Insights
Privacy and preparation
379 guides · page 4 of 8
Energy safety exclusion dataset: separating operational data from safety instructions
An energy safety exclusion dataset keeps lockout, switching and infrastructure details out of AI training. Use this reviewed scope framework for utility data.
English business dataset translation boundaries: keep source context
English business dataset translation boundaries: how to find non-English decision context in English records, keep originals, tag segments and audit the result.
English dataset currency normalization: why keep the original amount?
English dataset currency normalization done right: keep original amount and currency, disclose the rate, date and rounding, and validate every conversion.
English dataset mixed language exclusions: the workflow steps you lose
English dataset mixed language exclusions can delete whole workflow stages. Audit stage coverage and thread breakage before you cut non-English records.
English-led call dataset sampling: measure English by the task's unit
English-led call dataset sampling: how to measure English coverage by call, speech time, turn or case for the AI task, and why case-based samples hold up.
Enterprise archive AI contamination study: a measurement protocol
An enterprise archive AI contamination study protocol: date records, calibrate detectors on pre-AI text, correct for error and report intervals. No results.
Enterprise benchmark task family IDs: assign them before you split
Enterprise benchmark task family IDs, derived from procedure, template and system fields, with a schema, a passing and a failing record, and validator checks.
Event log actor role mapping: what an anonymous actor ID still needs
Event log actor role mapping: the role, authority and actor-kind fields a pseudonymized actor ID needs, with a schema, worked case and validator checks.
Event log reconstructed task validation: replay it with an operator
Event log reconstructed task validation: replay sampled cases with an operator, classify each discrepancy, fix the rules and retest, with a report template.
How to run an expert annotation fatigue audit: a study protocol
An expert annotation fatigue audit protocol: how to detect reviewer fatigue in dataset labels with seeded gold items, timing logs and a pre-registered plan.
Expert confidence labels for business tasks: know, infer or can't tell
Expert confidence labels for business tasks: a seven-code guide that separates stated facts, rules, inferences and unknowns, with evidence fields and audits.
Expert decision counterfactuals: how to capture what would have changed the call
Expert decision counterfactuals, done properly: vary one real factor at a time, have the expert decide blind, and check each stated boundary against past cases.
Expert decision rationale capture: how to record why the exception was made
Expert decision rationale capture, step by step: a four-pass interview, evidence links for every reason, uncertainty and review fields, and a worked example.
Expert disagreement in a dataset: how a business keeps the valid kind
How a business should document expert disagreement in a dataset: classify each case, fix slips and gaps, and keep legitimate answers with their conditions.
Expert judgment dataset pilot: how to capture verifiable expert decisions
An expert judgment dataset pilot in weeks: capture decisions with evidence, rationale and uncertainty, then measure whether the records are worth scaling.
Expert knowledge update dataset: recording a changed expert view
Expert knowledge update dataset schema: record when an expert changes their mind, the evidence that triggered it, and validator checks that catch hindsight.
Expert policy exception boundaries: recording where a legitimate exception ends
Expert policy exception boundaries: record the conditions that made an exception legitimate, pair it with an invalid counterexample and prove the outcome.
Expert review calibration workshop: calibrate reviewers before labeling
How to run an expert review calibration workshop before labeling thousands of cases: blind rounds, agreement gates and a disagreement log template.
Expert reviewer conflicts of interest: manage them with blinded review
Expert reviewer conflicts of interest in annotation: record incentives, choose recusal, blinding or double review, and test whether a conflict moved the labels.
Expert reviewer qualification evidence: show why this reviewer fits this task
Expert reviewer qualification evidence for business records: match role experience to the task, cite system records and disclose conflicts, with a profile.
Expert tacit knowledge as evidence: how to capture it without inventing it
Expert tacit knowledge becomes evidence when each claim is tied to a real case and its records, labeled by how it is known and checked by a second reviewer.
Expert task reconstruction interviews: replay the case before the task
Run expert task reconstruction interviews: replay a recorded case with the outcome hidden, log evidence gaps, then write a training task the records support.
Food safety data scope boundaries: what to exclude from an AI dataset
Food safety data scope boundaries for AI datasets: a screen for what food operations should exclude, transform or keep, plus a traceability replay test.
Free text redaction false negatives: testing detectors on signature blocks and footers
Free text redaction false negatives hide in signature blocks, footers and quoted replies. Use a stratified per-zone audit to find and count what detectors miss.
