{ "format": "bayeto-decision-record", "formatVersion": 3, "formatNote": "v3 adds decisionHash. Additive: v2 records remain valid and re-derivable — no v2 field changed meaning, moved, or was removed.", "record": { "action": "Route the low-output claude-sonnet traffic in Client Reporting down one tier to claude-haiku", "applicationId": "client-reporting", "beforeAfter": { "addedEdges": [], "addedNodes": [], "after": "Client Reporting: Route the low-output claude-sonnet traffic in Client Reporting down one tier to claude-haiku", "before": "Client Reporting: all traffic on the current premium model", "changedEdges": [ "model:anthropic/claude-sonnet" ] }, "benchmark": null, "businessValueEur": 1457.34, "complexity": "Low", "confidence": 0.65, "contextCompleteness": { "absent": [ "Codebase", "Quality evaluation" ], "present": [ "Telemetry", "Provider bill", "Per-app criticality", "Prior experiments" ], "presentCount": 4, "totalDims": 6 }, "costModelBasis": { "agentRunAllowanceEur": 20, "byHand": { "effortCostEur": 1600, "effortDays": 2, "engineeringDayRateEur": 800 }, "effortCostEur": 220, "rateBasis": "prior", "reviewHourRateEur": 100, "reviewHours": 2, "reviewHoursBasis": "prior", "version": "agent-assisted-v1-2026-08" }, "criticalReview": { "alternativeExplanations": [ "The large prompt may be an intentional cached system prefix (architecture), not model over-provisioning — the ratio alone can't tell these apart.", "The cost is spread across token types; a caching change would save little here." ], "assumptions": [ "Assumes claude-haiku preserves answer quality for this workload — quality is UNMEASURED here.", "Assumes the small answers reflect task simplicity; output is 55.7% of this segment's cost (a material share), so the tier premium is genuinely being paid on cheap work.", "Assumes the segment is cleanly switchable in full; the €1457 is net of a cache re-warm amortized over the observed reuse (12.5×)." ], "failureConditions": [ "Wrong if this workload needs claude-sonnet-tier reasoning for correctness (e.g. a compliance-critical answer).", "Wrong if the segment can't be switched as a whole (per-turn routing) — cache-thrash would erase the saving.", "Wrong if this workload's real quality bar is higher than assumed and the change degrades answers." ], "missingEvidence": [ "No quality evaluation on this segment — the downgrade's quality impact is a risk estimate, not a measurement.", "No independent quality evaluation — the quality impact is unmeasured." ] }, "demoted": "low-context-critical", "effort": "Low", "evidenceIds": [ "ev:client-reporting:overpowered_model" ], "expectedImpact": { "annualSavingsEur": 1457.34, "costDeltaPct": -0.4662, "latencyDeltaPct": 0, "qualityDeltaPct": null, "qualityRisk": "medium" }, "id": "rec:client-reporting:overpowered_model", "inferenceId": "inf:client-reporting:overpowered_model", "latencyValueEur": 0, "memoryNote": "No prior implementations yet — confidence derived from evidence only.", "objectiveKey": "highest_quality", "observationIds": [ "obs:client-reporting:avg_input_tokens" ], "pareto": true, "paybackDays": 56, "priceBasis": "scenario", "priorityScore": 947.27, "qualityOfEvidence": "Medium", "roi": { "affectedRequestsMonthly": 16140, "annualSavingsEur": 1457.34, "effortDays": 2, "monthlySavingsEur": 121.45, "paybackDays": 56, "verified": false }, "ruleId": "overpowered_model", "sampleSize": 14400, "targetNodeId": "model:anthropic/claude-sonnet", "title": "Right-size Client Reporting: move low-output claude-sonnet traffic to claude-haiku", "whyNot": [ { "costDeltaPct": -0.0035, "latencyDeltaPct": 0, "option": "Cut cache-write cost instead (keep claude-sonnet)", "paretoOptimal": false, "qualityDeltaPct": 0, "qualityRisk": "none", "reason": "Cache-writes