Get started

Stateless stack decision check

Source

Evaluate an inline workload, candidate model list and evaluation measurements against live Witness prices and provider compliance terms in memory, with no API key and zero retention.

text
POST /api/v1/decide
bash
curl -X POST https://vitna.ai/api/v1/decide \
  -H "content-type: application/json" \
  -d '{
    "workload": {
      "id": "ci/support-triage",
      "classId": "model",
      "incumbentSkuId": "openai/gpt-5.6-luna",
      "unitLabel": "tickets",
      "unitProfile": { "inputTokensPerUnit": 2000, "outputTokensPerUnit": 500 },
      "monthlyVolume": 100000,
      "constraints": {
        "qualityFloor": 94.0,
        "economicThresholdPct": 15,
        "edgeCaseVeto": ["pii-leak"],
        "requiresZeroRetention": true
      },
      "priority": "economics"
    },
    "candidates": ["anthropic/claude-sonnet-4-6"],
    "evidence": [
      { "skuId": "openai/gpt-5.6-luna", "measuredAt": "2026-09-14", "metrics": { "successRatePct": 96.0 } },
      { "skuId": "anthropic/claude-sonnet-4-6", "measuredAt": "2026-09-14", "metrics": { "successRatePct": 95.5 } }
    ]
  }'

How candidates are weighed

Every candidate runs through the same pure decide() engine used by aperture:decide:

  1. Eligibility against witness.provider_term (requiresZeroRetention, requiresNoTraining, requiresRegions and allowsProviders), three-provenance capability and runtime speed bounds (requiresCapabilities and minContextTokens over declared_mirror, minThroughputToksSec and maxTtftMs over measured_third_party, and requiresVerifiedCapabilities over customer verifiedCapabilities evidence carrying n and sourceRef), citing the exact term IDs, mirror readings, and customer evidence rows inspected.
  2. Quality floor (candidate.successRatePct >= constraints.qualityFloor).
  3. Critical edge-case vetoes (intersection of reported edgeCaseFailures with constraints.edgeCaseVeto, downgrading otherwise economic candidates to test).
  4. Cost per successful unit over the Witness market tape (savingsRatePct >= economicThresholdPct).

Inline workload profiles, candidate lists and measurements are evaluated purely in memory (Cache-Control: no-store) and never written to disk or the database.

Response shape

json
{
  "workload": {
    "id": "ci/support-triage",
    "name": "Stateless CI check",
    "classId": "model",
    "incumbentSkuId": "openai/gpt-5.6-luna",
    "incumbentName": "GPT-5.6 Luna",
    "unitLabel": "tickets",
    "monthlyVolume": 100000,
    "priority": "economics",
    "activeRequirements": ["requiresZeroRetention"]
  },
  "recommended": {
    "action": "switch",
    "skuId": "anthropic/claude-sonnet-4-6",
    "name": "Claude Sonnet 4.6",
    "savingsRatePct": 24.5,
    "annualSavingsUsd": 3420,
    "reason": "The candidate clears the quality floor with no critical failures and improves cost per successful task by 24.5%."
  },
  "evaluations": [
    {
      "skuId": "anthropic/claude-sonnet-4-6",
      "name": "Claude Sonnet 4.6",
      "evaluable": true,
      "outcome": "switch",
      "reason": "The candidate clears the quality floor with no critical failures and improves cost per successful task by 24.5%.",
      "figures": {},
      "thresholds": {},
      "checks": [],
      "inputs": {
        "pricePeriodIds": [101, 102],
        "enteredPriceIds": [],
        "evidenceIds": [],
        "evidenceSources": [],
        "termIds": [8]
      }
    }
  ],
  "privacy": "Evaluated in memory with zero retention; inline workload profiles, candidates, and evaluation figures are never stored.",
  "meta": {
    "generatedAt": "2026-09-14T18:35:00.000Z",
    "source": "database",
    "count": 1
  }
}