01 data intake

Empirical Data Intake — Routing-First Cleaning Skill

This skill is the upstream triage layer for empirical research. It sits between the user's raw data file and the four flagship analysis pipelines (00, 00.1, 00.2, 00.3). Its job is to convert "a file someone handed me" into "an analysis-ready dataset, a verified contract, and a routing decision".

Position in the repository

raw data file ──▶ 00.4 (this skill) ──▶ cleaned dataset + contract + routing ──▶ 00 / 00.1 / 00.2 / 00.3

This skill complements the flagships; it does not replace any of them. The flagship Step 1 references (00.1/references/01-data-cleaning.md etc.) remain the canonical execution manuals for advanced cleaning. This skill is the decision layer: what to do, when to ask, which flagship to send to.

When to use

Use this skill when any of these are true:

• User just received a raw data file and asks "where do I start?"
• User does not know which of 00 / 00.1 / 00.2 / 00.3 to use
• User is doing a public-health / epidemiology study (Mode A) — the four flagships' cleaning references do not cover epi-specific cleaning
• User says the data needs cleaning but cannot articulate what kind of cleaning
• Data has obvious issues (missing values, duplicates, ambiguous types) that the user has not explicitly diagnosed

Do NOT use this skill when:

• User has already produced an analysis-ready dataset and wants to run regressions → go straight to flagship
• User asks for a specific cleaning operation in isolation ("winsorize at 1/99%", "run MICE on missing wage") → flagship reference handles it
• User is asking for purely descriptive statistics on a dataset they already trust → flagship Step 3 handles it

---

The 5-slot conditional Q&A

The Q&A is a decision graph, not a fixed questionnaire. Each slot has three modes:

• AUTO — the data already answers it; do not ask
• CONFIRM — the data strongly suggests an answer; ask user to confirm in one yes/no
• ASK — the data cannot answer it; surface as multiple choice

Before any question, always inspect the data first (Section "Static inspection" below). The inspection result determines which slots are AUTO / CONFIRM / ASK.

Slot 1 — Discipline (always ASK)

你这份研究是 (A) 经济学 / 计量 / 金融 / 政治学等社科默认轨, 还是 (B) 公共卫生 / 流行病学 / 临床研究?

This determines:

• Whether to invoke Mode A epi-specific cleaning patterns (references/01-mode-a-epi-patterns.md)
• Whether the variable name is "treatment" (econ) or "exposure" (epi)
• Routing target: flagship default mode vs Mode A

Cannot be inferred from data alone. Always ask.

Slot 2 — Research design (CONFIRM if data suggests, ASK otherwise)

Inspection signals:

Data signal	Likely design	Action
Single (id, time) candidate key with high uniqueness, time has 5+ distinct values	Panel	CONFIRM ("This looks like panel data with units {id} over {time}. Correct?")
No date / time column at all	Cross-section	CONFIRM ("No time variable detected — this is a cross-section. Correct?")
Single time series (one entity, many time points)	Time series	CONFIRM
Discipline=epi + has columns matching `index_date	enrol	baseline
Time variable but no consistent (id, time) key	Repeated cross-section	CONFIRM
None of the above patterns	Unknown	ASK with all 5 options

If user says "I'm not sure" on confirmation, fall through to ASK with all options.

When discipline=epi, replace "panel" wording with "longitudinal cohort" in the confirm message.

Slot 3 — Unit of observation (CONFIRM)

After Slot 2 is set, inspect the strongest candidate primary key. Ask:

看起来每一行代表 一个 {推断单位} 的一个 {时间点} (主键候选: {id_col}, {time_col}). 对吗?

Where 推断单位 is one of: 个人, 家庭, 公司, 国家, 患者, 患者-访视, 学校, observation. The inference is by column name heuristic (e.g., worker_id → 个人; firm_id → 公司; patient_id → 患者) and uniqueness pattern. If no id_col is detectable, ASK directly with multiple choice.

