{
  "schemaVersion": 1,
  "measuredAt": "2026-09-12T15:07:04.268Z",
  "benchmark": "voice-writing-review-authored-v1",
  "catalogueVersion": "voice-writing-rules/0.2.1",
  "execution": {
    "offline": true,
    "providerRequestsSent": 0,
    "runnerExecutions": 0,
    "modelResponsesEvaluated": 0
  },
  "provenance": {
    "fixtureSha256": "caa45260f4c077434ec76098b8e9a57585111d7c604549fd61f6a4db22ca29e4",
    "runnerSha256": "010df8edeb97ac298af1e856ee59ea164d44737f51f086203259fb96a380a638",
    "librarySha256": "ce63fd5588ab37e945d27c224c66131e578464629568616358cb404f2f138eb7",
    "authoredFixtureNotice": "Fixtures authored for this offline benchmark. These are editorial test cases, not sampled model responses, a representative corpus, or independent ground truth.",
    "labelling": "expectedReview lists the exact passage and rule that an editor judged worth reviewing. An empty list means keep this fixture in its stated context. A scanner candidate on a keep fixture is a false positive for this editorial decision, not a scanner malfunction."
  },
  "tokenizer": {
    "status": "unavailable",
    "attempts": [
      {
        "package": "js-tiktoken",
        "status": "unavailable",
        "reason": "MODULE_NOT_FOUND"
      },
      {
        "package": "tiktoken",
        "status": "unavailable",
        "reason": "MODULE_NOT_FOUND"
      }
    ],
    "meaning": "No token estimates. UTF-16 code units and UTF-8 bytes are measured directly and are not token counts."
  },
  "promptMeasurement": {
    "audience": "People deciding how to use a local text-review product, including editors and application developers.",
    "goal": "Explain the current product behavior and the next useful action. Preserve factual technical explanations and explicitly quoted examples. Review wording in its supplied context; do not infer authorship.",
    "maxFindingsPerRequest": 3,
    "format": "JSON.stringify({ instructions: composed.prompt, source: { language, context, text } })",
    "sameCoverageComparisons": [
      {
        "separate": "all-rules-separate",
        "together": "all-rules-together",
        "sameRuleCoverage": true,
        "requestReduction": 570,
        "payloadUtf8BytesSaved": 741931,
        "payloadUtf8ReductionFraction": 0.5791652843208812,
        "repeatedSourceUtf8BytesSaved": 107806,
        "encodedTokensSaved": null
      },
      {
        "separate": "profile-separate",
        "together": "profile-together",
        "sameRuleCoverage": true,
        "requestReduction": 150,
        "payloadUtf8BytesSaved": 195245,
        "payloadUtf8ReductionFraction": 0.5048691056153742,
        "repeatedSourceUtf8BytesSaved": 28370,
        "encodedTokensSaved": null
      }
    ],
    "summaries": [
      {
        "id": "all-rules-separate",
        "label": "One request per rule, all rules",
        "ruleCountPerFixture": 20,
        "fixtureCount": 30,
        "totalRequests": 600,
        "totalInstructions": {
          "utf16CodeUnits": 1132905,
          "utf8Bytes": 1141155
        },
        "totalRepeatedSource": {
          "utf16CodeUnits": 112520,
          "utf8Bytes": 113480
        },
        "totalPayload": {
          "utf16CodeUnits": 1271825,
          "utf8Bytes": 1281035
        },
        "totalEncodedTokens": null
      },
      {
        "id": "all-rules-together",
        "label": "One request, all rules",
        "ruleCountPerFixture": 20,
        "fixtureCount": 30,
        "totalRequests": 30,
        "totalInstructions": {
          "utf16CodeUnits": 522435,
          "utf8Bytes": 526980
        },
        "totalRepeatedSource": {
          "utf16CodeUnits": 5626,
          "utf8Bytes": 5674
        },
        "totalPayload": {
          "utf16CodeUnits": 534511,
          "utf8Bytes": 539104
        },
        "totalEncodedTokens": null
      },
      {
        "id": "profile-separate",
        "label": "One request per public-docs rule",
        "ruleCountPerFixture": 6,
        "fixtureCount": 30,
        "totalRequests": 180,
        "totalInstructions": {
          "utf16CodeUnits": 342195,
          "utf8Bytes": 344760
        },
        "totalRepeatedSource": {
          "utf16CodeUnits": 33756,
          "utf8Bytes": 34044
        },
        "totalPayload": {
          "utf16CodeUnits": 383871,
          "utf8Bytes": 386724
        },
