENRUADL (Fenrua Disk Loader) is the associated Fenrua-authored artifact and resource-planning format work. Its current format controls are original to Fenrua Labs Pty Ltd and are not a compatibility layer for an external runtime or model format.

Technical design summary

Field	Current controlled design record	Boundary
Intended public role	Knowledgeable AI assistant by Fenrua Labs Pty Ltd	Not a live assistant or capability claim
Primary language target	English (en)	No trained English-language behaviour is claimed
Intended modality	Left-to-right causal text generation	text-generation is an intended-task classification only; no runnable generation path is published
Architecture ID	fml-mosaic-527-design-v0.1	Symbolic design identifier, not a checkpoint ID
Mosaic schedule	Eight cycles × eight sparse stages = 64 stages	No runtime implementation or routing result
Expert fabric	32 symbolic expert modules per sparse stage	No trained expert modules
Routing cadence	48 ordinary stages × 2 selections + 16 junction stages × 3 selections = 144 selections per token	Not a derived active-parameter, memory, or throughput value
Hidden width	12,288	Proposed shape only; no materialised tensor
Attention / position controls	96 × 128 heads with a 64-wide Phase Braid pair plane	Symbolic shapes only; no attention implementation or cache
Forward-control policy	Gain-only norms, per-stage router controls, request-scoped 16-slot thread relays, renewal/anchor/audit/output controls	No runtime state, forward pass, or trained controls
Future execution contract	FML-authored mathematical definition for attention, routing, relays, anchors, audit outputs, and tied logits; a local four-part synthetic reference suite checks small witnesses	No full numerical executor, tensor, or model forward pass has run
Future training contract	FML-only fresh initialization, causal objective, optimizer, precision, and native-resume rules	Planning-only; it does not authorize training or create a checkpoint
Future tokenizer target	196,608 entries, with a separately bound FML-native target-scale profile	The current experimental trainer is capped at 65,536 entries and cannot be promoted as the target tokenizer; no tokenizer artifact is present
Anchor binding	96,000 proposed learned-anchor rows	No corpus, retrieval index, or trained table
Symbolic tensor inventory	42 named tensor groups / 7,170 symbolic instances; 7,170 ABI owners plus one tied alias	Zero stored tensors, .safetensors files, or weight shards
Symbolic ledger	527,044,706,304 base + 30,746,821 exact forward controls = 527,075,453,125 symbolic unique parameters	Design arithmetic only; no realised tensor census

The full topology, dimensions, and evidence boundary are in ARCHITECTURE_RECORD.md. The table is deliberately technical without converting an untrained design into a capability claim.

The pipeline_tag: text-generation metadata makes the intended future task discoverable on Hugging Face. It does not mean that this repository can currently generate text: there is no model configuration, tokenizer, tensor file, runtime, hosted inference provider, or executable inference path.

Design contract record

ARCHITECTURE_RECORD.md records the current Fenrua-authored symbolic topology and forward contract: its fixed stage cadence, named projection/control roles, tied-input/output design assumption, 96×128 attention geometry, Phase Braid position controls, request-scoped thread-slot boundary, and design-ledger boundaries. It is a reviewable description of a proposed architecture—not an executable model definition, checkpoint specification, or evidence that the architecture has been trained or validated.

FORWARD_CONTRACT_RECORD.md gives the exact control-plane accounting and symbolic order for independent review. It is a design contract, not a model implementation or training claim.

EXECUTION_AND_TRAINING_RECORD.md adds the separate FML-authored future mathematical execution contract and the from-scratch initialization/training contract. They lock the intended equations and future training controls while remaining explicitly non-executable and untrained. They do not create a model, tokenizer, corpus, cluster allocation, or training authority.

The current record is bound to symbolic-topology fingerprint 559a06f43bcfc2f2763a1f69a6b422d83bc37162767dbd461b27c371db745ac3, forward-contract fingerprint 632d30d660657cd87b2dd22552184330ecf79c7e5906dda89bc2797da8661186, and canonical-specification fingerprint 65df62c8796067e8f47c78eb3411c15cdbbaeab5da21c085e4eb3bb028653777. They identify controlled design records only; neither identifies a model artifact, trained tensor, tokenizer, or release approval.

WSL / Linux record verification

The public repository is documentation-only. The following verifies the exact published record; it does not download or run a model:

git clone https://huggingface.co/Fenrua-Labs/FML-Mosaic-527B
cd FML-Mosaic-527B
sha256sum --check SHA256SUMS
sed -n '1,220p' ARCHITECTURE_RECORD.md

For real local design-control commands and expected fail-closed results, see WSL_DEVELOPER_GUIDE.md. The guide has no inference, training, tokenizer, or model-loading command because no such artifact exists.

What is not present

• No model parameter files or converted weight shards.
• No pretrained, merged, copied, adapted, or imported model artifact.
• No external checkpoint, tokenizer checkpoint, or runtime.
• No training-data payload, evaluation payload, or remote-code dependency.
• No live generation, benchmark, throughput, safety, or deployment claim.

Conditions before a future model release

A future FML Mosaic 527B release may be described as a trained model only after all of the following have independent, reviewable evidence:

1. an approved named-file data manifest with recorded provenance and rights;
2. an FML-only tokenizer-genesis receipt;
3. a self-bound local cluster inventory reviewed against raw static-state lower bounds; this planning evidence does not establish cluster reachability, allocation, training admission, or hardware sufficiency;
4. a fresh Fenrua-owned training run with zero external model-weight inputs;
5. a numerical FML-native executor demonstrated to conform to the bound mathematical contract;
6. a complete realised tensor census and training receipt bound to the architecture; and
7. original FENRUADL artifact, shard-set, evaluation, and release evidence.

Until then, this is deliberately a design record rather than a functioning text-generation model. A model name and a design-scale target are not evidence of a delivered capability.

Company and contact

Fenrua Labs Pty Ltd
fenrua.ai · partnerships@fenrua.ai

ABN 62 700 182 663 · ACN 700 182 663 · NSW, Australia

Community use and integrity

The original materials currently distributed in this repository are available under Apache-2.0. That makes the public design record usable, forkable, and shareable by the community. It does not claim to license absent weights, checkpoints, tokenizers, datasets, runtimes, or any third-party material.

Fenrua Labs' organisation profile uses a general other licence label because individual Fenrua repositories may have different release terms. That profile label does not narrow the Apache-2.0 licence for the original materials actually distributed in this FML Mosaic 527B repository.

SHA256SUMS is the normal integrity lock for this exact public record. Run sha256sum --check SHA256SUMS after cloning to verify it. The digest supports provenance and reproducibility; it is not an extra restriction on lawful Apache-2.0 use or on a fork that records its own changed digest.