Slot 4 — Focal variables (CONDITIONAL on Slot 2 + 1)

This slot's content depends entirely on the prior answers:

Slot 2 answer	Slot 4 content
Cross-section + causal	Outcome + treatment + key controls
Panel + causal (econ)	Outcome + treatment + treatment start time + (id, time) key
Cohort + index date (epi)	Outcome (event) + exposure (not "treatment") + index date column + censoring column
Time series prediction	Target variable + frequency (D/W/M/Q/A) — no treatment
Repeated cross-section	Outcome + group identifier + time
Descriptive / inequality	Focal variable + group / stratification dimension — no treatment
Pure exploratory	SKIP entire slot — go to "exploratory mode"

Within each branch, use column-name pattern matching to pre-fill candidates:

User intent	Auto-suggest candidates from columns matching
Outcome (econ)	wage, income, earnings, output, employment, y
Outcome (epi)	event, mortality, outcome, death, incidence, complication
Treatment (econ)	binary 0/1 column with name in treat, policy, program, intervention
Exposure (epi)	binary or continuous in exposure, dose, treatment_initiated, drug_*
Index date (epi)	index_date, enrol*_date, baseline_date, t0
Censor (epi)	censor, lost_followup, last_seen

Format the question as: "Outcome 是 wage 吗? Treatment 是 training 吗? (y / 选其他列 / 我也不确定)". Combine into one question, do not ask serially.

If user says "我也不确定", offer the top 3 candidates by name pattern + dtype, plus "skip — flag for later".

Slot 5 — Software target (CONFIRM if file extension reveals, ASK otherwise)

File extension	Default target
.dta	Stata (route to 00.2)
.rds, .RData	R (route to 00.3)
.parquet, .feather	Python (route to 00.1 or 00 StatsPAI)
.csv, .xlsx, .sav, .sas7bdat	ASK

When ASK, present 4 options: Python (general), Python (StatsPAI), Stata, R.

When CONFIRM (e.g. .dta), the message is: "文件是 .dta, 默认走 Stata (00.2). 确认吗?".

---

Prerequisites (Python environment)

Required (hard fail if missing):

• pandas >= 2.0
• numpy
• pyyaml
• openpyxl (for .xlsx read and write — the always-on inspection workbook needs it)

Conditional:

• pyreadstat — only if source is .sav (SPSS) or .sas7bdat (SAS). Pandas's built-in read_stata handles .dta so pyreadstat is optional even there.
• scipy — only used for the MCAR Welch t-test on focal-variable missingness. If absent, fall back to a numpy manual implementation with the |t|>1.96 large-n approximation (see § "MCAR hint fallback" below).
• pyarrow — only if Slot 5 = Python (writes .parquet native). Skip silently if absent and the chosen target is not parquet.

If a hard-required package is missing, intake stops with a clear install hint. If a conditional package is missing AND its file type is requested, stop with a hint. If a conditional package is missing AND not needed for this run, proceed silently.

---

Static inspection (always run before any question)

Before Slot 1, always perform this inspection and store results for use in slot decisions. Choose the inspection language based on the file extension (does not require user choice yet):

# Run this regardless of user's eventual software target
from pathlib import Path
import pandas as pd

def inspect_file(file_path: str) -> dict:
    path = Path(file_path)
    ext  = path.suffix.lower()

    # --- Load ---
    if ext == ".dta":
        df = pd.read_stata(path, convert_categoricals=False)
    elif ext == ".sav":
        import pyreadstat
        df, _meta = pyreadstat.read_sav(path)
    elif ext == ".sas7bdat":
        import pyreadstat
        df, _meta = pyreadstat.read_sas7bdat(path)
    elif ext == ".parquet":
        df = pd.read_parquet(path)
    elif ext in {".csv", ".tsv"}:
        sep = "\t" if ext == ".tsv" else ","
        df = pd.read_csv(path, sep=sep, low_memory=False)
    elif ext == ".xlsx":
        # Multi-sheet handling — never silently take Sheet1 if there are alternatives
        xl = pd.ExcelFile(path)
        sheet = xl.sheet_names[0]
        if len(xl.sheet_names) > 1:
            print(f"[intake] WARNING: {len(xl.sheet_names)} sheets found ({xl.sheet_names}); "
                  f"defaulting to '{sheet}'. To override, pass sheet_name explicitly.")
        df = pd.read_excel(path, sheet_name=sheet)
    else:
        raise ValueError(f"Unsupported file: {ext}")

    # --- Helpers ---
    def is_string_like(s):
        # pandas 2.x: string columns may be 'object' or 'string'/'string[pyarrow]'
        return pd.api.types.is_string_dtype(s) or s.dtype == object

    def looks_like_id_name(c):
        cl = c.lower()
        return cl == "id" or cl.endswith("_id") or cl in {"uid", "uuid", "key"}

    def looks_like_time_name(c):
        cl = c.lower()
        return cl in {"year","date","time","ym","quarter","month","wave","period","t"}

