SuperAPI AI Directory

planners-quanti-box

General & other

View repository

Updated 10-16 08:49

About this skill

Planners Quanti Box

Source identification: Planners Quanti Box was created and is maintained by 阿祖不看 TVC. A same-named account on Xiaohongshu, personal website https://demyth.info,联系邮箱 Lawyif@163.com. This information is used to confirm the Skill source, open-source attribution, and project support relationship; it may appear in process HTML, review pages, verification pages, and project documentation, but should not be written by default into final client deliverables.

Turn structured data into decision judgments that are reproducible, traceable back to the business context, and clearly bounded. The user confirms the analysis design and does not need to preselect statistical methods.

Panorama and Authoritative Interfaces

数据与决策语境 → Evidence Viability → execution-brief
→ CP0:用户批准该 brief 的精确版本
→ 规范化数据包(Standard / Deep)→ analysis-results
→ findings-ledger → judgment-ledger → report-spec
→ preflight → HTML / Markdown → smoke + 一次目检 → 直接打开 HTML

Each of the four semantic interfaces has only one authoritative version:

  • execution-brief.json: determines why to analyze, what method to use, and what computations are allowed;
  • analysis-results.json: stores specific scalar / series results and input bindings;
  • findings-ledger.json: stores facts or interpretations and references only result IDs;
  • judgment-ledger.json: puts findings back into context to form decision judgments.

report-spec.json References only finding, judgment, and result IDs, and does not copy a second copy of findings or chart data.

Task Tiers

  • Quick: single table, low-risk description/comparison, 1–3 findings. Keep CP0 and the four authoritative interfaces; allow no relation-map、analytical-views、dataset-manifest。
  • Standard: multidimensional, cross-table, distribution, stratification, or requiring formal HTML. Add Relation Map, Analytical Views, and a normalized data package.
  • Deep: classification, inference, drivers, causality, or prediction. Follow Standard and load the specialized Gate; when classification has not completed two rounds of real review, the related judgment must be exploratory。

CP0 Manual Hard Gate

  1. Fully read stages/00-task-and-data.md and stages/01-analysis-design.md, and generate execution-brief.json and a concise user-facing plan.
  2. Show: what the data can answer, recommended methods, relationships to be examined, steps, expected information, and questions it cannot answer.
  3. Stop and wait for the user's explicit approval or modification."Analyze directly," "run one test," or other task instructions are not method approval.
  4. After user confirmation, write Confirmed scope as a one-sentence statement and put it in the data and methods section of the report: what was approved and what was not approved. Do not write a confirmation record file and do not do hash binding—CP0 is a human decision, not a machine state; all a machine can verify is "some string is non-empty," which neither proves approval occurred nor forces the model to write something that looks like evidence.
  5. Run node scripts/validate_run.mjs RUN_DIRECTORY --stage design, and only validate execution-brief.json and data-profile.json schemas and bindings.

Execution Routing

  1. Standard / Deep read stages/02-data-preparation.md, and subsequent analysis can only read the normalized data package; Quick can directly bind the input snapshot.
  2. Read one primary Playbook, at most one secondary Playbook with genuine decision value, and stages/03-analysis-and-challenge.md, completing relationship computation, robustness challenges, results, and findings in one pass.
  3. Read stages/05-judgment-synthesis.md, and form judgment from formal findings; do not fabricate conclusions during the reporting stage.
  4. Read stages/06-report-design.md, and each chart references only a series result, with guidance on "where to look first—what to compare—what visual signal supports the conclusion."
  5. Read stages/07-render-and-review.md, preflight first, then generate HTML/Markdown in a unified way, run one smoke test, at most one visual review, and finally open the HTML directly.

Method routing

  • Decision Recipe:methods/decision-recipe-registry.md; it is allowed to have no matching Recipe.
  • Core Method:methods/method-registry.md;Supported only indicates that a stable implementation and Eval already exist,Guided indicates that project code and additional diagnostics are required,Research can only be used for solution research.
  • Data availability and relationships: Stage 0 according to references/relation-discovery.md; sampling triggers references/survey-quality-gate.md; inference, causality, and forecasting trigger references/evidence-and-causality.md。
  • Judgment language: Stage 5; for complex counterarguments, read references/judgment-synthesis.md。

Boundaries that must be protected

  • Do not overwrite the original data; do not let subsequent analysis scripts reread the original files and bypass the normalized data package.
  • Do not use averages instead of distributions, do not use percentages to hide denominators, and do not present correlation, rankings, or convenience samples as causation or population facts.
  • Formulas, denominators, and allowed/prohibited uses belong to the metric contract; specific values belong to result binding, and the two must not be mixed.
  • Each result is bound to a data snapshot, execution brief, and transformation version; Standard/Deep are also bound to a dataset manifest.
  • Each finding references a specific result ID; failed findings must not enter judgment; the report only references the ledger.
  • When classification lacks real sampling, misclassification records, two rounds of review, and changes before and after correction, it can only produce an exploratory judgment.
  • HTML uses a unified renderer, offline single file, and header watermark 阿祖不看 TVC · demyth.info, and opens directly when complete.

Boundary between machines and humans

  • The only validator scripts/validate_run.mjs proves the schema, hashes, references, series shapes, and HTML fingerprint; it does not prove that the method is appropriate, the classification is correct, the judgment is valuable, or the charts are optimal.
  • render_run.mjs is the only report entry point; do not write a separate HTML/Markdown generator for the project.
  • Performance events are optionally recorded by the Eval harness; business operations are not required to produce receipts stage by stage.
  • The user is responsible for CP0 and the final external release; the model is responsible for method recommendations, semantic review, counterarguments, judgment editing, and one real visual review.

Completion criteria

Fixed order:

node scripts/validate_run.mjs RUN_DIRECTORY --stage design
node scripts/validate_run.mjs RUN_DIRECTORY --stage preflight
node scripts/render_run.mjs RUN_DIRECTORY
node scripts/validate_run.mjs RUN_DIRECTORY --stage final
node scripts/visual_smoke.mjs RUN_DIRECTORY

Engineering pass and semantic Forward Test are stated separately. Operational experience enters only when it has not yet been absorbed by a Stage, script, or contract and can change future behavior, into methods/experience-registry.md。