Drive IV Desk from your own code

// make-body.js - build the run body with the SAME engine the web page uses.
// Save https://iv-desk.skillsafe.ai/ivkit.js next to this file.
// Usage: node make-body.js loans.csv
const fs = require("fs");
const K = require("./ivkit.js");

const table = K.parseCsv(fs.readFileSync(process.argv[2] || "loans.csv", "utf8"));
const A = K.analyze(table, {
  oos: "partner_d",              // held-out organizations, or ""
  settings: {}                   // thresholds; {} = IvKit.DEFAULTS
});
const body = K.buildInput(A, {
  dataset: "retail-instalment-2025",
  context: "Unsecured instalment loans. Target = 1 if 60+ days past due " +
           "within 12 months of booking.",
  question: ""                   // non-empty: answered in tldr as "Answer: ..."
});
fs.writeFileSync("body.json", JSON.stringify(K.mustBeObject(body)));
console.log(A.flags.length, "flags; browser verdict", A.browser_verdict);
console.log("Idempotency-Key: iv-desk:review:" + K.hashInput(body) + ":a1");

{"ok": true, "data": {"job_id": "job_...", "status": "queued"}}
{"ok": false, "error": {"code": "payment_required", "message": "..."}}

# Every call is the same three things: the base URL, your bearer token,
# and a JSON body. Keep the token in a shell variable.
BASE="https://api.skillsafe.ai/v1/app-api"
SLUG="iv-desk"
TOKEN="$SKILLSAFE_TOKEN"   # from https://iv-desk.skillsafe.ai/tokens.html

call() {                  # call <path> [json-body]
  if [ -n "$2" ]; then
    curl -sS -X POST "$BASE/$1" \
      -H "Authorization: Bearer $TOKEN" \
      -H "Content-Type: application/json" \
      -d "$2"
  else
    curl -sS "$BASE/$1" -H "Authorization: Bearer $TOKEN"
  fi
}

# The token page is the shortest path. It shows the token this browser holds and
# hands you a ready-made shell export:
#
#   https://iv-desk.skillsafe.ai/tokens.html
#   export SKILLSAFE_TOKEN="..."
#
# To mint a guest token from the command line instead. A guest token is enough
# for /me and /estimate; a run needs a personal token from signing in.
curl -sS -X POST "https://api.skillsafe.ai/v1/app-api/guest" \
  -H "Content-Type: application/json" -d '{"slug":"iv-desk"}'
# {"ok":true,"data":{"token":"…","subject_type":"guest"}}

call me
# {"ok":true,"data":{"subject_type":"user","username":"you","credits":51234}}

# body.json is the input object itself - no {"input": ...} wrapper. Download it with
# the page's free "Run input .json" button, or build it with make-body.js above.
# estimate does not validate it, so check the shape first:
python3 - <<'EOF'
import json
b = json.load(open("body.json"))
assert isinstance(b, dict) and "input" not in b, "no {input: ...} wrapper"
assert all(isinstance(v, str) for v in b.values()), "every value is a string"
assert b.get("task") == "review", "the only lane is review"
assert b.get("facts", "").strip(), "facts is required"
assert isinstance(json.loads(b["facts"]), dict), "facts is a JSON string of an object"
EOF
INPUT=$(cat body.json)

call estimate "$INPUT"
# {"ok":true,"data":{"model":"...","model_alias":"gpt-terra",
#   "markup_bps":...,"hold_credits":...,"min_credits":...,
#   "sponsor_enabled":false,"warnings":[]}}
#
# estimate creates no job and charges nothing. hold_credits is RESERVED, not the
# price; charged_credits after the run is the actual cost, usually far lower.