    # --- Single-column primary-key candidates (filter out numeric measures) ---
    # An "ID-shaped" single column must look like an ID by name OR be integer/string,
    # not a continuous numeric measure that happens to be unique by coincidence.
    single_pkey_candidates = []
    for c in df.columns:
        if not (df[c].is_unique and df[c].notna().all()):
            continue
        is_id_name  = looks_like_id_name(c)
        is_int_like = pd.api.types.is_integer_dtype(df[c])
        is_str_like = is_string_like(df[c])
        if is_id_name or is_int_like or is_str_like:
            single_pkey_candidates.append(c)

    # --- Composite (id, time) primary-key candidate ---
    id_cols   = [c for c in df.columns if looks_like_id_name(c)]
    time_cols = [c for c in df.columns
                  if looks_like_time_name(c) or pd.api.types.is_datetime64_any_dtype(df[c])]
    composite_pkey = None
    for ic in id_cols:
        for tc in time_cols:
            if not df.duplicated(subset=[ic, tc]).any() and df[ic].notna().all() and df[tc].notna().all():
                composite_pkey = (ic, tc)
                break
        if composite_pkey:
            break

    # --- Binary 0/1 columns (treatment-candidate detection) ---
    binary_01_cols = [c for c in df.columns
                       if df[c].dropna().nunique() == 2
                       and set(df[c].dropna().unique()) <= {0, 1, True, False, 0.0, 1.0}]
    # Two-valued text columns (e.g. 东南侧/西北侧) — flag separately
    binary_text_cols = [c for c in df.columns
                         if is_string_like(df[c])
                         and df[c].dropna().nunique() == 2]

    return {
        "file_path":              str(path),
        "ext":                    ext,
        "n_rows":                 len(df),
        "n_cols":                 df.shape[1],
        "columns":                list(df.columns),
        "dtypes":                 df.dtypes.astype(str).to_dict(),
        "missing_rate":           df.isna().mean().to_dict(),
        "n_unique":               df.nunique().to_dict(),
        "single_pkey_candidates": single_pkey_candidates,    # ID-shaped unique cols only
        "composite_pkey":         composite_pkey,            # (id_col, time_col) or None
        "candidate_id_cols":      id_cols,
        "candidate_time_cols":    time_cols,
        "binary_01_cols":         binary_01_cols,
        "binary_text_cols":       binary_text_cols,
        "epi_signal_cols": [c for c in df.columns
                              if any(k in c.lower() for k in
                                     ["index_date","baseline","enrol","followup","censor","event_date","t0"])],
        "string_cols":  [c for c in df.columns if is_string_like(df[c])],
        "_df":          df,    # intentional: downstream cleaning needs the actual DataFrame
    }

Bug-fix log for this code (vs v0.1):

• Composite primary-key detection: v0.1 only found single-column unique cols, which on panel data wrongly flagged continuous numeric measures (e.g., Y, X) as "primary keys" because their 1970 floats happened to be unique. v0.2 filters single-column candidates to ID-shaped names / int / string only, AND separately searches for (id_col, time_col) composite keys.
• string_cols detection: v0.1 used dtype == object, which misses pandas 2.x string-dtype columns. v0.2 uses pd.api.types.is_string_dtype(s) or s.dtype == object.
• Binary detection: v0.1 only caught 0/1; v0.2 separately catches two-valued text columns (e.g. 东南侧/西北侧) as binary_text_cols for moderator-candidate identification.
• path variable: v0.1 referenced path without defining it; v0.2 wraps in inspect_file(file_path) with path = Path(file_path).
• xlsx multi-sheet: v0.1 silently read Sheet1; v0.2 prints a warning when multiple sheets exist.

This inspection is silent in user-facing output — show the user a one-page summary table (rows×cols, missing summary, top candidate keys, suspected design, suspected discipline), not the dict.

---

MCAR hint fallback (when scipy missing)

If scipy is not installed, use this numpy-only implementation for the focal-variable MCAR check:

import numpy as np

def welch_t_manual(a, b):
    """Welch's t-statistic without scipy. For large n (>30 each), |t|>1.96 ≈ p<0.05."""
    a = np.asarray(a)[~np.isnan(np.asarray(a, dtype=float))]
    b = np.asarray(b)[~np.isnan(np.asarray(b, dtype=float))]
    if len(a) < 2 or len(b) < 2:
        return np.nan
    se = np.sqrt(a.var(ddof=1)/len(a) + b.var(ddof=1)/len(b))
    return (a.mean() - b.mean()) / se if se > 0 else 0.0

Threshold: |t| > 1.96 flags non-MCAR at α≈0.05 with normal approximation (valid for n_miss > 30 and n_obs > 30, which is the usual case).