        "totalEncodedTokens": null
      },
      {
        "id": "profile-together",
        "label": "One request, public-docs profile",
        "ruleCountPerFixture": 6,
        "fixtureCount": 30,
        "totalRequests": 30,
        "totalInstructions": {
          "utf16CodeUnits": 181545,
          "utf8Bytes": 183135
        },
        "totalRepeatedSource": {
          "utf16CodeUnits": 5626,
          "utf8Bytes": 5674
        },
        "totalPayload": {
          "utf16CodeUnits": 189841,
          "utf8Bytes": 191479
        },
        "totalEncodedTokens": null
      }
    ],
    "limitations": [
      "The six-rule profile has a smaller scope than all twenty rules. Their cost difference is not an equal-coverage comparison.",
      "The per-request finding cap is identical, so separate requests can yield more total findings. No output length or review quality was measured.",
      "Counts include repeated source context and actual composed instructions, but no provider message framing, caching, retries, latency, hidden reasoning or output.",
      "Prompt batching can reduce repeated input while changing reviewer behavior. This offline run does not establish the best batch size or model strategy."
    ]
  },
  "surfaceEvaluation": {
    "fixtureCount": 30,
    "candidateCount": 24,
    "candidatesAgreeingWithEditorialLabel": 16,
    "candidatesNotAgreeingWithEditorialLabel": 8,
    "candidatePrecisionAgainstAuthoredLabels": 0.6666666666666666,
    "lexicalProblemCases": 16,
    "lexicalProblemCasesWithExpectedCandidate": 16,
    "keepOrCleanCases": 10,
    "keepOrCleanCasesWithCandidate": 8,
    "semanticProblemCases": 4,
    "semanticProblemCasesWithNoCandidate": 4,
    "byLanguage": [
      {
        "language": "en",
        "cases": 15,
        "candidates": 12,
        "agreeingCandidates": 8,
        "coverage": {
          "language": "en",
          "scannedRuleIds": [
            "meta-framing",
            "concrete-subject",
            "unsupported-contrast",
            "calibrated-uncertainty",
            "unsupported-causality",
            "claims-evidence",
            "word-choice",
            "reported-evidence"
          ],
          "unscannedRuleIds": [
            "reader-goal",
            "current-state",
            "process-history",
            "invented-labels",
            "repetition",
            "format-fit",
            "complete-sentences",
            "task-relevance",
            "unearned-praise",
            "false-balance",
            "exhaustive-checklist",
            "forced-template"
          ],
          "semanticReviewRequired": true,
          "maxSignals": 50,
          "truncated": false
        }
      },
      {
        "language": "de",
        "cases": 15,
        "candidates": 12,
        "agreeingCandidates": 8,
        "coverage": {
          "language": "de",
          "scannedRuleIds": [
            "meta-framing",
            "concrete-subject",
            "unsupported-contrast",
            "calibrated-uncertainty",
            "unsupported-causality",
            "claims-evidence",
            "word-choice",
            "reported-evidence"
          ],
          "unscannedRuleIds": [
            "reader-goal",
            "current-state",
            "process-history",
            "invented-labels",
            "repetition",
            "format-fit",
            "complete-sentences",
            "task-relevance",
            "unearned-praise",
            "false-balance",
            "exhaustive-checklist",
            "forced-template"
          ],
          "semanticReviewRequired": true,
          "maxSignals": 50,
          "truncated": false
        }
      }
    ],
    "limitations": [
      "These authored examples are deliberately balanced for demonstration. Candidate precision is local to this fixture set and is not model accuracy or representative product quality.",
      "The scanner sees only text and language, not the supplied editorial context. Quoted examples and warranted causal statements therefore remain candidates.",
      "Missed semantic cases require a reviewer. No candidates never establishes that a document is good, correct or ready to publish.",
      "The catalogue currently exposes lexical patterns for eight rules. The returned unscannedRuleIds identify the twelve rules without lexical coverage."
