You might not need a frontier AI model.

🧭 Route setup Use SQL, a spreadsheet formula, or a short script. Ask a model to write the operation only if you need help.

💡 Why The answer is deterministic. A language model adds variability where ordinary code can be exact and repeatable.

🛡️ Keep frontier when The hard part is deciding what the numbers mean, reconciling ambiguous definitions, or designing the analysis—not executing the arithmetic.

✅ Verify it Re-run the operation from the source data; include row counts and invariants.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Start local or efficient hosted with a strict output schema. Escalate only for difficult layouts or domain judgment.

💡 Why The input and required fields bound the job, and schema validation gives you a cheap, objective first check.

🛡️ Keep frontier when Scans are poor, tables span pages, fields require legal/accounting interpretation, or missing values have costly consequences.

✅ Verify it Validate the schema, require source spans or page references, and sample against the original.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use a small local or efficient model against a fixed label set. Let low confidence abstain.

💡 Why Classification is bounded and high-volume—the shape where efficient models and task-specific tuning are strongest.

🛡️ Keep frontier when Labels are subjective, policy-sensitive, adversarial, or trigger irreversible action without human review.

✅ Verify it Measure precision and recall on held-out, human-labeled examples; monitor abstentions and drift.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient hosted model—or a tested local model for privacy and offline work.

💡 Why Official efficient-model guidance explicitly includes high-volume translation with constrained output.

🛡️ Keep frontier when The text is legal, medical, literary, dialect-heavy, safety-critical, or must preserve subtle brand voice.

✅ Verify it Back-translate a sample, preserve named terms with a glossary, and use native-speaker review for consequential copy.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Google — Gemma 4 E4B instruction-tuned model

🧭 Route setup Start with a local speech model or an efficient multimodal model. Keep the audio private when that matters.

💡 Why Transcription is a specialized, bounded transform; current efficient and on-device model guides name it directly.

🛡️ Keep frontier when Audio is noisy, speakers overlap, vocabulary is specialized, or a verbatim record has legal or safety significance.

✅ Verify it Spot-check names, numbers, speaker changes, and low-confidence segments against the audio.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Google — Gemma 4 E4B instruction-tuned model

🧭 Route setup Use a local or efficient model with explicit facts to preserve and a short style constraint.

💡 Why On-device guidance demonstrates rewriting and voice-driven editing; the task does not require frontier discovery.

🛡️ Keep frontier when The copy carries legal commitments, crisis communications, sensitive negotiation, or a distinctive creative voice that needs expert judgment.

✅ Verify it Diff facts, names, numbers, dates, promises, and links—not just grammar.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use local for privacy or efficient hosted for speed. Specify the audience, length, and facts that must survive.

💡 Why Document processing and summarization are named efficient-model tasks; the supplied source bounds the answer.

🛡️ Keep frontier when Synthesis spans conflicting sources, requires current external research, or omission could materially change a decision.

✅ Verify it Require page or section references and compare every decision-relevant claim with the source.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Google — Gemma 4 E4B instruction-tuned model · Anthropic — Choosing the right model

🧭 Route setup Use local or efficient hosted. Tell it never to invent an owner or date; mark missing fields as unknown.

💡 Why This is structured extraction plus light summarization, with a clear abstention rule for information the notes do not contain.

🛡️ Keep frontier when The meeting is politically sensitive, commitments are disputed, or deciding the real priority requires context outside the notes.

✅ Verify it Link each item to a quote or timestamp and leave absent owners/dates explicitly unknown.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas

🧭 Route setup Use an efficient model with retrieval, citations, and an explicit “not in the source” response.

💡 Why Good retrieval narrows the reasoning surface. The model should explain the supplied material, not invent missing policy.

🛡️ Keep frontier when The corpus is huge or contradictory, the question needs cross-document inference, or the answer drives high-stakes action.

✅ Verify it Open every cited passage; test unanswerable questions and conflicting-source cases.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · SamRouter — How SamRouter classifies model demand

🧭 Route setup Use an efficient multimodal model for the first pass, then human-review context and purpose.

💡 Why The visual input is bounded and the output short. The expensive part is often review, not frontier reasoning.

🛡️ Keep frontier when Images are medically, legally, or technically specialized, or subtle relationships are essential to accessibility.

