AI Quoting: ABI semantics, 1k-unit tiers, and Transfer guardrails

Locks down three classes of model error in the Dealops 2 AI quoting pipeline — ABI shorthand, malformed price tiers, and Transfer rail/risk/mode bundles — at both the prompt and deterministic-guardrail layers.

1. Why this exists

The AI quoting agent (Dealops 2, apps/server/src/dealops2/aiQuoting/pipeline) was producing three recurring classes of bad quotes that the existing guardrails couldn't catch:

The fix is symmetric: add the rules to the three prompts (quote_generation.v1.md, quote_validation.v1.md, quote_validate_correct.v1.md) so the model knows them, and add deterministic checks in validateAiRecommendationAgainstRules.ts and generateV2Quote.ts so the model can't slip past them.

2. The three new rule families

2.1 ABI means exactly three products

Detection lives in aiQuoteRuleHelpers.ts:

export function contextMentionsAbiProducts(text: string): boolean {
  const lower = text.toLowerCase();
  return (
    /\babi\b/.test(lower) ||
    (/\bauth\b/.test(lower) &&
      /\bbalance\b/.test(lower) &&
      /\bidentity\b/.test(lower))
  );
}

The validator walks ABI_PRODUCT_NAMES = ['Auth', 'Balance', 'Identity'], looks each up in the catalog, and emits a hard issue per missing product. The fixture transcripts in fetchOpportunityContext.ts were updated so the demo opp explicitly says "ABI scope: Auth, Balance, and Identity only" to exercise the rule end-to-end.

2.2 Price tiers must use strict 1,000-unit bands

The old check in generateV2Quote.ts only verified two things: tiers don't overlap, and the highest tier covers up to a 1M sentinel. Anything in between — like 0–999 / 1001–10M — slipped through.

The new rule, in validateStrictPriceTiers, demands:

Both isTierShapeValid (Stage 5 mapping) and hasValidPriceTiers (Stage 4 validator) now delegate to the shared helper, so the rule is enforced exactly once in code.

2.3 Transfer rail × risk × pricing mode

Transfer SKUs have a structured name that encodes three dimensions. The new parsers in aiQuoteRuleHelpers.ts split them out and the validator reasons about the bundle as a whole.

// Matches: "Transfer (Standard ACH) - Vanilla Originators - bps"
/^Transfer \((Standard ACH|Same-day ACH)\) - (Vanilla Originators|High Risk Customers) - (bps|Fixed)$/i

// Matches: "Transfer (Incoming Wire) - Vanilla Originators"
/^Transfer \((All other returns|Incoming Wire|NOC|Unauthorized returns)\) - (Vanilla Originators|High Risk Customers)$/i

Given a parsed base product with a comboKey of rail::riskTier, four checks run:

The pricing-mode band:

< $100   |   $100  ←  bps band  →  $600   |   > $600

ATS extraction is regex-based against the assembled context text (opp name, primary product, use case, secondary product, technical use cases, plus every Gong call's subject/brief/key points/next steps/full description). Three patterns catch variants like "average transaction size is $250", "txn size: ~250 avg", and "$250 avg ticket".

3. Where the rules fire

Each rule is enforced at three layers — prompt, LLM validator, deterministic code — so a regression at one layer doesn't ship a bad quote.

4. Transfer bps → percent pricing wiring

The other half of the Transfer story isn't validation — it's mapping. Previously every product became a quotePriceFlat. bps Transfer SKUs need quotePricePercent with floor/ceiling caps, and an avgTransactionSize term carried alongside.

The mapping path now branches on a new catalog field calculationSpec.bpsPercentPricing:

function buildProductPricingInput(params) {
  // ...
  if (isBpsPercentCatalogProduct(productEntry)) {
    return {
      quotePricePercent: buildQuotePricePercent({ /* rate + floor + ceiling */ }),
      ...(terms ? { terms } : {}),
    };
  }
  return {
    quotePriceFlat: buildQuotePrice(/* ... */),
    ...(terms ? { terms } : {}),
  };
}

The resulting product payload for a Transfer bps SKU looks like:

{
  quotePricePercent: {
    type: 'percentWithLimits',
    value:        { type: 'percent',  value: 0.1 },
    perTxnFloor:  { type: 'currency', value: 0.01, currency: 'USD' },
    perTxnCeiling:{ type: 'currency', value: 2,    currency: 'USD' },
  },
  terms: [{
    type: 'termInput',
    id: 'avgTransactionSize',
    variableId: 'avgTransactionSize',
    userValue: { type: 'currency', value: 250, currency: 'USD' },
  }],
}

ProductInput in generateV2Quote.ts gained two new optional fields (quotePricePercent, terms) and finalizeAiQuote.ts propagates them through quote creation. ProductCatalogEntry in generateAiQuote.ts gained a structured calculationSpec carrying requiresTransactionSize, bpsPercentPricing, bpsPerTxnFloor, bpsPerTxnCeiling, and a transactionSizeBand.

5. New issue codes and what they tell the model

Each new code maps to a human-readable fallback in generateValidatedAiRecommendation.ts and a "feedback" string the corrector sees on the next pass:

6. New shared module: aiQuoteRuleHelpers.ts

The PR consolidates rule logic that was previously duplicated between Stage 4 (validation) and Stage 5 (V2 mapping) into one file. Both stages now import from it.

7. Test coverage added

Two test files grow significantly; one is rewritten to match the new tier rules.

8. What it doesn't change

9. Risks and open questions