Skip to main content
Five layers, dispatched per tier. The cascade is what turns a raw prompt into a structured intent + reasoning evidence.

​
The layers

  1. L1 — Lexical + adversarial. Always runs. Keyword + regex scoring + injection detection. Cheap; can short-circuit when one intent dominates.
  2. L2 — Embedding similarity. Cosine similarity against per-intent centroids built from the card’s reference prompts.
  3. L3 — Contrastive scoring. Ranks survivors by discriminative margin against the negative set.
  4. L4 — LLM judgment with reasoning. Adjudicates ambiguous cases; produces structured supporting / disqualifying evidence + alternatives_considered.
  5. L5 — Per-org fine-tuned classifier. Ultraviolet only. Trained on your org’s classifier_corrections.

​
Per-tier dispatch

  • Violet — L1 only.
  • Deep Violet — L1 → L2 → L3 → L4.
  • Ultraviolet — L1 → L2 → L3 → L4 → L5.

​
The evidence document

Every dispense + every intent_event row carries a classification_evidence JSONB blob. It captures:
  • The final intent_id + confidence
  • Which layers were consulted
  • Supporting / disqualifying evidence (human-readable strings)
  • Alternatives considered
  • Cost + latency rolled across all layers
  • Adversarial verdict from L1
The Decision Feed on your dashboard renders this as a “Why was this picked?” panel under each row.

​
Confidence thresholds

Each intent card carries a classification block with match_threshold + ambiguous_floor. Confidence above match_threshold accepts directly; between the two it routes through approval (when route_ambiguous_to_approval is true); below the floor it denies.

​
Corrections feed

Reviewers can correct any classification straight from the Decision Feed. Corrections stream into classifier_corrections. Once an Ultraviolet org crosses 250 unique-prompt corrections (or 100 + 30 days since the last fine-tune), the nightly fine-tune driver triggers a new L5 model.