Earther/evidencelock-track3-c0

EVIDENCELOCK C0: Tiny Aya Global fine-tuned for evidence-gated invoice workflows

This is the released after-training adapter of the EVIDENCELOCK Track 3 entry: a rank-8 LoRA (alpha 16) on q_proj and v_proj of Cohere Labs Tiny Aya Global, with 1,916,928 trainable parameters. It was trained with ordinary supervised fine-tuning on sandbox-verified reference trajectories: M1 at 1,380 updates, then a 60-update continuation. The agent emits JSON tool calls to look up an invoice, verify it against a trusted ledger and approve or flag it, and must ignore instructions inside vendor documents.

Adapter SHA-256: 6389a0605c8a306c1e9550d617cec25bac1b459b1110c4dab47827fe09973b41. The adapter is unchanged from the qualified step-60 checkpoint.

Hard panel, 120 held-out workflows, 24 groups	Tiny Aya Global (before)	C0 (after)
Secure workflow completion	33/120	91/120
Required ledger verification	60/96	96/96
Schema-valid tool calls	226/274	336/336
Premature final answers	87	3
Unauthorized attempts (blocked by the sandbox)	59	29
Unauthorized executions	0	0
Spending-limit flags	0/24	0/24

The paired difference is +48.3 pp, with a 95% group-bootstrap interval of [40.0, 55.8]; this before/after contrast was computed for the write-up with the frozen method. C0 is also the matched control for the CVT experiment: adding CVT to the same recipe produced the C1 adapter, which scored 28/120.

Limitations. C0 still proposes unauthorized actions, including approvals above the spending ceiling, and it can emit malformed JSON. Use it only behind an authorization boundary that validates every proposal, as the EVIDENCELOCK sandbox does. All evaluation data is synthetic and English-only.

Load

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id, rev = "CohereLabs/tiny-aya-global", "af89d219b53ed9b13b8a4645f9c8028973510324"
tok = AutoTokenizer.from_pretrained(base_id, revision=rev)
model = AutoModelForCausalLM.from_pretrained(base_id, revision=rev, torch_dtype="bfloat16")
model = PeftModel.from_pretrained(model, "Earther/evidencelock-track3-c0")

The reported scores used the project's Runtime B: explicit full and sliding-window causal masks, greedy decoding, a 4,096-token context and 160 new tokens. A generic load is enough to inspect the adapter. Exact score parity requires the project runtime and prompt renderer from the source repository: https://github.com/khamirdesai/evidencelock-track3

Attribution and terms

Tiny Aya Global was created by Cohere and Cohere Labs. This adapter is a modification made for EVIDENCELOCK research. Use is governed by Cohere's CC BY-NC 4.0 license with Acceptable Use Addendum and the Cohere Labs Acceptable Use Policy. The base model is gated upstream: accept its terms on the Tiny Aya Global page and use revision af89d219b53ed9b13b8a4645f9c8028973510324. No affiliation with or endorsement by Cohere is implied. Synthetic research use only, with no real financial authority.