---

Execution scope — the 80% intake does

Auto-execute (no user choice needed) once Slots 1–5 are resolved:

Action	What	Why auto
Column rename	janitor::clean_names equivalent — strip whitespace, lowercase, snake_case, dedupe column names	Mechanical, no judgment
Dtype coercion (unambiguous)	Numeric strings with [0-9.,$%] → numeric; \d{4}-\d{2}-\d{2} strings → datetime	Mechanical
Whitespace + encoding cleanup	Strip leading/trailing whitespace, fix mojibake on UTF-8 round-trip	Mechanical
Duplicate detection	Report exact duplicates and panel-key duplicates	Inform, then ask only for resolution
Primary-key validation	If user's Slot 3 unit + Slot 2 design implies a key, assert uniqueness; stop with error if violated	Hard prerequisite
Panel structure check	If panel: compute coverage, gaps, entry/exit; report (do not force balance)	Diagnostic, no decision
Missing-rate inventory	Compute per-column missing rate, classify into <5% / 5–30% / >30% buckets, flag focal-variable missingness for MCAR hint	Diagnostic
Outlier flag	Compute z-score and IQR flags on numeric variables; only add *_outlier_z4 and *_outlier_iqr columns, do NOT winsorize or trim	Decision is method-dependent — flagship handles
Sample log initialization	Create sample_log = [("0. raw", n_rows)]; record any rows the intake itself drops	Reproducibility

Conditional-execute (ask only when data leaves it ambiguous):

Trigger	Question
Duplicates on panel key detected	"How to resolve? (A) keep most recent (B) aggregate within key (C) redefine key (D) abort and inspect"
String column with > 100 unique values that looks categorical	"Is {col} a free-text field or a categorical variable?"
Date string with ambiguous format (01/02/2020 could be Jan 2 or Feb 1)	"Date format: MDY or DMY?"
Mixed encoding detected	"Detected mixed encoding in column {col} — fix to UTF-8?"

---

What intake does NOT do (hand off to flagship)

Explicitly delegate the following to the matched flagship's Step 1 reference:

• Multiple imputation (MICE / mi / mice) — flagship has full setup; intake only flags missingness pattern
• Winsorize / trim / cap — Step 2 of flagship; intake only flags outliers
• Heckman selection / IPW for MNAR — flagship Step 5; intake only hints at MCAR plausibility
• Detailed merge with auxiliary data — flagship Step 1 has validate= / assert() / relationship patterns; intake handles only the focal dataset
• Event-study time alignment — flagship Step 2; intake only validates that event time can be constructed
• Within-group outlier detection (e.g. by industry-year) — flagship; intake does global only
• Survey weights handling — out of scope for v0; flag and note in unresolved_decisions

If user attempts any of these inside intake, redirect: "This belongs in {flagship} Step {N}. Intake will hand off the contract first."

---

Mode A — Public-health / epidemiology cleaning patterns

This is where intake does more than the flagships, because their Step 1 references omit it entirely. Full content in references/01-mode-a-epi-patterns.md. Triggered when Slot 1 = epi.

Mode A intake additionally checks:

Check	What	Why
Index date present and parseable	index_date (or named equivalent) is non-null for all rows; parsed as date	All time calculations relative to t0
Time-zero alignment	Person-time starts at index date, not at calendar enrollment, not at first record	Avoid immortal time bias
Censoring vs missing	event_date missing → check censor_date exists; if both null, raise	Survival analysis cannot run on ambiguous status
Person-time construction	If long format, verify (id, start, end) intervals do not overlap or gap	Cox / KM require valid risk sets
Washout period	If user declared washout, drop person-time before washout end and log	Standard exposure-window convention
ICD / CPT / ATC code normalization	Strip dots, leading zeros, hyphens; flag codes that do not match a known regex	Diagnosis codes are the #1 epi data-quality issue
Immortal-time-bias detection	If exposure is time-varying and person-time before exposure is coded as exposed → raise	Methodological red flag

These checks are run automatically when Slot 1 = epi; user is shown results, asked to confirm interpretation only when ambiguous.

---

Data evaluation (always runs after Mode A, before output)

Once the 80% mechanical cleaning (and Mode A checks if applicable) finish, intake always runs a built-in data evaluation phase that grades the cleaned dataset against a fixed rule library. The phase is not opt-in — it is part of intake's standard output.

