analyze

# Analyze Shared reasoning lens library. This skill recommends lenses; it does not own project routing thresholds, domain evidence standards, or final workflow decisions. Project workflows decide: - whether analysis is required - which evidence threshold counts - when a result changes a decision, falsifier, rerun, monitor, or research queue ## Lens Selection | Lens | Use when | Load | |---|---|---| | null/base-rate | The observation may be normal, seasonal, sector-wide, or mean-reverting | `lenses/null-base-rate.md` | | causal-attribution | The user asks why something happened or what caused a change | `lenses/causal-attribution.md` | | dag-adjustment | Regression, controls, confounders, causal effect, bad controls | `lenses/dag-adjustment.md` | | hypotheses-ach | Multiple explanations, anomaly triage, adversarial workup | `lenses/hypotheses-ach.md` | | fraud-error | A claim, dataset, poll, paper or audit may be fabricated or misrepresented, or a claim alleges fraud; supplies the slots and tests for ACH (record's noun and status rung, vetting funnel, raw-record and too-clean tests, outcome-level check) | `lenses/fraud-error.md` | | weakest-link-audit | A causal story sounds plausible but may have one unsupported link | `lenses/weakest-link-audit.md` | | spirit-audit | A work-product is judged against a contract (eval verdict, backtest, gate, benchmark, claim) and may be letter-true but spirit-false (gaming, invalid gold, errors-scored-as-results, confounds, leakage) | `lenses/spirit-audit.md` | | proxy-validity | Estimating a latent quantity (ability, credibility, capability) from observable proxies, or applying a proposed bias-discount to such an estimate | `lenses/proxy-validity.md` | | measurement-pipeline | The object is a published statistic built from administrative or survey records (crime, deaths, cases, unemployment, incidents, migration, inequality) and may not be the quantity it names; run before any cross-group, cross-place or cross-year comparison | `lenses/measurement-pipeline.md` | | decision-impact-stop | Analysis may not change any action | `lenses/decision-impact-stop.md` | Detailed reference files remain under `references/` for DAG algorithms, templates, worked examples, and ACH formats. Adjunct: when the object under analysis is a published quantitative claim (a headline number, %, ranking, or model output), also run the construction checks — ledger, unit, base, window, gross/net, measurement-vs-model — from the research skill's `references/quant-bias-checklist.md` (items 1–33). When an institution produced the number (agency, court filing, journal, commissioned report) or the claim is about consensus or trust, also run its section I (items 34–52). ## Mode Recommendation 1. Define the observation precisely, as bare facts separate from any explanation that arrived with it. If it is a published administrative or survey statistic, run measurement-pipeline first: the number may not be the quantity it names. 2. Run null/base-rate before causal story generation. 3. If residual remains: - shape-matched "why" question -> causal attribution - regression/control question -> DAG adjustment - 3+ live explanations -> hypotheses/ACH - plausible chain with uncertain mechanism -> weakest-link audit 4. If the result would not alter a decision, falsifier, monitor, data rerun, or research queue item, label it `NON-ACTIONABLE` and stop. ## Output Contract Every analysis should end with: - leading explanation or "null explains it" - top alternative - falsifier or discriminating evidence - decision impact - next action, or `NON-ACTIONABLE` Do not use this shared skill to encode project-private alpha, donor/sample context, private file paths, or domain-specific thresholds. Keep those adapters inside the project workflow that invoked the lens.