# Always send an Idempotency-Key derived from the input. A retried request with
# the same key returns the SAME job instead of billing a second run.
# One lane, so the key is iv-desk:review:<hash>:a<attempt>.
KEY="iv-desk:review:$(printf '%s' "$INPUT" | shasum -a 256 | cut -c1-16):a1"

JOB=$(curl -sS -X POST "$BASE/run" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $KEY" \
  -d "$INPUT" | python3 -c 'import sys,json;print(json.load(sys.stdin)["data"]["job_id"])')

while :; do
  OUT=$(call "jobs/$JOB")
  STATUS=$(printf '%s' "$OUT" | python3 -c 'import sys,json;print(json.load(sys.stdin)["data"]["status"])')
  [ "$STATUS" = "succeeded" ] && break
  [ "$STATUS" = "failed" ] && echo "$OUT" && exit 1
  sleep 2
done

# {"ok":true,"data":{"job_id":"job_...","status":"succeeded",
#   "output":{"output":"{\"task\":\"review\",\"verdict\":\"revise\", ...}"},
#   "charged_credits":...,"truncated":false}}
printf '%s' "$OUT" | python3 -c \
  'import sys,json;print(json.load(sys.stdin)["data"]["output"]["output"])' > reply.json

# Server-sent events. `delta` events carry chunks of the reply; `done` carries the
# status, charged_credits and the truncated flag. Ignore `tick` heartbeats.
curl -N -X POST "$BASE/run-stream" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $KEY" \
  -H "Accept: text/event-stream" \
  -d "$INPUT"

# event: job    {"job_id":"job_..."}
# event: delta  {"text":"{\"task\":\"review\",\"verdict\":\"revise\",\"headline\":\"This"}
# event: done   {"status":"succeeded","charged_credits":...,"truncated":false}

# reply.json holds data.output.output from step 5. Strip any fence, keep the object:
python3 - <<'EOF'
import json, re
t = open("reply.json").read().strip()
t = re.sub(r"^```(?:json)?\s*", "", t, flags=re.I)
t = re.sub(r"\s*```\s*$", "", t)
r = json.loads(t[t.index("{"):t.rindex("}") + 1])
print(r["task"], r["verdict"], "-", r["headline"])
for s in r["step_reviews"]:
    print("STEP", s["step"], s["verdict"], "|", s["note"])
for c in r["feature_calls"]:
    print("CALL", c["feature"], c["call"], "(screen:", c["engine"], "iv", c["iv"], ")")
for s in r["leakage_suspects"]:
    print("LEAK?", s["feature"], "| check:", s["check"])
for a in r["threshold_advice"]:
    print("SET", a["parameter"], a["current"], "->", a["suggested"])
EOF