are a small share here, so a caching change saves little." }, { "costDeltaPct": 0, "latencyDeltaPct": 0, "option": "Keep claude-sonnet (no change)", "paretoOptimal": false, "qualityDeltaPct": 0, "qualityRisk": "none", "reason": "Quality-safe but forgoes the saving; justified only if this traffic genuinely needs top-tier reasoning." } ] }, "decisionHash": "sha256:1efa322c757fc2aa0883da1b6eb708bfda76e92a6df0e989b07f3e3dfb930bfd", "narrative": { "note": "Phrased by the AI Fabric from deterministic drafts. Not part of the decision; changing it cannot change any number above. `provenance` records which provider source control approved, whether it served, and why not — source:\"draft\" means you are reading the engine's own deterministic wording.", "business": "Client Reporting pays claude-sonnet rates to produce tiny answers where output is a material 55.7% of cost — the hallmark of an over-provisioned model. Right-sizing to claude-haiku keeps the same work at a lower tier. Estimated €1,457/yr at the current run-rate, payback 56 days, confidence 65%.", "technical": "Route this segment to claude-haiku. Expected quality change is UNKNOWN — not measured; validate with a canary / Simulate on a traffic slice before full rollout, and roll back if graded quality drops. Prompt caches are per-model, so switch the WHOLE segment at once (never per-turn) to avoid cache-thrash. Affects ~16,140 req/mo; expected cost -46.6%, latency +0%, quality unmeasured (medium risk).", "phrasedBy": "mock", "provenance": { "narrativeProvenanceVersion": 1, "attempted": "mock", "policyVersion": "fabric-v1-2026-08", "source": "provider", "degraded": false } }, "ledgerState": "held", "ledgerVersion": "22c35e346df765c3", "overlapGroup": { "applicationId": "client-reporting", "memberIds": [ "rec:client-reporting:prompt_cache_adoption", "rec:client-reporting:overpowered_model" ], "leadId": "rec:client-reporting:prompt_cache_adoption" }, "simulation": { "formulaVersion": "sim-v1", "formula": "realization = 0.7 + 0.3 × confidence (scenario assumption)", "inputs": { "baseAnnualSavingsEur": 1457.34, "confidence": 0.65 }, "intermediates": { "realization": 0.895 }, "outputs": { "annualSavingsEur": 1304.32, "downsideAnnualSavingsEur": 1020.14, "upsideAnnualSavingsEur": 1457.34 } }, "measured": null, "run": { "engineVersion": "0.1.0", "workspaceSourceKind": "demo", "stateVersion": "none|none|0|0:none", "datasetHash": "c8030438f36955eebab44a8b2fe9a468df7d9592992ba71e8145de0a9965ce23", "pricing": { "catalogVersion": "thornbury-scenario-v0", "unresolved": null, "rows": [ { "family": "claude-haiku", "rates": { "inputPer1M": 1, "outputPer1M": 5, "cacheReadPer1M": 0.1, "cacheWritePer1M": 1.25 }, "provenance": { "source": "bayeto-scenario", "capturedAt": "2026-07-21", "quality": "scenario" } }, { "family": "claude-sonnet", "rates": { "inputPer1M": 2, "outputPer1M": 10, "cacheReadPer1M": 0.2, "cacheWritePer1M": 2.5 }, "provenance": { "source": "bayeto-scenario", "capturedAt": "2026-07-21", "quality": "scenario" } } ] }, "windowDays": 24, "windowStart": "2026-06-06T00:00:00.000Z", "windowEnd": "2026-06-29T23:58:51.723Z", "coverage": { "rows": 124800, "activeDays": 24, "unattributedRows": 0 }, "fxRates": { "USD_EUR": 0.92, "GBP_EUR": 1.17, "effectiveDate": "2026-07-06", "basis": "fixed reference rates (determinism; never live FX)" }, "implementationCostModel": { "version": "agent-assisted-v1-2026-08", "reviewHourRateEur": 100, "agentRunAllowanceEur": 20, "basis": "agent-assisted priors (review hours × rate + agent-run allowance); byHand comparison at the blended day rate" }, "engineeringDayRateEur": 800, "evidenceTaxonomyVersion": "evtax-v1", "formulaVersions": { "accuracy": "pvm-v1", "simulation": "sim-v1" }, "generatedAt": "2026-08-28T18:27:33.628Z" } }