Healthcare admin correction chronology: tracing one fix downstream
A healthcare admin correction chronology tracks one corrected account detail through eligibility, claims, remittances and statements. Task spec and example.
Healthcare admin document missingness: testing absent-form tasks
Healthcare admin document missingness: how to record absent forms as dated states tied to payer rules and test an agent's next administrative action.
Healthcare admin human escalation labels: which cases to hand off now
Healthcare admin human escalation labels: which administrative cases an AI should hand off at once, a tiered rubric, label fields and clinical detail to omit.
Healthcare admin multi-system joins: one case across scheduling, billing and support
Healthcare admin multi-system joins: link scheduling, billing, remittance and support into one case with random tokens, a reviewed schema and validator checks.
Healthcare admin retention versus reuse: keeping is not permission
Healthcare admin retention versus reuse: why a duty to keep records never settles AI reuse, with an issue map for claims, remits, calls and authorizations.
Minimum necessary healthcare administrative data for an AI task
Healthcare administrative data and minimum necessary: write the AI task first, then map each field to keep, transform or exclude with privacy review.
What should a buyer brief for healthcare administrative data specify?
A healthcare administrative dataset buyer brief template: the task, nonclinical boundary, field necessity, exclusions and permitted uses to require first.
Healthcare administrative dataset contamination: keep clinical detail out
Healthcare administrative dataset contamination: where clinical detail hides in admin records, how to audit the assembled package and how to log each exclusion.
Healthcare administrative dataset sample review: what to inspect
A healthcare administrative dataset sample review protocol: what privacy reviewers inspect, what evidence to record and which findings stop an extract.
Healthcare no-show and reschedule datasets: defining each outcome
Build a healthcare no-show and reschedule dataset with labels that hold: nine outcome definitions, a reconciliation procedure and a minimum-necessary field list
Healthcare scheduling exception dataset without clinical notes
Build a healthcare scheduling exception dataset that tests agents without clinical notes: a privacy reviewed task inventory, field rules and date handling.
Healthcare vendor ticket dataset: what privacy review comes before licensing?
Building a healthcare vendor ticket dataset? Check business associate terms, find PHI in screenshots and logs, and apply a PHI exclusion checklist first.
Insurance document redaction utility: keeping the coverage question
Insurance document redaction utility depends on what you keep: shifted dates, role tokens, schedule links. A tradeoff plan, utility test and worked case.
Insurance loss run data quality: the context a training set needs
Insurance loss run data quality depends on context: valuation dates, policy links, reserves and gaps. Use this audit worksheet before loss runs train a model.
Insurance underwriting exception dataset: why exceptions beat approvals
What makes an insurance underwriting exception dataset useful: an annotation guide for authority, guideline version, evidence conflicts and rationale.
International data supplier fact sheet: what should it contain?
An international data supplier fact sheet: record origin country, language tags and share, locale, rights and custody per source, with a template.
International dataset entity names: keeping links across languages
International dataset entity names: map renames, translations and transliterations to one entity key, with a variant table, link tests and a worked example.
International dataset numeric format errors: checking decimals and units
Fix international dataset numeric format errors before they ship: a column-level checklist for decimal commas, digit grouping, units and currency codes.
International dataset pilot coverage: test the local cases the main sample misses
International dataset pilot coverage: stratify by country and workflow, set per-cell floors, force local variants in and keep original formats and locale tags.
IT incident timeline dataset: what to keep beyond the closed ticket
An IT incident timeline dataset schema that links alerts, responder decisions, approvals, runbooks and measured recovery, with redaction rules and examples.
Which IT records to inventory before a data licensing fit check
IT operational data sample readiness: inventory tickets, incidents, alerts, changes and runbooks by source, rights, secrets and linkage before a fit check.
IT runbook exception data: what the engineer did after a step failed
IT runbook exception data: annotate the failed step, evidence checked, recovery move, approval and outcome, with secrets replaced by typed placeholders.
IT service desk escalation dataset: link each escalation to its fix
Build an IT service desk escalation dataset that links each hand-off's cause to the final fix and runbook version, showing where first-line procedures failed.
How to record a privilege exclusion without reproducing privileged content
Legal data privilege exclusion evidence: log who decided, under which screen rule, how many items and which topic, without the log repeating what was excluded.