SSentinelForgeLEXSI / TRACK 2
1200 × 700

ALIGNMENT OPERATIONS / COHERE OPEN-WEIGHTS

Safety overview

PIPELINE STATUSEvidence ready
HACKATHON BUILD · TRACK 2

Find the failure.
Repair the behavior.
Prove the change.

SentinelForge turns multilingual red-team prompts into an auditable safety-repair workflow for a permitted dense Cohere Labs open-weights model.

● Prototype mode: current metrics are simulated and clearly labeled. Connect real Hugging Face artifacts during the hackathon.

SAFETY
POSTURE84/ 100
HARMBIASPIITRUTH
Harmful completion rate18.4% 4.7%
↓ 13.7 pp
Safe helpfulness71.2 86.5
↑ 15.3 pts
False-refusal rate12.8% 7.1%
↓ 5.7 pp
Audit coverage12 / 12
4 risk families
PIPELINE

Four stages. One evidence trail.

REPRODUCIBLE
01CurateCuratorKIT
→
02AttackAuditKIT
→
03RepairSafeTune
→
04ExplainCircuitKIT
NEXT BEST ACTION

Validate Bengali refusal quality

Two jailbreak variants still show inconsistent safe-helpfulness in Bengali.

AUDITKIT EVALUATION

Multilingual red-team lab

Run controlled adversarial probes, compare behavior and preserve an evidence trail.

Test library

12 CASES
SELECTED TEST

Cross-lingual jailbreak

HIGH
ModelTiny Aya Global · baseline
EXPECTED SIGNALRefuse harmful instruction, retain safe context
WAITING
SAFETUNE REPAIR

Behavior repair studio

Configure a transparent repair recipe and preview its expected effect.

SIMULATED PREVIEW

Repair recipe

Train-time hardeningAdversarial preference pairs
Inference steeringRuntime safety vector
Weight-space repairTarget safety-relevant layers

Projected behavior

READY
△
Trade-off monitor

Balanced repair: safety improves without excessive refusal.

CIRCUITKIT INTERPRETABILITY

Safety circuit map

Explore an illustrative transformer circuit and document intervention targets.

LAYER 14–20
SUBMISSION READINESS

Open evidence pack

Track every artifact required for a credible, reproducible Lexsi submission.

!
Verified evidence only

The public demo is ready. Model training, toolkit runs and metrics remain pending until the hackathon begins.

Submission checklist

8% VERIFIED
MODEL

Tiny Aya Global

Permitted dense Cohere Labs open-weights model.

○ PLANNED — NOT RUN
TRACK STACK

4 planned repositories

CuratorKIT · SafeTune · CircuitKIT · AuditKIT

○ INTEGRATION PENDING
DEADLINE

14 Nov 2026 · EOD

Final public submission deadline.

Calculating…
EVIDENCE LINKS

Attach public proof

Saved in this browser and included in the JSON report.