# assert-reply.py - the page's reconciliation checks (recon.js), for body.json
# and reply.json from the steps above.
import json
body = json.load(open("body.json")); facts = json.loads(body["facts"])
t = open("reply.json").read(); r = json.loads(t[t.index("{"):t.rindex("}") + 1])
KEYS = ["task", "verdict", "headline", "tldr", "step_reviews", "feature_calls",
        "leakage_suspects", "threshold_advice", "next_steps", "memo", "prescan_responses"]
STEPS = ["sample", "months", "missing", "iv", "psi", "noise", "corr"]
missing = [k for k in KEYS if k not in r]
assert not missing, "keys missing: " + ", ".join(missing)
assert r["task"] == "review", "task is " + str(r["task"])
assert r["verdict"] in ("ready", "revise", "blocked")
# Steps: exactly one review per step, in order.
assert [s["step"] for s in r["step_reviews"]] == STEPS, "one review per step, in order"
assert all(s["verdict"] in ("agree", "adjust", "question") for s in r["step_reviews"])
# Flags: each answered exactly once, none invented.
flags = {f["id"]: f for f in facts["flags"]}
answered = [p["ref"].upper() for p in r["prescan_responses"]]
assert sorted(answered) == sorted(flags), "every flag answered exactly once"
assert all(p["verdict"] in ("confirmed", "dismissed") for p in r["prescan_responses"])
resp = {p["ref"].upper(): p for p in r["prescan_responses"]}
# Features: real names, the screen's keep/drop and IV quoted exactly.
feats = {f["name"]: f for f in facts["features"]}
known = set(feats) | set(facts["survivors"]) | {n for f in facts["flags"] for n in f["features"]}
calls = {c["feature"]: c for c in r["feature_calls"]}
named = list(calls) + [s["feature"] for s in r["leakage_suspects"]]
assert all(n in known for n in named), "a feature not in your file"
for c in r["feature_calls"]:
    assert c["call"] in ("keep", "drop", "investigate")
    f = feats.get(c["feature"])
    if not f:
        continue                      # named in a flag but not detailed
    assert c["engine"] == ("kept" if f["drops"] == "kept" else "dropped"), c["feature"]
    if c["iv"] is not None:
        assert "iv" in f and abs(c["iv"] - f["iv"]) <= 0.00015, "IV misquoted: " + c["feature"]
# Survivors covered; a leakage suspect is never kept.
assert all(n in calls for n in facts["survivors"][:30]), "a survivor has no call"
suspects = {s["feature"] for s in r["leakage_suspects"]}
assert not [n for n in suspects if calls.get(n, {}).get("call") == "keep"], "suspect kept"
# Confirmed high/medium leakage flags carried (or explicitly cleared in the note).
for fid, f in flags.items():
    p = resp[fid]
    if f["category"] != "leakage" or f["severity"] == "low" or p["verdict"] != "confirmed":
        continue
    for n in f["features"]:
        kept = calls.get(n, {}).get("call") == "keep"
        assert n in suspects or (n in calls and not kept) or (kept and n in p["note"]), n
# Threshold advice names a real setting and quotes its current value.
S = facts["settings"]
for a in r["threshold_advice"]:
    assert a["parameter"] in S, "no such setting: " + a["parameter"]
    cur, real = a["current"], S[a["parameter"]]
    if isinstance(real, str):
        assert str(cur).replace(" ", "") == real.replace(" ", ""), a["parameter"]
    else:
        assert abs(float(cur) - real) <= max(1e-6, abs(real) * 0.002), a["parameter"]
# Verdict: never looser than the flags left standing.
rank = {"ready": 0, "revise": 1, "blocked": 2}
floor = 0
for fid, f in flags.items():
    if resp[fid]["verdict"] == "dismissed":
        continue
    lvl = 2 if f["severity"] == "high" and f["category"] == "sample" else 1 if f["severity"] != "low" else 0
    floor = max(floor, lvl)
if not facts["survivors"]:
    floor = 2
assert rank[r["verdict"]] >= floor, "verdict looser than the flags left standing"
overrides = [c["feature"] for c in r["feature_calls"] if c["call"] == "keep" and c["engine"] == "dropped"]
print("ok:", r["verdict"], len(r["feature_calls"]), "calls;", "overrides:", overrides or "none")

{"task":"review",
 "verdict":"ready|revise|blocked",
 "headline":"...",
 "tldr":["...","..."],
 "step_reviews":[{"step":"sample","verdict":"agree|adjust|question","note":"..."}],
 "feature_calls":[{"feature":"...","call":"keep|drop|investigate",
                   "engine":"kept|dropped","iv":0.1234,"reason":"...","ref":"F2"}],
 "leakage_suspects":[{"feature":"...","evidence":"...","check":"..."}],
 "threshold_advice":[{"parameter":"psi_max_orgs","current":6,"suggested":2,"why":"..."}],
 "next_steps":["...","..."],
 "memo":["...","..."],
 "prescan_responses":[{"ref":"F1","verdict":"confirmed|dismissed","note":"..."}]}

