debjyotikm/siliconshield-rtl-contract-2

SiliconShield-RTL Contract-2 (temporal specialist)

Contract-2 judges RTL against security requirements over FIFO, counter, and arbitration behavior: occupancy, backpressure, saturation, commit/hold, grant handshakes, and missing implementation context. It returns one three-way verdict per requirement: satisfies_requirement, violates_requirement, or insufficient_context, with all three scores.

Pair it with Contract-1, which owns local guard, register-write, and access-control 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-2",
    revision="b96211872d573ae48b6dabe128c0d96aad291156",
)

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

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

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

Overall 2,510/5,589. In-domain (fifo/counter/arbitration): 203/234 with perfect 78/78 missing-context judgment and 51/78 whole triplets, with 8 of 20 mechanisms perfect. Violation recall there is 61/78 with 17 false-safes.

Full per-area and per-mechanism detail (classes, false-safes, false violations, abstentions, triplets, NLL, inference settings) is in benchmark-results.json in this repo.

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-2 is strongest on FIFO occupancy control and counter behavior under explicit requirements, with 97/108 and 92/108 on those benchmark areas and perfect missing-context judgment in domain. Arbitration support is narrow at 14/18.

Limitations

• Narrow specialist. Local safety, transfer behavior, access-control bugs, write paths, and plain-English policy reasoning are at or near chance. Route those to Contract-1.
• Misses one in five in-domain violations (17 false-safes on the focused set). Commit-hold, arbiter-tree, empty-data-blanking, and stored-word-hold families are unreliable.
• Owner-selected research checkpoint. It did not pass the frozen promotion floor (violation recall, false-safe, and NLL gates). The trainer's selector held the parent instead.
• 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 a temporal-pilot parent export through the SiliconShield security task: existing residual adapters and decision head with the encoder frozen. Two single-GPU arms ran and the triplet arm produced this checkpoint. Training used 2,700 solver-verified FIFO, counter, and arbitration rows in complete triplets. Learning rate 1e-7 with 5-update warmup and hold-decay to 0.2, batch 48, seed 7502, 300-update horizon with evaluation every 10, row cross-entropy plus worst-member triplet loss. The owner selected this checkpoint for satisfies and context recovery at minimal violation cost. It did not pass the frozen promotion floor. Question contract siliconshield-requirement-context-v1, max length 1536.

Provenance

Field	Evaluated value
Release	SiliconShield-RTL Contract-2
Weights SHA-256	f916444ff45cfc67bc5a67c81a36a770a20709fcd7a85fbfc0258fb411144ed8
Question contract	siliconshield-requirement-context-v1
Contract file SHA-256	dd1f38ee223abac1f75b19d9086fd26d90865c9aa43097674fffdd61dbbf8bd3
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.