    ]
  },
  "modelPricing": {
    "schemaVersion": 1,
    "checkedAt": "2026-09-12",
    "scope": "Candidate text-review APIs and public USD token rates. Documentation research only: no inference requests, account-access checks or quality benchmark were run.",
    "qualification": "All candidates remain unqualified for Voice's English and German writing rules. Provider positioning selects a test cohort; it does not establish detection accuracy or a quality ranking.",
    "sources": [
      {
        "id": "openai-models",
        "url": "https://developers.openai.com/api/docs/models",
        "read": "Current model catalogue and efficiency/flagship positioning."
      },
      {
        "id": "openai-luna",
        "url": "https://developers.openai.com/api/docs/models/gpt-5.6-luna",
        "read": "Model ID, snapshot, context limits, supported endpoints, features and reasoning levels."
      },
      {
        "id": "openai-astra",
        "url": "https://developers.openai.com/api/docs/models/gpt-6-astra",
        "read": "Model ID, snapshot, context limits, supported endpoints and pricing notes."
      },
      {
        "id": "openai-pricing",
        "url": "https://developers.openai.com/api/docs/pricing",
        "read": "Standard and Batch text-token tables, including short/long context and cache writes."
      },
      {
        "id": "openai-cache",
        "url": "https://developers.openai.com/api/docs/guides/prompt-caching",
        "read": "GPT-5.6-and-later cache minimum, prefix matching, explicit breakpoints, write charging and lifetime."
      },
      {
        "id": "openai-batch",
        "url": "https://developers.openai.com/api/docs/guides/batch",
        "read": "Asynchronous processing, discount, 24-hour completion window and expiration state."
      },
      {
        "id": "openai-reasoning",
        "url": "https://developers.openai.com/api/docs/guides/reasoning",
        "read": "Reasoning-token billing and incomplete responses at output limits."
      },
      {
        "id": "claude-haiku",
        "url": "https://platform.claude.com/docs/en/models/haiku-4-5/overview",
        "read": "Pinned API ID, alias, active status, context limits, manual extended thinking and rates."
      },
      {
        "id": "claude-pricing",
        "url": "https://platform.claude.com/docs/en/about-claude/pricing",
        "read": "First-party Standard/Batch rates and cache-write/read prices."
      },
      {
        "id": "claude-cache",
        "url": "https://platform.claude.com/docs/en/build-with-claude/prompt-caching",
        "read": "Haiku cache minimum, TTLs, usage fields, concurrent cache availability and billing multipliers."
      },
      {
        "id": "claude-batch",
        "url": "https://platform.claude.com/docs/en/build-with-claude/batch-processing",
        "read": "Asynchronous processing, expiration and best-effort cache hits when batching."
      },
      {
        "id": "google-flash-lite",
        "url": "https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite",
        "read": "Model code, throughput positioning, input/output limits and structured output/caching capabilities."
      },
      {
        "id": "google-pricing",
        "url": "https://ai.google.dev/gemini-api/docs/pricing#gemini-3.5-flash-lite",
        "read": "Paid Standard/Batch text rates, thinking-token inclusion, cache rate and storage charges."
      },
      {
        "id": "google-cache",
        "url": "https://ai.google.dev/gemini-api/docs/caching",
        "read": "Implicit caching, prefix guidance, API distinction and published minimum-token table."
      },
      {
        "id": "google-batch",
        "url": "https://ai.google.dev/gemini-api/docs/batch-api",
        "read": "generateContent-only batching, turnaround target and cache-pricing statement."