Full content in references/02-data-evaluation.md. Summary:

Hard constraints

• Zero external dependencies: no networking, no external skill calls, no LLM calls, no MCPs. Pure Python rule engine over the metrics intake already computed.
• Fully deterministic: same input contract → same evaluation. Reproducibility is preserved by construction.
• Read-only on data: only writes a markdown report and augments the contract with an audit field. Never modifies any data column.

What it produces

intake/data_evaluation.md — a structured markdown report with three sections:

1. 数据表现好的地方 (Strengths) — positive signals about the dataset (balanced panel, unique pkey, focal vars clean, sufficient sample, etc.)
2. 需要完善和优化的地方 (Optimizations) — risks that affect downstream analysis (non-MCAR missingness, severe outliers, multicollinearity, counter-intuitive correlations, sample imbalance, etc.), sorted by severity (critical / high / medium / low)
3. 综合评价 (Overall Verdict) — a letter grade (A / A- / B+ / B / B- / C+ / C) with a 2-4 paragraph summary that answers: how solid is this data, what must be fixed before flagship, what to discuss in the methods section, and what journal tier this data quality can target

Rule library

23 rules total: 10 strength rules + 13 optimization rules. Each rule is a deterministic check on the contract dict (e.g. panel_structure.balanced == True, mcar_hint contains "NOT MCAR", |cor(outcome, main_x)| < 0.10). Every triggered rule outputs concrete evidence (numbers, column names, ratios) — no vague "looks bad" messages.

Grading logic

strengths_count, optims_count, n_critical, n_high, ...

if n_critical >= 3: grade = "C"
elif n_critical >= 1 and strengths_count < 3: grade = "C+"
elif strengths_count >= 7 and optims_count <= 2 and n_critical == 0: grade = "A"
elif strengths_count >= 6 and n_high <= 2 and n_critical == 0: grade = "A-"
elif strengths_count >= 5 and (n_high + n_critical) <= 4: grade = "B+"
elif strengths_count >= 3 and optims_count <= 6 and n_critical <= 2: grade = "B"
else: grade = "B-"

Contract integration

A new top-level field data_evaluation is added to data_contract.yaml:

data_evaluation:
  generated_at: "2026-04-29T15:32:00Z"
  output_file: "intake/data_evaluation.md"
  grade: "B+"
  n_strengths_triggered: 8
  n_optimizations_triggered: 6
  n_critical: 1
  n_high: 2
  n_medium: 2
  n_low: 1
  rules_triggered: { strengths: [...], optimizations: [{rule: ..., severity: ..., target: ...}, ...] }

Downstream modules can read this field to gate their own behavior (e.g., a paper-writing module can refuse to draft a methods section until grade is at least B+).

Invariants

1. Always runs (unless the user explicitly disables with --no-evaluation)
2. Never networks, never calls external models, never invokes other skills
3. Same contract input → same evaluation output (full determinism)
4. Every triggered rule must include concrete evidence in the rendered message
5. Grade is a function of strengths and optimization severity, not just counts

---

Output artifacts (the contract)

Write five files to the user's working directory under intake/:

1. intake/cleaned_dataset.{dta|parquet|rds} + intake/cleaned_dataset.xlsx (always)

Two formats are written every time, regardless of Slot 5:

• Native format matching Slot 5: .dta (Stata) | .parquet (Python) | .rds (R)
• Always-on Excel inspection workbook: cleaned_dataset.xlsx with multi-sheet layout for visual spot-check before the user commits to flagship

The Excel workbook contains 7 sheets:

Sheet	Content
cleaned_data	Substantive variables only (no outlier flag cols — they live in a separate sheet)
outlier_flags	id + year + pro + all *_outlier_z4 / *_outlier_iqr columns
rename_map	original_name → new_name table (lets user verify rename decisions)
missing_inventory	column / missing_rate / n_missing
outlier_summary	per-variable z-flag and IQR-flag counts
unresolved	unresolved decisions surfaced for the flagship
contract_summary	the YAML contract flattened to key/value pairs for at-a-glance review

The native-format file (.dta / .parquet / .rds) always includes:

• All original columns (renamed to snake_case)
• Auto-coerced types
• Outlier flag columns (*_outlier_z4, *_outlier_iqr)
• Missing flag columns for focal variables ({focal}_missing)
• For epi mode: parsed index_date, computed time_to_event, validated censor indicator

Why always also write Excel? Empirical researchers — especially in Chinese economics academia — spot-check cleaned data visually in Excel before trusting it to a Stata .dta. Refusing to write .xlsx because Slot 5 = Stata erodes trust unnecessarily. The cost is one extra file (typically smaller than the source xlsx since outlier flags are int8). Always write both.