✅ Verify it Review with the page context; confirm the text conveys purpose rather than listing every visible object.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model

🧭 Route setup Use an efficient model when the pattern, files, constraints, and check command are explicit.

💡 Why The repository supplies the pattern and deterministic tools can catch formatting, type, and test failures.

🛡️ Keep frontier when The framework is unfamiliar, APIs are uncertain, architecture is changing, or generated code crosses security boundaries.

✅ Verify it Run formatter, typecheck, tests, and a focused review for copied assumptions.

📚 Evidence Anthropic — Choosing the right model · OpenAI — Using GPT-5.6 · SamRouter — How SamRouter classifies model demand

🧭 Route setup Start balanced; step down to efficient when the edit is mechanical and the verifier is strong.

💡 Why A clear boundary plus executable checks does more for safe downgrading than prompt length or a model leaderboard score.

🛡️ Keep frontier when The task is ambiguous root-cause debugging, a cross-system migration, security-sensitive, destructive, or long-horizon autonomous work.

✅ Verify it Inspect the diff, run the narrow test suite, and escalate on failure rather than asking the cheap model to improvise indefinitely.

📚 Evidence OpenAI — Using GPT-5.6 · Anthropic — Choosing the right model · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance · SamRouter — How SamRouter classifies model demand

🧭 Route setup Use a date parser in a spreadsheet, SQL query, or script with an explicit timezone and invalid-value rule.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a shell, file-renaming utility, or short script. Preview the rename map before applying it.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a standards-aware validator and return the failing row plus reason; use a model only to explain the rule.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a set operation, spreadsheet lookup, or SQL join and preserve counts for all three groups.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a URL parser or tested regular expression, then diff the result. Ask a model for the pattern only if needed.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use SQL, a dataframe, or a spreadsheet join with an explicit cardinality and unmatched-row report.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a pivot table, SQL GROUP BY, or dataframe aggregation so the calculation stays reproducible.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use the charting feature in your spreadsheet or analytics tool; let a model suggest chart types, not invent points.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use schema validation or explicit null checks and report missing fields by row.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a word-aware diff tool for the exact changes; use a model separately if you need a plain-language summary.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an image processor with fixed dimensions, quality limits, and metadata rules.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a deterministic converter such as Pandoc or the project’s existing build pipeline.

💡 Why The operation has a deterministic definition, so ordinary software can be exact, fast, and repeatable.

🛡️ Keep frontier when The real problem is choosing ambiguous rules, resolving conflicting definitions, or interpreting consequential results.

✅ Verify it Run the operation twice, preserve the source, and check counts or invariants before using the output.

📚 Evidence OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local model with a strict schema when the notes are private; let missing fields remain null.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use an efficient vision model with a receipt schema and confidence or source-box output.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use a balanced model to locate and quote clauses, then send interpretation to qualified counsel.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use an efficient model with an allowed-value schema and an explicit unknown state.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use structured page data first, then an efficient model for fields the markup does not expose.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use OCR or an efficient vision model with a tabular schema and an explicit unreadable marker.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use specialized OCR first, with an efficient vision model for layout recovery and uncertainty labels.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use a compact language detector or local model, with unknown for short or mixed-language text.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a small model against the fixed labels and calibrate it on examples from your own domain.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a tested local classifier as a review queue, not an automatic punishment system; let uncertain cases pass to humans.

💡 Why A constrained first pass can reduce review labor, but the decision remains sensitive enough to require abstention and human ownership.

🛡️ Keep frontier when The output directly determines access, employment, safety, compliance, money, or another consequential outcome.

✅ Verify it Use representative labeled cases, audit disparate errors, retain source evidence, and keep a qualified human in the decision.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use embeddings or a local model for candidate clusters, then have a human merge, split, and name them.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model only for a reversible first-pass queue, with evidence and abstention on unclear cases.

💡 Why A constrained first pass can reduce review labor, but the decision remains sensitive enough to require abstention and human ownership.

🛡️ Keep frontier when The output directly determines access, employment, safety, compliance, money, or another consequential outcome.

✅ Verify it Use representative labeled cases, audit disparate errors, retain source evidence, and keep a qualified human in the decision.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Combine deterministic pattern detectors with a local model for contextual candidates, then require security review.

💡 Why A constrained first pass can reduce review labor, but the decision remains sensitive enough to require abstention and human ownership.