      }
    ],
    "candidates": [
      {
        "id": "openai-luna",
        "provider": "OpenAI",
        "modelId": "gpt-5.6-luna",
        "documentedSnapshot": "gpt-5.6-luna",
        "role": "Low-cost candidate for bounded rule review; provider describes it as the cost-sensitive, high-volume tier.",
        "availability": "Listed in the current API catalogue. Access in this environment was not tested.",
        "context": {
          "windowTokens": 1050000,
          "maxInputTokens": 922000,
          "maxOutputTokens": 128000
        },
        "pricing": {
          "currency": "USD",
          "unit": "per_1000000_tokens",
          "scope": "First-party text API, Standard or Batch, input at most 272000 tokens; no tools or regional/priority modifiers.",
          "standard": {
            "input": 0.2,
            "cacheRead": 0.02,
            "cacheWrite": 0.25,
            "output": 1.2
          },
          "batch": {
            "input": 0.1,
            "cacheRead": 0.01,
            "cacheWrite": 0.125,
            "output": 0.6
          },
          "longContextStandard": {
            "inputAboveTokens": 272000,
            "input": 0.4,
            "cacheRead": 0.04,
            "cacheWrite": 0.5,
            "output": 1.8
          }
        },
        "cache": {
          "minimumPrefixTokens": 1024,
          "minimumLifetimeMinutes": 30,
          "note": "A matching eligible prefix is required. Writes replace the ordinary input rate for those tokens; they are not an additional charge on top of it."
        },
        "benchmarkStart": "Compare reasoning.effort none and low while keeping the input, rule scope and output contract fixed; measure reasoning-inclusive usage.",
        "sourceIds": [
          "openai-models",
          "openai-luna",
          "openai-pricing",
          "openai-cache",
          "openai-reasoning"
        ]
      },
      {
        "id": "claude-haiku",
        "provider": "Anthropic",
        "modelId": "claude-haiku-4-5-20251001",
        "alias": "claude-haiku-4-5",
        "documentedSnapshot": "claude-haiku-4-5-20251001",
        "role": "Anthropic efficiency-tier candidate for a separate provider baseline.",
        "availability": "Official page marks this snapshot active; account access was not tested. The page gives no retirement sooner than 2026-10-15, not a promised retirement date.",
        "context": {
          "windowTokens": 200000,
          "maxInputTokens": null,
          "maxOutputTokens": 64000
        },
        "pricing": {
          "currency": "USD",
          "unit": "per_1000000_tokens",
          "scope": "First-party Claude API; no partner, residency or other modifiers.",
          "standard": {
            "input": 1,
            "cacheRead": 0.1,
            "cacheWrite5m": 1.25,
            "cacheWrite1h": 2,
            "output": 5
          },
          "batch": {
            "input": 0.5,
            "output": 2.5,
            "cacheRead": 0.05,
            "cacheWrite5m": 0.625,
            "cacheWrite1h": 1
          },
          "batchCacheBasis": "Derived from the documented 50% Batch discount stacking with cache multipliers, not measured invoices."
        },
        "cache": {
          "minimumPrefixTokens": 4096,
          "lifetimeMinutes": [
            5,
            60
          ],
          "note": "Shorter marked prefixes do not cache. Concurrent requests may miss until the first response starts; Batch cache hits are best effort."
        },
        "benchmarkStart": "Start without optional extended thinking; test a separate, explicitly configured thinking condition only if it improves judged quality enough to justify its measured cost.",
        "sourceIds": [
          "claude-haiku",
          "claude-pricing",
          "claude-cache",
          "claude-batch"
        ]
      },
      {
        "id": "google-flash-lite",
        "provider": "Google",
        "modelId": "gemini-3.5-flash-lite",
        "documentedSnapshot": null,
        "role": "Google high-throughput candidate for short reviews and structured findings.",
        "availability": "Listed with a model code and paid API pricing; account access and an immutable snapshot ID were not established.",
        "context": {
          "windowTokens": null,
          "maxInputTokens": 1048576,
          "maxOutputTokens": 65536
        },
        "pricing": {
          "currency": "USD",
          "unit": "per_1000000_tokens",
          "scope": "Paid Gemini Developer API text requests; no grounding, priority or other tools. Output pricing includes thinking tokens.",
          "standard": {
            "input": 0.3,
            "cacheRead": 0.03,
            "cacheWrite": null,
            "output": 2.5,
            "cacheStoragePerHour": 1
          },
          "batch": {
            "input": 0.15,
            "cacheRead": null,
            "cacheWrite": null,
            "output": 1.25,
            "cacheStoragePerHour": 1
          },
          "unknowns": "No separate cache-write rate was established. Batch pricing lists cache reads at 0.02, but the Batch guide says standard cache rates (0.03 here); confirm this discrepancy before assuming either rate. Storage price applies when explicitly retaining cached content."
