Evaluate table extraction without losing row relationships
Check cell text, row and column association, merged headers, units, and downstream totals rather than accepting a table that merely looks plausible.
Research briefing · Sources includedResearch & field notes
Research library — page 3
Read a guide, follow its sources, and check the limits before applying a finding. Search the full library from the research desk, or browse by topic.
Evaluate numerical reasoning with units and tolerances
Validate source quantities, operations, units, rounding, and acceptable error separately with worked numeric controls and explicit nonfinite handling.
Research briefing · Sources includedEvaluate generated code with tests and execution boundaries
Compare generated code using held-out functional cases, boundary inputs, dependency constraints, and isolated execution while retaining failed and incomplete attempts.
Research briefing · Sources includedEvaluate AI code review for useful findings and false alarms
Measure whether AI review finds reproducible defects in a diff, locates them accurately, and avoids unsupported or duplicate findings on clean changes.
Research briefing · Sources includedEvaluate generated search queries by the evidence they retrieve
Compare generated queries on a fixed index using required evidence coverage, precision at a stated depth, query cost, and preservation of user constraints.
Research briefing · Sources includedEvaluate answer completeness against a required-facts checklist
Build required-point checklists from the user task, distinguish optional detail from essential coverage, and pair completeness checks with factual support.
Research briefing · Sources includedMeasure answer relevance without rewarding unsupported detail
Judge whether an answer addresses the user’s actual request, preserves constraints, and avoids distracting additions while checking support separately.
Research briefing · Sources includedEvaluate refusals for safety, precision, and useful alternatives
Test prohibited and permitted requests together, verify whether restricted actions actually stop, and distinguish appropriate refusal from unnecessary blocking.
Research briefing · Sources includedEvaluate tone and style without confusing preference with correctness
Turn brand and audience preferences into reviewable tone anchors, use controlled alternatives, and keep politeness from masking unsupported or unusable answers.
Research briefing · Sources includedEvaluate AI-generated content for accessibility tasks
Check AI-written alternatives, headings, links, and instructions in their real interface context with task-specific evidence and assistive-technology review.
Research briefing · Sources includedTest long-context admission, truncation, and separated evidence
Verify what context actually reaches an LLM, detect truncation at input boundaries, and test separated premises under a declared task budget.
Research briefing · Sources includedPlan evaluation power around a practical decision
Specify a useful effect, sampling unit, test, dependence assumptions, and decision threshold before budgeting an LLM comparison sample.
Research briefing · Sources includedUse paired bootstrap comparisons for LLM experiments
Resample task identities together for baseline and candidate, define the statistic and interval method, and preserve cluster structure and missing outcomes.
Research briefing · Sources includedUse Elo-style LLM rankings with explicit limits
Interpret preference rankings as relative evidence from a defined comparison pool, preserving task mix, uncertainty, ties, and application-specific checks.
Research briefing · Sources includedCompare LLM latency with percentiles and incomplete runs
Measure time to useful completion, first output, retries, and timeouts with a fixed workload and explicit percentile calculation instead of averages alone.
Research briefing · Sources includedPrepare an evaluation review packet a release owner can inspect
Assemble scope, version manifests, task outcomes, uncertainty, critical failures, and a decision record that a release owner can trace back to cases.
Research briefing · Sources includedChoose what evaluation evidence to retain and for how long
Map evaluation artifacts to investigation needs, minimize sensitive copies, and test deletion across raw outputs, labels, exports, caches, and derived records.
Research briefing · Sources includedChoose RAG chunk sizes with an evidence-focused trial
Compare chunk sizes using evidence coverage, boundary failures, and answer quality while holding retrieval budgets and corpus versions fixed.
Research briefing · Sources includedTest whether chunk overlap helps retrieval or adds noise
Test overlap with boundary questions, duplicate-evidence measures, and a fixed context budget before increasing index size.
Research briefing · Sources includedEvaluate document parsing before tuning a RAG retriever
Audit document parsing with source anchors, reading order, missing symbols, and table relationships before tuning retrieval.
Research briefing · Sources included