Model realistic deployment

Choose which rules to deploy in each phase, set per-rule target compliance, and see revenue build up phase by phase. Save multiple scenarios and compare them side by side for board presentations.

Quick start: Select a prospect run, click "+ New scenario" to create from defaults, then adjust per-rule parameters. Revenue updates live as you edit.

Client portfolio

Each row is a veterinary group using Sweetfish. Click Dashboard to see performance data, upload monthly reports, and get Gemini analysis.

Class (ADR-035) tells you what the entity is: CUSTOMER (signed pilot or contract — popups flowing), PROSPECT (evaluating; uplift modelled, popups not yet), CORPUS (anonymised reference market for cross-market mining).

Reporting is a separate operational signal — green REPORTING means a monthly performance report has been uploaded; — means awaiting the first upload.

What is this?

Every clinic has tens of thousands of products. Most don't matter for billing accuracy — but a few hundred do, and getting their concept assignment right unlocks every rule downstream. This page ranks the top canonicals to review, so you spend the vet's time on the highest-value ones first.

How the rank is computed: 50% by money-weight (how often it's billed × typical price), 30% by audit flag (the rule auditor sees a problem with this concept), 20% by low embedding-confidence (it doesn't sit close to anything else, so the auto-mapper guessed). Vet-confirmed canonicals are excluded — no point re-reviewing them.

Bulk-edit is dry-run only in Phase 0. The pattern panel below shows what would change without writing anything. Phase 1 adds a commit option with cascade-revoke wiring.

What is this?

Gemini Pro reads stratified samples of real invoices and asks "what is clinically missing from this basket?" It surfaces patterns below the statistical threshold of pattern mining but above what a senior vet notices by inspection. Each finding below is an aggregated cluster — review and decide whether to promote it to a proposed clinical assertion (which can later become a rule), reject it, defer it, or flag it for the next Nic session.

Promote does NOT create a rule — it creates a proposed clinical assertion. Authoring a rule from a confirmed assertion is a separate step with its own dual-gate review.

What is this?

Every night the Stage 4a model scores yesterday's invoices and flags baskets where a rule's required item is probably missing. Each card below is one prediction. Decide whether the model got it right (accept), got it wrong (reject), or you'd like to come back to it later (defer). Every click writes a label that the next weekly retrain consumes — this is the closed-loop channel.

Plain English: p_accept = how confident the model is that this is a genuine miss (0 = definitely fine, 1 = definitely a miss). Predicted value = the typical line value of the missing item on this client. Rule = which rule fired. Missing concept = what the model thinks should have been on the invoice.

Keyboard: J / K to move between cards. A / R / D to accept / reject / defer the focused card. ? for help.

What is this?

Every night the Cross-Market engine reads acceptance data from other live markets (default: IVC ES + IVC PT) and flags rules where the target client's compliance is materially below what other markets have already achieved. Each card is one suggestion: "raise this rule's target — here's the gap-share estimated uplift if you do."

Plain English: Source acceptance = how often the popup is accepted in the source markets. Target current = the target client's current compliance for this rule. Proposed target = what we suggest raising it to. Estimated uplift = the §13 gap-share value (full_potential × (target − current) / (1 − current)) — never full × target. Norwegian-trap score = Build 7 trap-risk filter (we only show suggestions with score < 0.5).

Click ✓ Accept to write the proposed target through to the active scenario. Click ✗ Reject to dismiss. Click Compose Notion ask to copy a Q-S14-NN-style block to the clipboard for posting.

What is this?

Each rule can be scoped to a specific list of clients (markets). A rule with no scope is universal — it fires on every client. A rule with an explicit scope fires only on the listed clients; the engine, the simulator, the cross-market suggester, and the auditor all skip clients NOT in the list.

When to scope: when the rule's base concept is empty on a client (e.g. ivft on IVC Spain, dispensing_fee on EMPET-Headers) and the right fix is "this rule doesn't apply here" rather than "populate the concept".

What this is NOT for: intentional clinical decisions ("Mediterranean parasites don't apply on Nordic clients"). Those need Senior Vet sign-off and a different audit trail; raise them via Notion.

Audit trail: every change emits an audit.rule_events row with action rule.client_scope_changed + before/after JSONB + the reason you supply.

The brain behind the rules

Clinical assertions from your veterinary experts and guidelines. Every assertion feeds into Sweetfish's AI — improving rule generation, blind discovery triage, prospect analysis, and concept mapping.

How to use: Add a knowledge source (e.g. Nick), then paste a conversation transcript or upload a guideline. Gemini extracts structured assertions. Review and confirm each one — confirmed assertions immediately improve every AI call.

What is this?

Statistical mining finds billing patterns nobody wrote as rules. For every concept pair, we compute co-occurrence confidence, lift, and support across all invoices. Gemini then triages each candidate with real invoice samples to classify it as leakage, bundle, clinical variation, or unclear.

Click any row to drill into the evidence: trigger/base products, sample invoices (compliant and violating), and Gemini's reasoning.

What this surface is for

The system has flagged thousands of concept mappings as below 80% confidence, plus 6,765 unmapped products. To make this tractable we filter down to the genuinely impactful rows: mappings whose concept appears in an active or candidate rule, where the product has been billed at least once, plus all unmapped products that show up on real invoices.

Each row is one explicit decision. No bulk apply, no embedding propagation. Choices are: Keep (concept is right, just confirm) · Move (concept is wrong, pick a better one) · Unmap (this product shouldn't be in any concept). Every action lands in audit.rule_events and is reversible.

What are rule anomalies?

If a rule's acceptance rate suddenly jumps or dips (by 10 percentage points or more week-over-week), something changed — a locum joined, a new accounting code appeared, a PMS update mangled a product category. Anomalies surface those jumps so you can investigate.

How to use: click Diagnose on a row; Gemini reads the before/after stats for that rule and returns a plain-English explanation. You don't need clinical knowledge — the diagnosis is written for an admin audience.

Discovery Lab

Statistical mining combined with Gemini triage surfaces candidate billing patterns. Each card below is a pattern found in client data. Promote a pattern to make it a live rule, Investigate to drill deeper, or Reject to dismiss it.