        },
        "cache": {
          "minimumPrefixTokens": null,
          "note": "Implicit caching is documented for Gemini 2.5 and later, but the reviewed minimum-token table does not list this Flash-Lite model. Do not infer its threshold from Flash. Explicit cache objects require generateContent; Interactions supports implicit caching only."
        },
        "benchmarkStart": "Use the documented model code and a provider-supported low-thinking configuration as a recorded condition; verify exact request parameters before a paid run.",
        "sourceIds": [
          "google-flash-lite",
          "google-pricing",
          "google-cache",
          "google-batch"
        ]
      },
      {
        "id": "openai-astra-reference",
        "provider": "OpenAI",
        "modelId": "gpt-6-astra",
        "documentedSnapshot": "gpt-6-astra",
        "role": "Higher-capability reference candidate, selected from the provider's flagship tier. Its judgments are not ground truth.",
        "availability": "Listed in the current API catalogue. Access in this environment was not tested.",
        "context": {
          "windowTokens": 1050000,
          "maxInputTokens": 922000,
          "maxOutputTokens": 128000
        },
        "pricing": {
          "currency": "USD",
          "unit": "per_1000000_tokens",
          "scope": "First-party text API, Standard or Batch, input at most 272000 tokens; no tools or regional/priority modifiers.",
          "standard": {
            "input": 10,
            "cacheRead": 1,
            "cacheWrite": 12.5,
            "output": 50
          },
          "batch": {
            "input": 5,
            "cacheRead": 0.5,
            "cacheWrite": 6.25,
            "output": 25
          },
          "longContextStandard": {
            "inputAboveTokens": 272000,
            "input": 20,
            "cacheRead": 2,
            "cacheWrite": 25,
            "output": 75
          }
        },
        "cache": {
          "minimumPrefixTokens": 1024,
          "minimumLifetimeMinutes": 30,
          "note": "Use a stable eligible prefix; retain actual write/read counters instead of presuming a warm cache."
        },
        "benchmarkStart": "Start with reasoning.effort low on the same held-out cases as cheaper candidates. Use human review to adjudicate disagreements; escalate production cases only after that policy is qualified.",
        "sourceIds": [
          "openai-models",
          "openai-astra",
          "openai-pricing",
          "openai-cache",
          "openai-reasoning"
        ]
      }
    ],
    "pricingNotes": [
      {
        "id": "billing-formula",
        "text": "Estimate USD from disjoint billed categories: ordinary input × its rate + cache writes × their applicable rate + cache reads × their rate + billed output × its rate, all divided by 1000000; add explicit cache storage and tool fees separately. Normalize each provider's usage fields before applying this formula.",
        "sourceIds": [
          "openai-cache",
          "claude-cache",
          "google-pricing"
        ]
      },
      {
        "id": "measure-actual-tokens",
        "text": "Character counts and one tokenizer cannot establish cross-provider cost. Record the provider's input, cached, written and output usage, including billed reasoning/thinking. Keep visible answer length separate; an incomplete response may consume tokens without a usable finding.",
        "sourceIds": [
          "openai-reasoning",
          "openai-cache",
          "google-pricing"
        ]
      },
      {
        "id": "cache-economics",
        "text": "Cache reuse depends on an unchanged prefix, eligibility, lifetime and scheduling. Put a useful stable rubric before changing source passages and mark its boundary where supported. Measure cold runs and repeated runs separately. Adding filler merely to reach a cache threshold changes both cost and model behavior.",
        "sourceIds": [
          "openai-cache",
          "claude-cache",
          "google-cache"
        ]
      },
      {
        "id": "provider-batch-versus-rule-bundle",
        "text": "Bundling six rules into one synchronous request reduces repeated request content. Provider Batch APIs instead schedule many independent requests asynchronously at discounted token rates. Their roughly 24-hour processing windows suit offline evaluation and queued CI work, not an immediate review interaction; requests can expire.",
        "sourceIds": [
          "openai-batch",
          "claude-batch",
          "google-batch"
        ]
      },
      {
        "id": "excluded-costs",
        "text": "These are public API list prices, not Codex/Claude subscription allowances, a spend authorization or an invoice estimate. Output size, retries, cache misses, storage, tools, regional processing, taxes and negotiated terms can change the total. Unknown values are null rather than zero.",
        "sourceIds": [
          "openai-pricing",
          "claude-pricing",
          "google-pricing"
        ]
      }
    ],
    "benchmarkRecommendations": {
      "status": "Proposed protocol; no model-quality results exist from this task.",
      "selectionDecision": "Use the three efficiency candidates and the flagship reference as an experimental cohort only after canonical Token Economy resolution and explicit host admission. Unknown or unsupported identities require a catalogue/policy update there or no execution. Choose a production route from matched-coverage quality, latency and actual total cost; provider price or positioning does not establish the winner.",
      "variants": [
        {
          "id": "single-rule-calls",
          "shape": "One selected rule per call, sharing the same source and task context. Run six calls for a six-rule profile or twenty calls for full coverage.",
          "hypothesis": "Narrow instructions may improve focus, while repeated source/context and per-call output can increase cost.",
          "fairComparison": "Compare against a bundle containing the identical rule set; deduplicate overlapping findings and count all calls, retries and tokens."