{
 "task": "review",
 "dataset": "retail-instalment-2025",
 "context": "Unsecured instalment loans, three lenders plus one partner held out of sample. Target = 1 if 60+ days past due within 12 months of booking. Features are meant to be known at application time.",
 "facts": "{\"roles\":{\"date\":\"apply_date\",\"label\":\"target\",\"org\":\"org_info\",\"keys\":[\"loan_id\"]},\"settings\":{\"sentinels\":\"-1, -999, -1111\",\"min_ym_bad\":10,\"min_ym_…",
 "question": "Which of these would you take into a logistic scorecard?"
}

{
 "roles": {"date": "apply_date", "label": "target", "org": "org_info", "keys": ["loan_id"]},
 "settings": {
  "sentinels": "-1, -999, -1111", "min_ym_bad": 10, "min_ym_total": 500,
  "missing_ratio": 0.6, "overall_iv": 0.1, "org_iv": 0.1,
  "max_low_iv_orgs": 2, "psi": 0.1, "psi_months_ratio": 0.3333,
  "psi_max_orgs": 6, "psi_min_month": 100, "null_perms": 10,
  "max_corr": 0.9, "top_n_keep": 20
 },
 "sample": {
  "rows_in": 22273, "invalid_label": 0, "invalid_date": 0,
  "duplicates": 0, "rows_valid": 22273, "oos_rows": 2899,
  "modeling_rows_before_months": 19374, "modeling_rows": 19374, "modeling_bads": 2122,
  "modeling_bad_rate": 0.1095, "months_total": 12, "months_dropped": 0
 },
 "organizations": [
  {"org": "bank_a", "type": "modeling", "rows": 7592, "bads": 818, "bad_rate": 0.1077, "months": 12},
  {"org": "bank_b", "type": "modeling", "rows": 6647, "bads": 655, "bad_rate": 0.0985, "months": 12},
  {"org": "broker_c", "type": "modeling", "rows": 5135, "bads": 649, "bad_rate": 0.1264, "months": 12},
  {"org": "partner_d", "type": "oos", "rows": 2899, "bads": 329, "bad_rate": 0.1135, "months": 12}
 ],
 "months": [
  {"ym": "202501", "rows": 1643, "bads": 191, "dropped": false, "reason": ""},
  {"ym": "202502", "rows": 1591, "bads": 176, "dropped": false, "reason": ""}
   … 10 more months, 202503 to 202512, all with dropped false
 ],
 "step_counts": {"constant": 1, "missing": 1, "iv": 13, "psi": 0, "noise": 8, "corr": 0},
 "features_total": 20,
 "features_detailed": 20,
 "features": [
  {"name": "overdue_days_m3", "type": "numeric", "missing": 0, "iv": 16.2826, "drops": "kept", "psi_max": 0.0219, "null_margin": 16.2709},
  {"name": "bureau_score", "type": "numeric", "missing": 0, "iv": 0.5756, "drops": "kept", "psi_max": 0.11, "null_margin": 0.5633},
  {"name": "util_revolving", "type": "numeric", "missing": 0, "iv": 0.14, "drops": "kept", "psi_max": 0.1154, "null_margin": 0.1259, "top_corr": "util_revolving_pct 0.9998"},
  {"name": "util_revolving_pct", "type": "numeric", "missing": 0, "iv": 0.1393, "drops": "kept", "psi_max": 0.117, "null_margin": 0.1226, "top_corr": "util_revolving 0.9998"},
  {"name": "noise_a", "type": "numeric", "missing": 0.0002, "iv": 0.0155, "drops": "iv+noise", "low_iv_orgs": 3, "psi_max": 0.0943, "null_margin": 0, "sentinel_hits": 4},