2. intake/data_contract.yaml

Cross-language generalization of StatsPAI's data_contract() dict. The full canonical schema (with all fields, invariants, and version evolution) is at docs/contract-spec.md. The example below shows the v0.3 shape for an econ panel run:

intake_version: "0.3"
generated_at: "2026-04-29T02:30:50.876538+00:00"   # ISO-8601 UTC
source_file: "raw/panel.dta"
source_sheet: "Sheet1"                              # only for .xlsx; null otherwise
n_rows_raw: 12345
n_rows_clean: 12000
n_cols_raw: 26
n_cols_clean: 56                                    # raw + outlier flag cols added
n_outlier_flag_cols_added: 30

discipline: econ                                    # econ | epi
research_design: panel                              # cross_section | panel | time_series | cohort_index_date | repeated_cs | descriptive | exploratory
unit_of_observation: "city-year"
software_target: stata                              # python | python_statspai | stata | r

primary_key: [id, year]
key_uniqueness: 1.0                                 # must equal 1.0 (intake hard-asserts; otherwise stops)

panel_structure:
  balanced: true
  n_units: 197
  n_periods: 10
  year_range: [2014, 2023]
  coverage: 1.0
  units_with_gaps: 0

focal_vars:
  outcome: y
  main_x: x                                          # for econ
  alt_outcome: y1                                    # null if not declared
  controls: [edu, ind, fin, pepo, open, inv, hum, ind1, ind2, pgdp]
  heterogeneity: [city_cluster, resource_city, central_city]
  treatment: null                                    # null if no binary treatment; "exposure" name for epi
  treatment_note: "No binary treatment — Slot 4.3 = TWFE continuous-X"

design:
  type: two_way_fixed_effects_continuous             # or did_staggered | iv | rdd | cohort_iptw | km_cox | ...
  specification: "y = beta*x + Gamma*Controls + city_FE + year_FE + epsilon"
  cluster_level: "id (city)"

missing_pattern:                                     # flat: {col: rate}; high-missing flagged in unresolved
  city_cluster: 0.005076
  entropy_idx: 0.060914

mcar_hint_outcome: "no missingness on `y`"           # one of: "no missingness" | "likely MCAR" | "NOT MCAR — ..."
mcar_hint_alt_outcomes:                              # only for cols that have missingness
  entropy_idx: "NOT MCAR — entropy missingness differs on ind (|t|=2.81) ..."

outlier_flags:                                       # flat: {var: {n_z4, n_iqr}}
  y:    {n_z4: 31, n_iqr: 279}
  x:    {n_z4: 17, n_iqr: 105}

sample_log:                                          # list of [label, n] tuples
  - [raw, 12345]
  - [drop_missing_panel_key, 12300]
  - [drop_exact_duplicates, 12000]

# Mode A only — present iff discipline=epi
epi_checks:
  index_date_col: enrol_date
  index_date_valid: true
  n_with_missing_t0: 0
  index_date_range: ["1995-03-12", "2018-11-04"]
  time_zero_aligned: true
  n_pre_t0_events: 0
  n_pre_t0_exposures: 0
  immortal_time_risk: false
  censoring_unambiguous: true
  n_events: 487
  n_censored: 1942
  median_follow_days: 1095
  follow_time_col: follow_time
  status_col: status
  code_normalization:
    dx_code: {scheme: ICD-10, n_total: 12000, n_invalid: 7}

renames_applied:                                      # original → new column names
  Y: y
  所属地域: region

unresolved_decisions:                                 # human-readable list, surfaced for flagship
  - "tenure has 8% missingness and NOT MCAR — use MI in flagship Step 1.5"
  - "Survey weights not detected; declare in flagship before regression if applicable"

