MONTREAL.AI / SKILLOS
Autonomous RSI Silicon Verification Proof
Recursive self-improvement on RTL bug triage, assertion selection, and fix-plan recommendation.
Current status
PASSED_AUTONOMOUS_RSI_SILICON_VERIFICATION_MARKET_PROOF
No human review. No emails. No invoices. No CloudOps or cyber reuse. No customers. No private data. No API keys. Deterministic holdout benchmark.
+85.8 ptsfully-correct gain
100.0%SEV1 recall
97.9%debug-time reduction
$2,817,512,477.9synthetic cost avoided
Recursive self-improvement curve
Before / after on holdout RTL verification cases
| Metric | Baseline | SkillOS RSI |
|---|---|---|
| Fully correct decisions | 14.2% | 100.0% |
| Bug-class accuracy | 14.2% | 100.0% |
| Assertion accuracy | 14.2% | 100.0% |
| Fix-plan accuracy | 14.2% | 100.0% |
| SEV1 recall | 11.1% | 100.0% |
| Design escape rate | 44.4% | 0.0% |
| Avg debug days | 19.797 | 0.425 |
| Avg cost | $3968946.96 | $55735.19 |
Final learned verification skills
- skill_fifo_underflow — Detect reads when FIFO is empty and fix with empty-flag gating plus valid-state preservation.
- skill_off_by_one_counter — Detect terminal-count boundary errors and fix the counter comparison/increment boundary.
- skill_handshake_deadlock — Detect valid/ready circular waits and fix with registered-ready progress rule.
- skill_reset_state_leak — Detect unreset state leakage and fix with complete reset initialization and coverage.
- skill_arbitration_starvation — Detect starvation under contention and fix arbitration with fairness/aging.
- skill_cache_coherence — Detect coherence-state violations and fix invalidation/dirty-state transition.
- skill_address_alias — Detect overlapping address decode windows and fix masks/priority.
- skill_cdc_metastability — Detect unsynchronized CDC paths and fix with synchronizer or async FIFO.
- skill_sign_extension — Detect signed-width extension errors and fix with explicit casts and width normalization.
- skill_endian_swap — Detect byte-lane mapping errors and fix endian/protocol adapter mapping.
- skill_credit_underflow — Detect credit underflow and fix decrement/return path accounting.
- skill_war_hazard — Detect write-after-read pipeline hazards and fix with interlock/forwarding.
- skill_power_transition — Detect bad power-state sequencing and fix retention/power-gate sequence.
- skill_interrupt_edge — Detect lost interrupt edges and fix with latched edge-until-ack behavior.
- skill_packet_length — Detect packet length/payload mismatch and fix field-counter binding.
- skill_timing_constraint — Detect timing-constraint mismatch and fix constraint/pipeline strategy.
- skill_fifo_overflow — Detect writes when FIFO is full and fix with full-flag gating plus depth guard assertion.
- skill_clean_no_bug — Recognize clean cases and avoid unnecessary changes.
Proof gates
- ✅ not email workflow
- ✅ not invoice workflow
- ✅ not cloudops workflow
- ✅ not cyberdefense workflow
- ✅ no human review required
- ✅ no emails sent
- ✅ no customers contacted
- ✅ no private data used
- ✅ no api keys required
- ✅ deterministic reproducible benchmark
- ✅ recursive self improvement releases at least 8
- ✅ rsi validation improves monotonically
- ✅ train cases at least 350
- ✅ validation cases at least 175
- ✅ holdout cases at least 700
- ✅ final rules at least 18
- ✅ fully correct gain at least 70 points
- ✅ bug class accuracy at least 99 percent
- ✅ assertion accuracy at least 99 percent
- ✅ fix plan accuracy at least 99 percent
- ✅ sev1 recall at least 99 percent
- ✅ design escape rate zero
- ✅ debug time reduction at least 80 percent
- ✅ cost reduction at least 80 percent
- ✅ synthetic cost avoided positive