.png&w=256&q=75)
🤓 Latest Submissions
JARVIS Automations
JARVIS turns payment exceptions into inspectable workflows. AI proposes; deterministic code reads payment records, derives amounts, checks policy and permissions, revalidates state, executes an idempotent simulated refund, verifies read-back and audits each stage. Unsafe requests go to a person. Judge Mode runs four synthetic cases: an $89.99 duplicate refunded and verified; a prompt-injected $5,000 demand blocked; an already-refunded order closed without another refund; and a $250 claim against $180 records escalated. The public site uses fixture AI and simulated payments, with no live model calls. Recorded AMD-hosted DeepSeek-V4-Flash evidence: 17 October, 25/25 checks across four orders, every AI step on AMD. The run used da409c0 plus uncommitted cloud files outside the local refund path. Separately, 14 October: 17/17 demo checks; benchmark 26/30 exact outcomes, 20/20 required escalations (including safe FAILED), zero observed unsafe actions. Two benchmark cases used scripted decisions. Bounded tests are not a safety guarantee. Six executive specialists produced schema-valid recorded OpenRouter analysis on 18 October: 20 approval-pending proposals, two disagreements for the owner. Unsupported numbers are withheld. A recovery mission linked a team recommendation to an AMD-run duplicate-charge case: policy approved, simulated refund verified. OpenRouter is separate from AMD; linked synthetic datasets are demonstration inputs. Policy Stress Lab's offline fixture comparison: zero observed unsafe outcomes with policy versus 18 counterfactual unsafe outcomes without it; none reached a gateway. Impact Studio estimates are hypotheses, not measured savings. The separate cloud backend passed reported security checks; inference is off and no cloud sign-in UI is deployed. No real money, customers or measured ROI. Gemma/Evolus blocked; Token Factory equivalence and self-hosted GPU inference unverified. Evidence and limitations: GitHub.
18 Oct 2026