        },
        {
          "id": "bundled-six",
          "shape": "One call with an existing six-rule profile, audience, goal, complete relevant passage and evidence.",
          "hypothesis": "A small bundle may offer a useful cost/quality balance; this is unmeasured.",
          "fairComparison": "State the six covered rules. Do not compare its total recall directly with all twenty rules or report omitted rules as passed."
        },
        {
          "id": "full-twenty",
          "shape": "One call with all twenty rules and the same evidence/context.",
          "hypothesis": "It increases instruction coverage; extra prompt length or competing instructions could affect cost and review quality.",
          "fairComparison": "Use per-rule labels and the same response constraints; record findings suppressed by output/findings limits as a coverage limitation."
        },
        {
          "id": "lexical-prefilter-and-semantic-review",
          "shape": "Run local surface matching first, then supply exact candidate spans with context and a separately scheduled semantic review for the requested scope.",
          "hypothesis": "Local matching can avoid model work for deterministic cue discovery; reduced paid scope must still be validated for missed issues.",
          "fairComparison": "Only eight catalogue rules have surface patterns. Zero lexical matches does not clear those rules semantically or the other twelve. Sample negative cases and retain a semantic pass for uncued issues."
        }
      ],
      "referenceEscalation": "As a separate variant, let a cheaper reviewer propose findings and send only defined uncertain or conflicting cases to the reference model. Count both stages and independently sample non-escalated cases to measure routing misses.",
      "fixtures": "Use Git-owned EN/DE passages with source hashes, audience, goal, supporting facts and accepted decisions. Include problematic cases, valid uses of each pattern, clear keep decisions, long-page context and human-edited AI drafts. Keep a held-out set outside prompt tuning.",
      "judgment": "Human reviewers label rule-specific problems and acceptable keep/change outcomes. Rate false alarms, missed issues, preserved facts/intent and distinct warranted alternatives separately. Resolve or record disagreement; neither a larger model nor author acceptance alone supplies ground truth.",
      "measurements": [
        "Per-rule precision and recall with denominators and coverage",
        "False changes on legitimate counterexamples",
        "Fact and author-intent preservation",
        "Structured-output validity and exact quoted-span alignment",
        "Input/cache-write/cache-read/billed-output usage and total USD",
        "Latency p50/p95, retries, incomplete responses and provider errors"
      ],
      "controls": "Record exact requested and returned model IDs, date, API endpoint, provider settings, rule IDs/catalogue version, prompt hash, cache mode, source hash, output cap and retry policy. Run repeated cases to expose variation. Compare both cold and reused prefixes; keep prompt and model changes separate when attributing effects.",
      "releaseGate": "Set acceptable false-change and missed-issue thresholds before reviewing results. Publish fixture coverage and observed failures with the selected configuration. Preserve the keep-as-written option and host validation in production. No benchmark result should become an AI-authorship score."
    },
    "runtimeIntegration": {
      "status": "research-only",
      "canonicalReference": "docs/model-strategy.md",
      "boundary": "This dated official-documentation snapshot is not a runtime model catalogue, price resolver, allowlist, access grant or qualification record. Resolve candidates against the installed Token Economy catalogue; the host explicitly admits provisional experiments. Unknown or unsupported identities require an update to that catalogue/policy or no execution. API list prices can differ from CLI/subscription routing."