🛡️ Keep frontier when The output directly determines access, employment, safety, compliance, money, or another consequential outcome.

✅ Verify it Use representative labeled cases, audit disparate errors, retain source evidence, and keep a qualified human in the decision.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Start balanced with quoted policy evidence, conservative abstention, and human review for removals or sanctions.

💡 Why A constrained first pass can reduce review labor, but the decision remains sensitive enough to require abstention and human ownership.

🛡️ Keep frontier when The output directly determines access, employment, safety, compliance, money, or another consequential outcome.

✅ Verify it Use representative labeled cases, audit disparate errors, retain source evidence, and keep a qualified human in the decision.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use local embeddings to propose duplicate groups, while retaining each original ticket for review.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local or efficient classifier with a fixed label set, confidence threshold, and other queue.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local model against the frozen codebook, allowing multiple labels and an unclear state.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model to extract stated intent, not to infer protected traits or automatically deny service.

💡 Why A constrained first pass can reduce review labor, but the decision remains sensitive enough to require abstention and human ownership.

🛡️ Keep frontier when The output directly determines access, employment, safety, compliance, money, or another consequential outcome.

✅ Verify it Use representative labeled cases, audit disparate errors, retain source evidence, and keep a qualified human in the decision.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use a local model with the supplied taxonomy and let it suggest new themes in a separate review field.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a balanced model for evidence-linked extraction only; keep hiring judgment and protected-trait controls with people.

💡 Why A constrained first pass can reduce review labor, but the decision remains sensitive enough to require abstention and human ownership.

🛡️ Keep frontier when The output directly determines access, employment, safety, compliance, money, or another consequential outcome.

✅ Verify it Use representative labeled cases, audit disparate errors, retain source evidence, and keep a qualified human in the decision.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use a local model with controlled vocabularies and review low-confidence or multi-topic items.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient multi-label classifier and evaluate each aspect separately on real reviews.

💡 Why A fixed label set turns an open-ended language task into a bounded decision that smaller models can be evaluated on.

🛡️ Keep frontier when Labels are subjective, adversarial, policy-sensitive, or trigger a costly action without human review.

✅ Verify it Test precision, recall, and abstentions on held-out human-labeled examples, including rare and ambiguous cases.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local model with a hard length limit and an explicit list of facts that must survive.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use a local model and instruct it to mark ambiguous fragments instead of completing them imaginatively.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use a local grammar tool or model with a minimal-edit instruction and review the diff.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use deterministic lint rules where possible, with a local model for the contextual remainder.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use an efficient model to generate candidates from the article, then have an editor select and revise.

💡 Why The model is producing candidates from supplied facts and constraints; selection and factual review remain human work.

🛡️ Keep frontier when The campaign needs original strategy, sensitive positioning, regulated claims, or a distinctive creative concept rather than variants.

✅ Verify it Reject invented claims, compare every fact with the brief, and have the responsible editor select and revise the final.

📚 Evidence Anthropic — Choosing the right model · OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model for variants, preserving the brief’s facts and banning unsupported scarcity.

💡 Why The model is producing candidates from supplied facts and constraints; selection and factual review remain human work.

🛡️ Keep frontier when The campaign needs original strategy, sensitive positioning, regulated claims, or a distinctive creative concept rather than variants.

✅ Verify it Reject invented claims, compare every fact with the brief, and have the responsible editor select and revise the final.

📚 Evidence Anthropic — Choosing the right model · OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model, require a source section for each answer, and omit unsupported questions.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use a local model and a readability checker, then review domain terms that cannot safely be simplified.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use a local model for the draft, but retain the original policy as authoritative and cross-check every requirement.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use an efficient model for variants, then check dates, venue, links, accessibility text, and platform constraints.

💡 Why The model is producing candidates from supplied facts and constraints; selection and factual review remain human work.

🛡️ Keep frontier when The campaign needs original strategy, sensitive positioning, regulated claims, or a distinctive creative concept rather than variants.

✅ Verify it Reject invented claims, compare every fact with the brief, and have the responsible editor select and revise the final.

📚 Evidence Anthropic — Choosing the right model · OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model with the verified fact sheet and a hard ban on inferred benefits or certifications.

💡 Why The model is producing candidates from supplied facts and constraints; selection and factual review remain human work.

