nomos-v1-nano-g1

nomos-v1-nano-g1 is a small local agentic tool-routing co-processor. Given a user objective, current agent state, and legal candidate registry, it ranks the tools most likely to be useful for the next step.

It is not an answer generator, tool executor, or fixed-vocabulary classifier. It sits beside an agent as a fast candidate-reduction layer, so the primary LLM can reason over a short ranked list instead of every tool description in a large registry.

Nomos v1 ranks candidates through semantic metadata rather than treating tool names as output classes. The same checkpoint can therefore rank registries supplied by different agents, including tools with previously unseen names. This is the first Nomos v1 nano generation and a compact base for future tool-routing models.

Native V1 Output

The Fitz-Tool reference wrapper returns an ordered candidate list:

Field	Value	Intended use
tool_id	Registry-local tool identifier	The legal candidate proposed to the agent.
tool_family	Semantic tool family	Routing analysis and family-level holdouts.
semantic_fingerprint	Stable metadata fingerprint	Candidate identity across aliases and registry presentations.
score	Normalized cosine similarity	Relative ranking within the supplied candidate set.

Output Contract

The raw Hugging Face model returns normalized 384-dimensional embeddings. It does not emit a tool name from a closed label set. A router embeds the current decision state and every legal candidate, computes cosine similarity, and returns the highest-scoring candidates:

[
  {
    "tool_id": "inspect_record",
    "tool_family": "code_inspection",
    "semantic_fingerprint": "a1b2c3d4...",
    "score": 0.81
  },
  {
    "tool_id": "search_catalog",
    "tool_family": "search",
    "semantic_fingerprint": "e5f6a7b8...",
    "score": 0.63
  }
]

Scores are relative to the legal candidate set supplied for that decision. Candidate filtering, argument validation, execution, provenance checks, side- effect policy, and recovery paging remain responsibilities of the surrounding agent runtime.

Intended Use

Use this model when an agent runtime needs fast local signals for:

• reducing a large legal tool registry to a short top-k candidate list,
• ranking tools from their capabilities and schemas rather than memorized names,
• routing across registries owned by different agents,
• proposing a fresh candidate page when an agent rejects the first page,
• improving tool selection for language models with weaker agentic behavior,
• keeping routing local on CPU-only systems.

This model is not intended to choose arguments, execute tools, verify facts, authorize side effects, or replace deterministic runtime validation.

Input Format

The query side should describe the objective and relevant decision state:

Objective: Find the implementation corresponding to this SDK symbol.
Current need: Inspect code structure before opening a specific definition.
Agent state: active execution; source inventory known; schema unknown.
History: searched documentation by exact symbol; no implementation inspected.
Governance: read-only operations are allowed.

Each candidate should carry semantic metadata rather than relying on its name:

Inspect symbols and code structure in a source repository.
Capabilities: inspect code structure.
Accepts: code, text. Returns: symbols.
Evidence role: observation. Prerequisites: repository available.
Constraints: read only. Side effects: none.
Arguments: query (string required), scope (string optional).

Useful candidate fields include the description, capabilities, accepted and returned modalities, evidence role, prerequisites, constraints, side effects, and argument schema. Production integrations should preserve one consistent serialization for both training and inference.

Quick Start

from sentence_transformers import SentenceTransformer

MODEL_ID = "yafitzdev/nomos-v1-nano-g1"
model = SentenceTransformer(MODEL_ID)

state = """Objective: Find the implementation corresponding to this SDK symbol.
Current need: Inspect code structure before opening a specific definition.
Agent state: active execution. Governance: read-only operations are allowed."""

candidates = [
    "Search source text for an exact pattern. Capabilities: exact pattern search.",
    "Inspect symbols and code structure. Capabilities: inspect code structure.",
    "Search public web pages. Capabilities: web search.",
]

query_embedding = model.encode([state], normalize_embeddings=True)
candidate_embeddings = model.encode(candidates, normalize_embeddings=True)
scores = (query_embedding @ candidate_embeddings.T)[0]

for index in scores.argsort()[::-1][:3]:
    print(float(scores[index]), candidates[index])

The Fitz-Tool repository defines the complete state serialization, tool registry, legal-candidate filtering, and runner contract used in evaluation.