    }
  },
  "organizationResearch": {
    "schemaVersion": 1,
    "checkedAt": "2026-09-12",
    "scope": "Research on prompt organization and model-cost trade-offs; API uses are Voice's assessment, not results established by the papers.",
    "sources": [
      {
        "id": "batch-prompting-2023",
        "title": "Batch Prompting: Efficient Inference with Large Language Model APIs",
        "url": "https://aclanthology.org/2023.emnlp-industry.74.pdf",
        "kind": "peer-reviewed industry paper",
        "read": "Sections 2–4, section 6 and Appendix B/Table 6; checked 2026-09-12.",
        "finding": "Groups independent questions under shared demonstrations and returns indexed answers. Main experiments use code-davinci-002 and ten QA, arithmetic and NLI/NLU datasets. Reusing demonstrations reduces repeated input. Larger batches often reduce accuracy; the best size varies by task.",
        "limits": "No EN/DE editorial review, current-model comparison or rewrite-quality measurement. Output-heavy work and long inputs can weaken the benefit. This tests multiple samples, not multiple rules on one sample.",
        "apiUse": {
          "availableNow": [
            "selectWritingRules",
            "composeWritingReviewPrompt"
          ],
          "currentUse": "Freeze a consistent rubric and reproduce the prompt for each condition.",
          "proposed": "composeWritingReviewBatch needs passage IDs and per-passage results; evaluateReviewPrompts should measure omissions, wrong assignments, quality and total cost across batch sizes.",
          "releaseFunctionIds": [
            "review-batching",
            "prompt-benchmarks"
          ]
        }
      },
      {
        "id": "lost-middle-2024",
        "title": "Lost in the Middle: How Language Models Use Long Contexts",
        "url": "https://aclanthology.org/2024.tacl-1.9.pdf",
        "kind": "peer-reviewed TACL paper",
        "read": "Sections 2–4 and section 5 retrieval case study; checked 2026-09-12.",
        "finding": "Moves relevant information within controlled multi-document QA and key-value retrieval inputs. Models include MPT, LongChat, GPT-3.5 and Claude-1.3. QA performance commonly declines for central evidence. Claude performs almost perfectly on synthetic retrieval; repeating the query helps retrieval more than QA.",
        "limits": "Older models and retrieval tasks do not establish failure of a twenty-rule editorial prompt. Context-window capacity alone does not establish that every relevant constraint is used.",
        "apiUse": {
          "availableNow": [
            "selectWritingRules",
            "composeWritingReviewPrompt"
          ],
          "currentUse": "Make intended coverage explicit and reproduce the selected instructions.",
          "proposed": "evaluateReviewPrompts should move important facts, rules and keep cases between beginning, middle and end while holding content constant. composeWritingReviewBatch should retain page context and measure position sensitivity.",
          "releaseFunctionIds": [
            "prompt-benchmarks",
            "review-batching"
          ]
        }
      },
      {
        "id": "robatch-2026",
        "title": "Towards Cost-effective LLMs Routing with Batch Prompting",
        "url": "https://arxiv.org/html/2605.28268v1",
        "kind": "May 2026 preprint",
        "read": "Sections 2–6, including implementation and ablations; checked 2026-09-12.",
        "finding": "Jointly selects the model and number of independent queries per prompt. Tests Qwen3 4B/14B/32B and Gemma3 4B/12B/27B on six benchmarks with separate training, validation and test partitions. Reports improved cost–accuracy trade-offs in many settings. Batching tolerance varies by model and task; larger batches can produce malformed outputs.",
        "limits": "Not validation of Voice reviews or current proprietary models. Calibration requires evidence. Scheduling time excludes model API latency, and incomplete batches complicate shared-input cost accounting.",
        "apiUse": {
          "availableNow": [
            "selectWritingRules",
            "composeWritingReviewPrompt"
          ],
          "currentUse": "Freeze rules and prompt versions across paired model comparisons.",
          "proposed": "evaluateReviewPrompts should compare model and prompt structure jointly. composeWritingReviewBatch can produce execution candidates; routing remains in Token Economy and the host.",
          "releaseFunctionIds": [
            "prompt-benchmarks",
            "review-batching",
            "usage-budget"
          ]
        }
      }
    ]
  }
}