  {"name": "fraud_score_v2", "type": "numeric", "missing": 0, "iv": 0.0147, "drops": "iv+noise", "low_iv_orgs": 3, "psi_max": 0.3709, "psi_unstable_orgs": 3, "null_margin": -0.0018},
  {"name": "bal_change_3m", "type": "numeric", "missing": 0.0691, "iv": 0.0126, "drops": "iv+noise", "low_iv_orgs": 3, "psi_max": 0.1035, "null_margin": -0.0046, "sentinel_hits": 1339},
  {"name": "const_flag", "type": "categorical", "missing": 0, "drops": "constant"}
   … 12 more features: dti, inq_6m, income_monthly, employment_months, social_score, app_channel, residence_type, age, loan_amount, noise_b, device_os, term_months
 ],
 "survivors": ["bureau_score", "dti", "inq_6m", "util_revolving", "util_revolving_pct", "overdue_days_m3"],
 "survivor_count": 6,
 "correlated_pairs": [
  {"a": "util_revolving", "b": "util_revolving_pct", "r": 0.9998}
 ],
 "noise_screen_ran": true,
 "flags": [
  {"id": "F1", "severity": "medium", "category": "psi", "message": "The PSI rule needs 6 unstable organizations to drop a feature, but the modeling sample has only 3 - the PSI step cannot drop anything as set.", "features": []},
  {"id": "F2", "severity": "high", "category": "leakage", "message": "2 feature(s) have overall IV above 0.5 - bureau_score (0.5756), overdue_days_m3 (16.2826). IV that strong usually means the feature was observed after the outcome or encodes the label.", "features": ["bureau_score", "overdue_days_m3"]},
  {"id": "F3", "severity": "medium", "category": "missing", "message": "Sentinel values were treated as missing in feature(s) that also hold other negative numbers, so a real value may have been blanked: noise_a, bal_change_3m.", "features": ["noise_a", "bal_change_3m"]},
  {"id": "F4", "severity": "low", "category": "sample", "message": "1 constant column(s) dropped before screening: const_flag.", "features": ["const_flag"]},
  {"id": "F5", "severity": "medium", "category": "psi", "message": "Unstable month over month in every modeling organization, yet kept by the PSI rule as set: fraud_score_v2 (max PSI 0.3709).", "features": ["fraud_score_v2"]},
  {"id": "F6", "severity": "medium", "category": "corr", "message": "1 pair(s) above the correlation threshold were left in place because both sides are among the top 20 features by IV: util_revolving ~ util_revolving_pct (r 0.9998).", "features": ["util_revolving", "util_revolving_pct"]}
 ],
 "browser_verdict": "revise",
 "method_notes": [
  "IV bins: best-first decision tree, at most 5 leaves, 1% minimum leaf, missing as its own bin (toad dt recipe, re-implemented).",
  "Noise screen: univariate null-IV test against label permutations - a stand-in for the reference's LightGBM null-importance gain, which does not run in a browser.",
  "Correlation drop keeps the higher-IV feature of each pair (the reference uses LightGBM gain).",
  "PSI: month-over-month within each organization, 10 quantile bins of the earlier month plus a missing bin; values above the earlier month's maximum go to the top bin."
 ]
}