# v0.3 — always present (data evaluation runs every time)
data_evaluation:                                      # see contract-spec.md and references/02-data-evaluation.md for full semantics
  generated_at: "2026-04-29T15:32:00Z"
  output_file: "intake/data_evaluation.md"
  grade: "B+"
  n_strengths_triggered: 8
  n_optimizations_triggered: 6
  n_critical: 1
  n_high: 2
  n_medium: 2
  n_low: 1
  rules_triggered:
    strengths:  [balanced_panel, unique_pkey, clean_focal_vars, ...]
    optimizations:
      - {rule: not_mcar_alt_outcome, severity: high, target: entropy_idx}
      - {rule: pgdp_dominates_x, severity: medium}

routing_recommendation:
  flagship: "00.2-Full-empirical-analysis-skill_Stata"
  mode: default                                       # default | mode_a_epi | mode_b_ml_causal
  reason: "Slot 5=Stata + balanced panel + econ + TWFE continuous → 00.2 default"
  next_step_in_flagship: "Step 1.5 (advanced missing) → Step 2 → Step 5 (reghdfe)"

Schema invariants (any deviation is a contract violation, intake should refuse to write):

• key_uniqueness == 1.0 always — if not, intake stops before writing
• n_rows_raw - n_rows_clean == sum of (drops in sample_log after [raw, ...]) — sample log must reconcile
• If discipline == "epi", epi_checks block MUST be present
• If software_target == "stata", all column names in renames_applied.values() MUST be ASCII and ≤32 chars (Stata variable name limit)
• missing_pattern only includes columns with non-zero missing rate (zero-missing cols are omitted to keep the contract small)
• data_evaluation MUST always be present in v0.3+ contracts (unless intake was invoked with --no-evaluation)
• data_evaluation.grade MUST be one of: A | A- | B+ | B | B- | C+ | C (no other values allowed)
• The triggered rules in data_evaluation.rules_triggered MUST be deterministic given the contract — same input → same triggered set

3. intake/routing_recommendation.md

A human-readable version of the routing portion above, with explicit links to the flagship's relevant sections. Format:

# Intake routing recommendation

You should now invoke flagship: **00.2 Stata** (mode: default).

## Why this flagship

- File extension `.dta` → Stata native
- Detected panel structure → flagship Stata Step 5 has `reghdfe` + `csdid` + `did_imputation`
- Discipline = econ + research design = causal identification → default AER mode

## What's already done (don't redo)

- Column rename, dtype coercion, primary-key validation, duplicate handling, outlier flagging, panel structure check.

## What flagship should do next

- **Step 1.5 (Multiple imputation)**: tenure has 8% missingness, NOT MCAR. Use `mi impute chained`.
- **Step 2 (Variable construction)**: winsorize wage at 1/99% (intake only flagged outliers).
- **Step 3 (Table 1)**: balance table on training, with focus on tenure and education.
- **Step 5 (Baseline)**: `reghdfe wage training age edu tenure, absorb(worker_id year) vce(cluster worker_id)`.

## Open issues you must address in the flagship

- 23 units with year gaps. Decide whether to drop or accept unbalanced panel.
- Survey weights not detected — if applicable, declare before regression.

4. intake/data_evaluation.md

Produced on every run. Markdown report with three sections — strengths, optimizations (sorted critical → high → medium → low), and an overall letter-grade verdict with a 2–4 paragraph summary. Rendered deterministically from the rule engine described above.

Required structure:

# 数据评价报告

**生成时间**: <ISO-8601 UTC>
**数据**: <source_file> (<sheet>, <n_rows> 行 × <n_cols> 列, <n_units> 单位 × <n_periods> 期)
**研究设计**: <research_design> | <discipline> | <unit_of_observation>

## 一、数据表现好的地方
1. **<strength rule name>**: <evidence-bearing message>
...

## 二、需要完善和优化的地方

### 关键问题（必须处理）
<critical-severity items>

### 重要问题（强烈建议处理）
<high-severity items>

### 一般问题（论文写作时讨论）
<medium- and low-severity items>

## 三、综合评价

**评级**: <A / A- / B+ / B / B- / C+ / C>

<2-4 paragraphs answering: solid base? what must be fixed?
 what to discuss in methods? recommended journal tier?>

Each rule's rendered line MUST contain concrete evidence (a number, a column name, a ratio) — no "looks bad" or other vague claims.

5. (Mode A only) intake/epi_checks_report.md

For Mode A (epi) runs, an additional human-readable report summarizing the 7 epi-specific checks (index date validation, time-zero alignment, censoring vs missing, person-time, washout, code normalization, immortal-time-bias screen). The structured outcomes are already inside data_contract.yaml > epi_checks; this file is the readable companion.

For econ (default) runs, this file does not exist.

