Skip to main content

Changelog

Release history for the Quint platform. Each entry maps to a tagged release and documents what was built, what changed from spec, and what’s next.

Intent Model (Preview, in development)

On-device intent model that judges whether an agent’s stated intent licenses its actions. Model proposes, scope disposes; conversation text stays on the endpoint; per-policy-key calibration. Leak-free evaluation methodology established; on-device encoder upgraded to a larger, more capable model. Pre-GA, no accuracy claims yet.

v1.0.3 — Session Attribution & Edge Stability

session_id + process_pid on every audit row, decoded timeline API (/api/sessions/timeline), NE CPU watchdog fix (99% loop → 50ms poll budget), SSE chunked-framing fix, historical sessions visible in viewer after process exit.

v1.0.2 — Rule Engine + 143-dim Features

GraphReasoner: 90 inference rules, 11 compliance frameworks, baseline-aware scoring, capability-based tool detection. GNN features extended from 38-dim to 143-dim with rule firing bits. 0.133ms per event, 282 tests.

v1.0.1 — Accuracy Hardening

Multi-level GNN detector (4 levels), learned ensemble via logistic regression, 50K synthetic training sessions. Reported AUROC 1.000 on that synthetic set — memorisation of the generator, not deployment accuracy; see the banner on the release page. The GNN will not be built as described.

v1.0.0 — Graph Intelligence Engine

P5 complete. Memgraph graph database, VGAE autoencoder, 5-stage BI Service pipeline, signature distillation, full intelligence loop (proxy → graph → GNN → signature → fleet). 191 BI tests, 11 bugs fixed via bug bounty, 6,577 events/sec throughput.

v0.9.0 — Behavioral Intelligence Engine

P2-P4 complete. Full behavioral scoring pipeline (Gates 1-3, flow sketches, group envelopes), multi-proxy sync (Redis, NATS, serialization), Behavioral Intelligence Service (Python), threat signatures, baseline floors, CA encryption. 38 tickets, 28,700 LOC, 513 tests.

Upcoming