{ "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": "Prune the 12002 un-cached tokens Meeting Summaries re-sends on every call — retrieved context, conversation history, repeated blocks", "applicationId": "meeting-summaries", "beforeAfter": { "addedEdges": [], "addedNodes": [], "after": "Meeting Summaries: Prune the 12002 un-cached tokens Meeting Summaries re-sends on every call — retrieved context, conversation history, repeated blocks", "before": "Meeting Summaries: full-length system prompts on every request", "changedEdges": [ "app:meeting-summaries" ] }, "benchmark": { "context": "Average input tokens per request (lower is leaner prompts). (illustrative)", "median": 2200, "metric": "avg_input_tokens", "topQuartile": 1400, "unit": "tokens", "yourValue": 12002 }, "businessValueEur": 9573.02, "complexity": "Low", "confidence": 0.7139, "contextCompleteness": { "absent": [ "Codebase", "Quality evaluation" ], "present": [ "Telemetry", "Provider bill", "Per-app criticality", "Prior experiments" ], "presentCount": 4, "totalDims": 6 }, "costModelBasis": { "agentRunAllowanceEur": 20, "byHand": { "effortCostEur": 2400, "effortDays": 3, "engineeringDayRateEur": 800 }, "effortCostEur": 320, "rateBasis": "prior", "reviewHourRateEur": 100, "reviewHours": 3, "reviewHoursBasis": "prior", "version": "agent-assisted-v1-2026-08" }, "criticalReview": { "alternativeExplanations": [ "The large prompt may be load-bearing context the model actually uses; trimming it would degrade quality, not just cost." ], "assumptions": [ "Assumes ~15% of the prompt is redundant boilerplate that can be pruned without changing answers." ], "failureConditions": [ "Wrong if the trimmed context was genuinely used by the model for correctness.", "This figure counts ONLY the un-cached surface. Shortening the cached prefix instead reclaims far less — those tokens already bill at the cache-read rate — and forces a re-warm at the write premium, which can cost more than the trim saves.", "Wrong if this workload's real quality bar is higher than assumed and the change degrades answers." ], "missingEvidence": [ "No before/after quality evaluation — the compressible fraction is a heuristic, not a measured safe-to-cut amount.", "No independent quality evaluation — the quality impact is unmeasured." ] }, "demoted": null, "effort": "Low", "evidenceIds": [ "ev:meeting-summaries:prompt_compression" ], "expectedImpact": { "annualSavingsEur": 6952.22, "costDeltaPct": -0.1333, "latencyDeltaPct": -0.05, "qualityDeltaPct": null, "qualityRisk": "medium" }, "id": "rec:meeting-summaries:prompt_compression", "inferenceId": "inf:meeting-summaries:prompt_compression", "latencyValueEur": 2620.8, "memoryNote": "Learned caution: only 1/3 prior implementations succeeded (N=51,000); confidence penalized.", "objectiveKey": "lowest_cost", "observationIds": [ "obs:meeting-summaries:avg_input_tokens" ], "pareto": true, "paybackDays": 17, "priceBasis": "scenario", "priorityScore": 5293.99, "qualityOfEvidence": "High", "roi": { "affectedRequestsMonthly": 30000, "annualSavingsEur": 6952.22, "effortDays": 3, "monthlySavingsEur": 579.35, "paybackDays": 17, "verified": false }, "ruleId": "prompt_compression", "sampleSize": 24000, "targetNodeId": "app:meeting-summaries", "title": "Compress the un-cached context Meeting Summaries re-sends every call", "whyNot": [ { "costDeltaPct": -0.28, "latencyDeltaPct": -0.1, "option": "Aggressively truncate context", "paretoOptimal": false, "qualityDeltaPct": null, "qualityRisk": "high", "reason": "Bigger cost cut, but an aggressive prune carries a materially higher unmeasured quality risk on this workload — the quality effect of removing context has never been measured here." }, { "costDeltaPct": 0.05, "latencyDeltaPct": 0.2, "option": "LLM summarization pre-pass", "paretoOptimal": false, "qualityDeltaPct": 0, "qualityRisk": "none", "reason": "Adds an extra model call — raises cost and latency, defeating the purpose." } ] }, "decisionHash": "sha256:d06cf5bfe8e73257454b0322e3b10273cd4b2a632690567d06a068598537ae73", "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": "Meeting Summaries re-sends a large un-cached context on every call and pays full input rate for it; pruning ~15% of that surface cuts 13.3% of this application’s spend. Estimated €6,952/yr at the current run-rate, payback 17 days, confidence 71%.", "technical": "Dedupe retrieved context, cap conversation history, and remove repeated blocks from the PER-CALL portion of the prompt — measure quality before/after. Affects ~30,000 req/mo; expected cost -13.3%, latency -5%, quality unmeasured (medium risk).", "phrasedBy": "mock", "provenance": { "narrativeProvenanceVersion": 1, "attempted": "mock", "policyVersion": "fabric-v1-2026-08", "source": "provider", "degraded": false } }, "ledgerState": "overlapping_alternative", "ledgerVersion": "22c35e346df765c3", "overlapGroup": { "applicationId": "meeting-summaries", "memberIds": [ "rec:meeting-summaries:migration_laggard", "rec:meeting-summaries:overpowered_model", "rec:meeting-summaries:prompt_compression" ], "leadId": "rec:meeting-summaries:migration_laggard" }, "simulation": { "formulaVersion": "sim-v1", "formula": "realization = 0.7 + 0.3 × confidence (scenario assumption)", "inputs": { "baseAnnualSavingsEur": 6952.22, "confidence": 0.7139 }, "intermediates": { "realization": 0.914 }, "outputs": { "annualSavingsEur": 6355.51, "downsideAnnualSavingsEur": 4866.55, "upsideAnnualSavingsEur": 6952.22 } }, "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": "gpt-4-turbo", "rates": { "inputPer1M": 10, "outputPer1M": 30, "cacheReadPer1M": 10, "cacheWritePer1M": 0 }, "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": "2026-08-28T18:51:37.397Z" } }