---

Operating instructions for Claude

When this skill is invoked:

1. First: identify the data file from user message or working directory. If unclear, ask for the path.
2. Run static inspection silently. Show user a one-page summary table of: rows × cols, missing summary, top candidate keys, suspected design, suspected discipline.
3. Walk through Slots 1–5 in order. For each slot:
   • Compute mode (AUTO / CONFIRM / ASK) from inspection
   • Skip if AUTO
   • Show single-line confirmation if CONFIRM
   • Show multiple-choice if ASK
   • One question per turn — never batch unrelated slots
4. Execute the 80% auto-cleaning pipeline. Print [intake] log lines for every row drop.
5. Run Mode A checks if Slot 1 = epi.
6. Run data evaluation on the in-memory contract. Execute all 23 rules (10 strengths + 13 optimizations), compute the grade, render intake/data_evaluation.md, and augment the contract with the data_evaluation audit field. This step always runs; it does not call any external tool, network, or model.
7. Write the five output files to <parent>/intake/. Show the user the file paths. The files are: cleaned_dataset.{dta|parquet|rds}, cleaned_dataset.xlsx (always), data_contract.yaml, routing_recommendation.md, and data_evaluation.md. Mode A runs additionally produce epi_checks_report.md.
8. Print the routing message ("now invoke flagship 00.X — see intake/routing_recommendation.md") and the one-line evaluation summary ("Data evaluation grade: B+ — see intake/data_evaluation.md for details").

Never:

• Silently drop rows (always print count + reason)
• Winsorize / impute (those are flagship's job)
• Skip Slot 1 (discipline is never inferable from data)
• Write to anywhere other than intake/ and the user's specified output path

If the user's data triggers a hard error (no rows, no columns, fully duplicated keys with no resolution path), stop and report — do not produce a contract.

Output directory convention

By default, write the four output files to <parent-of-source-file>/intake/ (sibling to the data folder, not inside it). For example, /path/to/yjn/data/foo.xlsx → outputs at /path/to/yjn/intake/. If intake/ already exists from a previous run, overwrite without prompting but log [intake] overwriting previous run at {path}. If the user wants to preserve previous runs, instruct them to rename the old intake/ to intake_backup/ before re-invoking.

Stata-target column-name guard

When software_target == "stata", after rename, assert every column name is:

• ASCII only (no Chinese, no whitespace, no punctuation except _)
• Starts with a letter
• ≤ 32 characters
• Not a Stata reserved word (if, in, using, _n, _N, _merge, etc.)

If any column fails, append a numeric suffix (col_1, col_2) and log the change to renames_applied. Never silently truncate — Stata silently truncating to 32 chars has caused real-world publication bugs.

---

Version

• v0.3 (2026-04-29) — Adds always-on data evaluation phase:
  • New phase between Mode A and Output: deterministic rule-engine evaluation of the cleaned dataset.
  • 23 rules total (10 strengths + 13 optimizations), each with concrete evidence rendering and a severity label (critical / high / medium / low) for optimizations.
  • Letter-grade verdict (A / A- / B+ / B / B- / C+ / C) computed from a fixed function of strengths and optimization severity.
  • New output file intake/data_evaluation.md (5th file alongside the 4 v0.2 outputs).
  • New contract field data_evaluation recording grade, triggered rules, and severity counts — readable by downstream modules.
  • Zero external dependencies: no networking, no LLM calls, no MCP, no calls to other skills. Pure Python over already-computed metrics. Reproducibility 100%.
  • See references/02-data-evaluation.md for the full rule library and grading function.
  • Earlier v0.3 work-in-progress (literature consultation, opt-in academic search) was rolled back in favor of this self-contained design — researchers asked for a direct evaluation rather than an external advisory layer.
• v0.2 (2026-04-29) — Bug-fix release after first real-data test:
  • cleaned_dataset.xlsx 7-sheet inspection workbook now always written (in addition to Slot-5 native format).
  • Inspection code rewritten as inspect_file() function — fixes path undefined bug.
  • Composite primary-key detection added — v0.1 misidentified continuous numeric measures as primary keys.
  • string_cols now uses pd.api.types.is_string_dtype — pandas 2.x string-dtype columns were silently missed by dtype == object.
  • binary_text_cols separately captured (e.g. 东南侧/西北侧) — was previously lost.
  • xlsx multi-sheet warning added — v0.1 silently read Sheet1.
  • YAML schema in this file now matches actual production output (was inconsistent in v0.1).
  • Prerequisites section added; scipy fallback documented.
  • Output directory convention pinned to <parent>/intake/.
  • Stata-target column-name guard added (32-char limit, reserved words).
• v0.1 (2026-04-29) — first cut. Audit report at audit-flagship-cleaning.md.