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, fraud-vs-error, anomaly triage, adversarial workup |
lenses/hypotheses-ach.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 |
| 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 (32 items, historical anchors).
Mode Recommendation
- Define the observation precisely.
- Run null/base-rate before causal story generation.
- 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
- 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.
1---2name: analyze3description: Use when: why/root-cause, confounder, regression, ACH, DAG adjustment, weakest-link. Lenses: null, causal, dag, hypotheses, audit, stop. NOT ideation (/brainstorm).4---56# Analyze78Shared reasoning lens library. This skill recommends lenses; it does not own9project routing thresholds, domain evidence standards, or final workflow10decisions.1112Project workflows decide:1314- whether analysis is required15- which evidence threshold counts16- when a result changes a decision, falsifier, rerun, monitor, or research queue1718## Lens Selection1920| Lens | Use when | Load |21|---|---|---|22| null/base-rate | The observation may be normal, seasonal, sector-wide, or mean-reverting | `lenses/null-base-rate.md` |23| causal-attribution | The user asks why something happened or what caused a change | `lenses/causal-attribution.md` |24| dag-adjustment | Regression, controls, confounders, causal effect, bad controls | `lenses/dag-adjustment.md` |25| hypotheses-ach | Multiple explanations, fraud-vs-error, anomaly triage, adversarial workup | `lenses/hypotheses-ach.md` |26| weakest-link-audit | A causal story sounds plausible but may have one unsupported link | `lenses/weakest-link-audit.md` |27| 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` |28| 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` |29| decision-impact-stop | Analysis may not change any action | `lenses/decision-impact-stop.md` |3031Detailed reference files remain under `references/` for DAG algorithms,32templates, worked examples, and ACH formats.3334Adjunct: 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` (32 items, historical anchors).3536## Mode Recommendation37381. Define the observation precisely.392. Run null/base-rate before causal story generation.403. If residual remains:41 - shape-matched "why" question -> causal attribution42 - regression/control question -> DAG adjustment43 - 3+ live explanations -> hypotheses/ACH44 - plausible chain with uncertain mechanism -> weakest-link audit454. If the result would not alter a decision, falsifier, monitor, data rerun, or46 research queue item, label it `NON-ACTIONABLE` and stop.4748## Output Contract4950Every analysis should end with:5152- leading explanation or "null explains it"53- top alternative54- falsifier or discriminating evidence55- decision impact56- next action, or `NON-ACTIONABLE`5758Do not use this shared skill to encode project-private alpha, donor/sample59context, private file paths, or domain-specific thresholds. Keep those adapters60inside the project workflow that invoked the lens.
markusstrasser