CPU Runtime

This repository contains the native SentenceTransformers checkpoint. No ONNX export is included in this release. Candidate embeddings can be cached when a registry is loaded, leaving only the current decision state to encode during warm routing.

Unoptimized PyTorch CPU measurements on the release workstation:

Legal candidate pool	Warm p50	Warm p95	Cold registry + query
10 tools	187 ms	200 ms	0.71 s
30 tools	292 ms	295 ms	1.07 s
100 tools	190 ms	203 ms	2.66 s

The representative state text differs between pool sizes, so warm latency is not expected to rise monotonically after candidate embeddings are cached. Results are workstation-specific and should be remeasured in each deployment.

Evaluation

All figures below measure raw encoder ranking. No downstream LLM choice, repair controller, deterministic override, or tool execution is included.

Frozen evaluation	Recall@1	Recall@3
sealed unseen-style states	0.9625	0.9750
independent ToolRet sample	0.6333	0.8000
final multiview suite	0.9257	0.9865
promotion multiview suite	0.9737	1.0000

Additional frozen generic-registry Recall@3 is 0.9440. These suites exercise opaque tool names, changed registry presentation, legal-candidate filtering, hard negatives, and held-out workflows. ToolRet contains 60 independently sampled queries, so its result has substantial sampling uncertainty. None of these benchmarks establishes universal tool-routing accuracy.

Training Data

Field	Value
Backbone	BAAI/bge-small-en-v1.5
Parameters	approximately 33M
Embedding dimension	384
Maximum sequence length	512 tokens
Training states	40,181 unique answer-present states
Main objective	in-batch multiple-negative ranking loss
Final checkpoint	90% established router + 10% full-replay specialist interpolation
Similarity	normalized cosine similarity

Training component	States	Role
Established replay lineage	35,081	Generic, agentic, ToolRet, transition, and opaque-contrast routing.
Balanced hard subset	5,100	Difficult examples selected from a separate 25,000-row scaling cohort.
Total	40,181	Answer-present states used for the final specialist branch.

Frozen evaluation questions were not added to training. Synthetic teacher proposals were filtered through the repository's validation and data-lineage workflow rather than being treated automatically as ground truth.

Artifacts

This repository contains:

• model.safetensors: standalone SentenceTransformers checkpoint,
• tokenizer and BERT configuration files,
• pooling and normalization modules,
• nomos_training_manifest.json: inherited training-lineage metadata,
• nomos_interpolation_manifest.json: final 90/10 interpolation record,
• nomos_calibration.json: optional abstention-calibration metadata.

The SHA-256 digest of model.safetensors in this release is 429fb204e4a38c1514c5a232376600fb4d3c6d2e82e9e27951cdfdeada27fca8.

Limitations

1. Top-k retrieval, not execution. The checkpoint ranks candidates but does not select arguments, execute tools, or establish provenance.
2. English-centric synthetic training. Multilingual performance has not been established.
3. Independent evaluation is still small. The ToolRet result is useful but does not cover the full diversity of real agent tool registries.
4. Metadata quality matters. Missing or misleading tool descriptions and state fields can produce incorrect rankings.
5. Abstention is separate. The included calibration is an optional policy signal and should be validated again for each integration distribution.
6. Warm latency assumes caching. New or frequently changing registries pay the candidate-embedding cost before ranking.

License

Mixed-source research preview. The BGE backbone is MIT-licensed, while the complete release also reflects synthetic and benchmark-derived training and evaluation sources with their own obligations. Review the source-card and data terms before redistribution or commercial use.