SiliconShield-RTL Contract-1 (local-guard specialist)

SiliconShield-RTL Contract-1 is a locally executable decision model for specification-conditioned RTL security assessment. It evaluates an RTL region together with an English security requirement and explicit execution assumptions, then returns a structured satisfies, violates, or insufficient-context decision. On balanced, group-disjoint development families, Contract-1 classified 191 of 192 fully specified decisions correctly, including 95 of 96 violations, with zero observed false-safe predictions. It helps hardware engineers prioritize simulation, formal verification, and review while keeping sensitive designs on local hardware.

Pair it with Contract-2, which owns FIFO, counter, and arbitration behavior. Neither model alone covers every RTL area. The SiliconShield router consults each on its qualified scopes and reports the rest as thin-evidence.

Labels

Given a self-contained RTL region, an explicit security requirement, and stated execution assumptions, the model produces one of three structured answers. It runs locally and keeps design inputs on user hardware.

Label Description
satisfies_requirement The model predicts that the supplied RTL region looks consistent with the stated requirement under the supplied assumptions.
violates_requirement The model predicts a possible inconsistency between the RTL and the requirement.
insufficient_context The model predicts that omitted implementation detail, assumption, or policy information prevents a reliable determination.

Quickstart

from huggingface_hub import snapshot_download

checkpoint = snapshot_download(
    repo_id="debjyotikm/siliconshield-rtl-contract-1",
    revision="2f63a4a87d0d6010fca033b1637ff4142ff7cc60",
)
import json
from laya.agent import load

agent = load(checkpoint)
contract = json.load(open(checkpoint + "/question_contract.json"))
questions = {
    "security": {
        "type": "security",
        "instructions": contract["instructions"],
        "criteria": contract["criteria"],
        "option_order": [0, 1, 2],
    }
}
answers = agent.predict_batch(
    [state_text],
    questions,
    batch_size=16,
    sort_by_length=False,
    max_len=1536,
    head_max_len=256,
)
probs = answers[0]["answers"]["security"]["probabilities"]
label = max(probs, key=lambda name: float(probs[name]))

state_text is the RTL plus requirement and assumptions in the benchmark row format (see below). Scores round to four decimals. Treat exact-zero target scores as confident misses. The SiliconShield workflow conclusion combines model assessment, executable checks, a repair proposal, and post-patch verification. The decision model alone neither proves the violation nor accepts the repair.

Benchmark

Headline results on reviewed development populations:

Metric Contract-1
Fully specified decisions 191 of 192 at 99.48 percent
Satisfying cases 96 of 96 at 100 percent
Violating cases 95 of 96 at 98.96 percent
False-safes 0 of 96
Full three-way accuracy 257 of 288 at 89.24 percent

Counterbalanced qualification at 249 of 288 at 86.46 percent with zero false-safes on 96 violations and shifted evaluation at 330 of 756 at 43.65 percent with 25 false-safes define the measured operating envelope. Broad retention at 3682 of 4095 is reported separately. Full matrices and family breakdowns are in the companion limitations record. Sealed populations stay sealed.

Public benchmark comparison on the SiliconShield RTL benchmark (5,589 rows, 11 areas), scored with the pinned prompt above. Development evidence. Sealed qualification and test populations were never opened.

Area Contract-1 Contract-2
fifo 36/108 97/108
counter 41/108 92/108
arbitration 14/18 14/18
interrupt-wake 158/234 145/234
input-filter 91/162 94/162
ot-access-control 108/108 77/108
local-safety-cwe 2601/2610 933/2610
transfer-pairs 606/621 275/621
write-paths 179/432 244/432
guard-semantics 171/432 268/432
requirement-clarity 346/756 271/756

Overall 4,351/5,589 for Contract-1 and 2,510/5,589 for Contract-2.

Reproduce with the SiliconShield repo (no training needed):

python -m siliconshield.benchmark run --registry models.json \
    --out-dir bench-out

What it handles well

Contract-1 is strongest when the requirement names a visible state element or write path and relevant guards are present in the RTL and assumptions define reset and sampling precisely. Representative strengths include lock-gated write paths, privileged or authorized controls before register updates, explicit reset exemption from hold requirements, and distinct treatment of normal and debug paths.

Limitations

  • Narrow guard specialist. FIFO, counter, and arbitration behavior are outside its strength. Route temporal scopes to Contract-2.
  • Misses some multi-path violations. A debug write missing a privilege guard was returned as satisfying with 99.74 percent confidence. One correct path does not prove every path was checked.
  • Fragile to rewording. Shifted requirement and context evaluation scores 330/756 with 25 false-safes.
  • Overconfident errors. Wrong answers routinely carry very high scores. Scores are not calibrated risk estimates.
  • Small-fragment scope. Judges bounded RTL regions against one requirement at a time. No full-IP, SoC-level, FSM, clock-domain, bus-protocol, crypto, or firmware-interface reasoning.
  • Out of scope. Broad discovery, CWE classification, localization, calibrated risk scores, and automatic patch approval.
  • A verdict is not a proof. Confirm findings by simulation or formal verification before acting on them.

Training and lineage

Continued from the OpenTitan-expansion winner through the SiliconShield security task: full encoder, rank-256 residual adapters, and decision head with the action head frozen. Training sampled 576 reviewed rows in 192 complete triplets with a group-disjoint 288-row development set. Learning rate 2e-6 with 10-update warmup and cosine decay to 0.1, batch 48, seed 3301, 250 updates with evaluation every 25, plain row cross-entropy. The release checkpoint scored 257/288 with zero false-safes and 66/96 whole triplets under frozen development gates. A follow-up improvement study trained three arms on a wider mixture and promoted none, so no released weight comes from those rows. Question contract siliconshield-requirement-context-v1, max length 1536.

Provenance

Field Evaluated value
Release SiliconShield-RTL Contract-1
Weights SHA-256 f4130223ed19be5cc7e21b3da70cb7aebf944ad6ed4a8a80f11d8f59bcd05ee7
Question contract siliconshield-requirement-context-v1
Contract file SHA-256 dd1f38ee223abac1f75b19d9086fd26d90865c9aa43097674fffdd61dbbf8bd3
Adapter metadata 28 residual-adapter layers, rank 256, scale 8
Evaluation precision BF16 model inference, FP32 probabilities and metrics
Input budgets max_len 1536, head_max_len 256

Model bytes plus question contract plus runtime behavior plus label order are one identity. Marginal decisions can differ across hardware even when identities match, so reproducible claims must name runtime, precision, contract, checkpoint, and population.

License

Apache-2.0. LAYA foundation weights retain their upstream terms. See the SiliconShield repository for the exact pinned revision.

Downloads last month
33
Safetensors
Model size
0.4B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for debjyotikm/siliconshield-rtl-contract-1

Finetuned
(251)
this model