🛡️ Keep frontier when The campaign needs original strategy, sensitive positioning, regulated claims, or a distinctive creative concept rather than variants.

✅ Verify it Reject invented claims, compare every fact with the brief, and have the responsible editor select and revise the final.

📚 Evidence Anthropic — Choosing the right model · OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model grounded in the policy and order record; require approval before sending.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use a local model or deterministic formatter, then render the result and run a link check.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use a local model for the structural rewrite and compare claims sentence by sentence.

💡 Why The supplied text contains the facts; the model is transforming expression rather than discovering new information.

🛡️ Keep frontier when The copy creates legal commitments, handles a crisis or negotiation, or depends on a distinctive expert or creative voice.

✅ Verify it Diff names, numbers, dates, links, claims, and promises against the source before publishing.

📚 Evidence Google — Gemma 4 E4B instruction-tuned model · OpenAI — Using GPT-5.6

🧭 Route setup Use a balanced model with strict source fields, privacy limits, and human approval before sending.

💡 Why The model is producing candidates from supplied facts and constraints; selection and factual review remain human work.

🛡️ Keep frontier when The campaign needs original strategy, sensitive positioning, regulated claims, or a distinctive creative concept rather than variants.

✅ Verify it Reject invented claims, compare every fact with the brief, and have the responsible editor select and revise the final.

📚 Evidence Anthropic — Choosing the right model · OpenAI — Using GPT-5.6 · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local model to sequence the supplied topics, with named outcomes and realistic time boxes.

💡 Why The task is a bounded conversion from supplied operational material into a known structure or checklist.

🛡️ Keep frontier when Priorities are disputed, missing context changes commitments, or the output authorizes an irreversible external action.

✅ Verify it Trace each output item to the source, leave missing information unknown, and require a human approval before acting.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model with sender and message references, leaving disputed points explicitly disputed.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model grounded only in the ticket history, with ticket IDs or dates for each key point.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model after filtering the real commit set, then verify every claimed behavior in the diff or release.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a deterministic redline plus a balanced model for a cited summary; keep legal interpretation with counsel.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a balanced model with a comparison schema and paper-level citations; do not infer beyond the supplied abstracts.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use an efficient model with retrieval and a not-found response; keep the linked documentation visible.

💡 Why Retrieval narrows the answer to supplied evidence, so the model can explain a corpus instead of inventing missing facts.

🛡️ Keep frontier when The corpus is contradictory or very large, the question needs difficult cross-document inference, or the answer drives high-stakes action.

✅ Verify it Open each citation, test unanswerable questions, and require an explicit not-found response when the corpus is silent.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · SamRouter — How SamRouter classifies model demand

🧭 Route setup Use a local model to extract term-definition pairs with section references, omitting terms the handbook does not define.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use a local model grounded in the notes, and include a source heading or page for each answer.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local model and require answer keys with source references; discard ambiguous questions.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a balanced model with document and line references, keeping conflicting decisions separate.

💡 Why The task is a bounded conversion from supplied operational material into a known structure or checklist.

🛡️ Keep frontier when Priorities are disputed, missing context changes commitments, or the output authorizes an irreversible external action.

✅ Verify it Trace each output item to the source, leave missing information unknown, and require a human approval before acting.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a balanced model with retrieval, paired citations, and an uncertain category for merely different wording.

💡 Why Retrieval narrows the answer to supplied evidence, so the model can explain a corpus instead of inventing missing facts.

🛡️ Keep frontier when The corpus is contradictory or very large, the question needs difficult cross-document inference, or the answer drives high-stakes action.

✅ Verify it Open each citation, test unanswerable questions, and require an explicit not-found response when the corpus is silent.

📚 Evidence Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ · SamRouter — How SamRouter classifies model demand

🧭 Route setup Use an efficient model after a reproducible sampling or clustering step, preserving minority and dissenting themes.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Start balanced when the log and system context are supplied; demand diagnostic commands before any fix.

💡 Why The task still benefits from strong reasoning, but a tight scope and executable verifier can make a balanced model the economical first pass.

🛡️ Keep frontier when Root cause is ambiguous across systems, rollback is weak, production data is at risk, or repeated checks keep failing.

✅ Verify it Review the plan and diff, run the narrowest useful checks, and escalate on failed evidence instead of allowing improvisation.

