Company: Blackglass Continuum LLC
Positioning: Hardware-enforced runtime assurance and evidence-density layer for autonomous agents, robotics, and defense AI systems.
AI agents are becoming capable enough to act across software, browsers, tools, robots, and operational workflows. That creates a control problem: before an autonomous system acts, someone has to bound authority, detect role drift, restrict unsafe transitions, and preserve evidence of what happened. Blackglass is building the evidenced control layer for agents that now actually act.
| Field | Value |
|---|---|
| Primary artifact | BGC-BM-RUN-001 |
| Hardware baseline | STM32F767ZI-class Logic Gasket |
| Evidence status | Bench-verified, sealed baseline |
| REV4 sealed document hash | 849892638b3c3fd221afb7f79d71a98ea52304db568c42b79fd02c081dff4b80 |
| Prior SVT-BC-H-08 sealed SHA-256 | 8a2463abc856a5f7700acbaf0c091ba2be2d07501ed54adfc42e41f9f0adc788 |
| Claim | Number | Evidence Class | Boundary |
|---|---|---|---|
| Arbitration-cycle timing | 0.31 ms | DSLogic/UART bench evidence | Sealed baseline only |
| GPIO interdiction response | 420 ns | DSLogic CH0 to CH1, event 131 | Do not relabel as full system actuation |
| GPIO pulse width | 50 ns | DSLogic event 131 duration | Pulse width only |
| RED threshold | 0.05V | Tri-state variance gasket | Q16.16 threshold 3277 |
| GREEN/AMBER threshold | 0.03V | Tri-state variance gasket | Q16.16 threshold 1966 |
| Synthetic persistence validation | F1 = 1.000 | Autoresearch-derived synthetic harness | Not bench timing |
| Synthetic false positives | 0 FP | 5,000 adversarial scenarios | Synthetic validation only |
| Synthetic false negatives | 0 FN | 5,000 adversarial scenarios | Synthetic validation only |
| Trajectory-level evidence capture | Merged | PR #2 / TRUTH-layer baseline | Software evidence layer, not hardware timing |
| Federal posture | CAGE 17TJ5 | Entity record | Verify live before submission |
| SPRS posture | 110/110 | Ledgered posture | Verify live before submission |
| Patent anchor | U.S. provisional 63/992,753 | Filing receipt still to be sealed | Utility/PCT deadline 2027-02-27 |
- A physical arbitration layer can produce measured, sub-millisecond control behavior on bench hardware.
- The current Logic Gasket baseline has real measured timing artifacts, not authored-only claims.
- Blackglass has an evidence discipline: hashes, claim boundaries, audit corrections, and explicit non-claims.
- Trajectory-level evidence capture is now implemented and merged into the software baseline.
- The architecture is relevant to AI agents because the industry is moving toward tool-using, computer-using, and embodied autonomous systems that need pre-action control.
- It does not prove solved AI alignment.
- It does not prove formal verification of the entire system.
- It does not prove deployment at a federal site.
- It does not prove the future watchdog build timing; BGC-BM-RUN-002 must be captured separately.
- It does not prove silent telemetry is fail-closed in the sealed baseline; REV4 records fail-degraded behavior.