Adaptive routing

Provara's adaptive router doesn't pre-train a classifier on labeled data. Instead, every piece of quality feedback — user 1–5 ratings and LLM-as-judge scores — nudges a per-cell EMA (exponential moving average). Over time, models that actually produce good answers win more traffic in the cells they're good at.

Cells

A cell is the tuple (task_type, complexity). Provara classifies every incoming prompt into one of 15 cells (5 task types × 3 complexities):

• task_type: coding, creative, summarization, qa, general
• complexity: simple, medium, complex

The classifier is heuristic (keyword + length + pattern signals). It's intentionally simple — adaptive scoring is the layer that makes routing smart. If a model consistently wins on coding + complex, the router learns that from outcomes, regardless of whether the classifier's "complex" label is perfect.

EMA update

On every judged response:

new_score = alpha * judge_score + (1 - alpha) * old_score
sample_count += 1

alpha defaults to 0.1 — new samples nudge the score but don't flip it overnight. Stored in model_scores keyed by (tenant_id, task_type, complexity, provider, model), persisted across restarts.

Routing decision

Given a request and its classified cell, the router:

1. Filters candidates to models with sample_count >= MIN_SAMPLES (default 5)
2. Applies routing weights {quality, cost, latency} — defaults per routing profile (cost, balanced, quality), overridable per-token
3. Scores each candidate: weights.quality * norm(score) + weights.cost * norm(cost) + weights.latency * norm(latency)
4. Picks the top score

ε-greedy exploration bypasses the EMA with probability PROVARA_EXPLORATION_RATE (default 0.1) and picks uniformly at random, so one model can't win a cell forever without alternatives getting tested.

Stale cell detection

A cell is stale when its most recent update is older than PROVARA_STALE_AFTER_DAYS (default 30). Stale cells get a boosted exploration rate (PROVARA_STALE_EXPLORATION_RATE, default 0.5) — when traffic arrives, half the time we'll explore off the stale winner. Forces ground-truth refresh without silently trusting an EMA that hasn't updated in months. Dashboard renders stale cells with an amber badge.

Tuning

What this isn't

• Not a pre-trained classifier. No training step to maintain, no re-training when new models ship. Quality converges from real outcomes.
• Not federated. EMAs are tenant-scoped; no cross-tenant pooling.
• Not real-time ML. The EMA is a classical online-learning formula — transparent, cheap, auditable.