📚 Evidence OpenAI — Using GPT-5.6 · Anthropic — Choosing the right model · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance · SamRouter — How SamRouter classifies model demand

🧭 Route setup Use an efficient model when the complete function and relevant types are provided.

💡 Why An explicit pattern plus executable checks gives a smaller model a bounded target and makes most mistakes visible.

🛡️ Keep frontier when Architecture is changing, APIs are uncertain, security boundaries are involved, or the task requires long-horizon autonomous work.

✅ Verify it Inspect the diff and run the formatter, typecheck, narrow tests, and relevant security checks.

📚 Evidence Anthropic — Choosing the right model · OpenAI — Using GPT-5.6 · SamRouter — How SamRouter classifies model demand

🧭 Route setup Use an efficient model, then test the worked example in the spreadsheet itself.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local model when the locations, opening times, and travel constraints are already supplied and current.

💡 Why The task is a bounded conversion from supplied operational material into a known structure or checklist.

🛡️ Keep frontier when Priorities are disputed, missing context changes commitments, or the output authorizes an irreversible external action.

✅ Verify it Trace each output item to the source, leave missing information unknown, and require a human approval before acting.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Use a local model to mirror the supplied structure and cite the source section for every branch.

💡 Why The provided material bounds the answer, and explicit length, audience, and citation requirements constrain the transform.

🛡️ Keep frontier when The work must reconcile conflicting sources, make a consequential judgment, or incorporate current facts beyond the supplied context.

✅ Verify it Check every decision-relevant claim against the source and require section, page, line, or timestamp references where possible.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model · Hamel Husain and Shreya Shankar — LLM Evals FAQ

🧭 Route setup Extract date-event pairs with source references, then sort them deterministically and flag partial dates.

💡 Why The source and target fields bound the work, while a schema and source references make failures observable.

🛡️ Keep frontier when Inputs are degraded, layouts are unusually complex, or a field requires legal, medical, financial, or domain judgment.

✅ Verify it Validate the schema, retain source locations, and compare a representative sample with the original material.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Simon Willison — Structured data extraction from unstructured content using LLM schemas · Stanford University — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

🧭 Route setup Use an efficient vision model for the first pass and preserve an uncertain state for ambiguous details.

💡 Why The media input and requested metadata bound the output; the expensive step is usually review rather than frontier reasoning.

🛡️ Keep frontier when The media is medically, legally, or technically specialized, or subtle context materially changes the answer.

✅ Verify it Review against the original media, paying special attention to names, numbers, timing, and uncertain segments.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model

🧭 Route setup Use an efficient vision classifier with a fixed label set and human review for multi-subject images.

💡 Why The media input and requested metadata bound the output; the expensive step is usually review rather than frontier reasoning.

🛡️ Keep frontier when The media is medically, legally, or technically specialized, or subtle context materially changes the answer.

✅ Verify it Review against the original media, paying special attention to names, numbers, timing, and uncertain segments.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model

🧭 Route setup Use OCR or an efficient vision model, marking cropped and unreadable regions explicitly.

💡 Why The media input and requested metadata bound the output; the expensive step is usually review rather than frontier reasoning.

🛡️ Keep frontier when The media is medically, legally, or technically specialized, or subtle context materially changes the answer.

✅ Verify it Review against the original media, paying special attention to names, numbers, timing, and uncertain segments.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model

🧭 Route setup Use an efficient vision model for drafts, then verify people, location, and event context with the photo owner.

💡 Why The media input and requested metadata bound the output; the expensive step is usually review rather than frontier reasoning.

🛡️ Keep frontier when The media is medically, legally, or technically specialized, or subtle context materially changes the answer.

✅ Verify it Review against the original media, paying special attention to names, numbers, timing, and uncertain segments.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model

🧭 Route setup Use an efficient vision model for controlled product-type labels, not unsupported brand or model identification.

💡 Why The media input and requested metadata bound the output; the expensive step is usually review rather than frontier reasoning.

🛡️ Keep frontier when The media is medically, legally, or technically specialized, or subtle context materially changes the answer.

✅ Verify it Review against the original media, paying special attention to names, numbers, timing, and uncertain segments.

📚 Evidence Google — Gemini 3.5 Flash-Lite · Anthropic — Choosing the right model

🧭 Route setup Use a specialized local audio classifier and review boundary timestamps.