{
 "task": "review",
 "verdict": "revise",
 "headline": "This screen needs one leakage check before modeling: overdue_days_m3's near-perfect IV must be confirmed as application-time before the survivors go to binning.",
 "tldr": [
  "Answer: dti, inq_6m and bureau_score are safe for the scorecard today; hold overdue_days_m3 until its timing is confirmed, and use only one of util_revolving/util_revolving_pct.",
  "6 survivors from 19,374 modeling rows across 3 organizations (2,122 bads, 10.95% bad rate), with partner_d's 2,899 rows correctly held out of sample.",
  "overdue_days_m3 (IV 16.28, null_margin 16.27) is too strong to trust without a timing check; bureau_score's IV 0.58 is plausible for a bureau score and is cleared.",
  "PSI and correlation both have configuration gaps: psi_max_orgs=6 can never fire with only 3 organizations, and top_n_keep=20 equals the full feature count so the correlation guard never engages.",
  "Sentinel handling likely masks real values in bal_change_3m and noise_a - all of their recorded missingness is sentinel hits, though both are already IV-dropped so it costs nothing here."
 ],
 "step_reviews": [
  {"step": "sample",
   "verdict": "agree",
   "note": "22,273 rows in, no invalid labels/dates, no duplicates. 2,899 partner_d rows held out. Modeling sample is 19,374 rows, 2,122 bads (10.95%) across bank_a/bank_b/broker_c - plenty of bads for a stable screen."},
  {"step": "months",
   "verdict": "agree",
   "note": "12 months, none dropped. Monthly rows run 1,575-1,661 and bads 157-191, comfortably above min_ym_bad=10 and min_ym_total=500 - the calendar is sound."},
  {"step": "missing",
   "verdict": "adjust",
   "note": "The 0.6 cutoff correctly caught only social_score (0.7588 missing). But bal_change_3m's whole missing share (0.0691) and noise_a's (0.0002) equal their sentinel_hits exactly - every 'missing' there is a sentinel replacement, not a true gap; review the sentinel list."},
  {"step": "iv",
   "verdict": "agree",
   "note": "overall_iv=0.1 with org_iv=0.1/max_low_iv_orgs=2 correctly dropped 13 weak features and kept the 6 survivors plus bureau_score and overdue_days_m3, which the leakage check (F2) handles separately."},
  {"step": "psi",
   "verdict": "question",
   "note": "psi_max_orgs=6 exceeds the 3 modeling organizations that exist, so the rule structurally cannot drop a feature (F1); step_counts.psi=0 confirms it never has. fraud_score_v2 is unstable in all 3 orgs (PSI 0.3709) yet survived (F5)."},
  {"step": "noise",
   "verdict": "agree",
   "note": "The null-IV permutation screen correctly caught both deliberately noisy probes, noise_a and noise_b (null_margin ~0 and -0.0126), a good sanity check that the screen works as intended."},
  {"step": "corr",
   "verdict": "question",
   "note": "top_n_keep=20 equals features_total=20, so the 'both in the top N' protection is trivially true for every pair; util_revolving/util_revolving_pct (r=0.9998) were kept together with no real prioritization (F6)."}
 ],
 "feature_calls": [
  {"feature": "overdue_days_m3",
   "call": "investigate",
   "engine": "kept",
   "iv": 16.2826,
   "reason": "IV 16.2826 with null_margin 16.2709 is far beyond what an application-time feature should show; the name suggests a days-overdue window that may overlap the 12-month performance period used to build the target. Hold for a timing check before scorecard use.",
   "ref": "F2"},
  {"feature": "bureau_score",
   "call": "keep",
   "engine": "kept",
   "iv": 0.5756,
   "reason": "IV 0.5756 (null_margin 0.5633) is high but a genuine bureau score can legitimately reach that on its own; no leakage indicator beyond IV strength. Safe for the scorecard.",
   "ref": "F2"},
  {"feature": "util_revolving",
   "call": "keep",
   "engine": "kept",
   "iv": 0.14,
   "reason": "IV 0.14, the stronger side of a near-duplicate pair (r 0.9998 with util_revolving_pct). Keep this one.",
   "ref": "F6"},
  {"feature": "util_revolving_pct",
   "call": "drop",
   "engine": "kept",
   "iv": 0.1393,
   "reason": "Overriding the corr step: it kept this feature only because top_n_keep=20 equals the full feature count, so the correlation guard never actually screens anything here. r=0.9998 with util_revolving means it adds no information; drop the redundant copy.",
   "ref": "F6"}
   … 7 more calls: dti keep, inq_6m keep, noise_a drop, bal_change_3m drop, const_flag drop, fraud_score_v2 drop, income_monthly investigate
 ],
 "leakage_suspects": [
  {"feature": "overdue_days_m3",
   "evidence": "IV 16.2826, essentially unchanged after label permutation (null_margin 16.2709) - far outside the range a legitimate application-time feature should show.",
   "check": "Confirm overdue_days_m3 is computed strictly from data timestamped before apply_date and not from the 12-month performance window; rebuild it from pre-application snapshots only and rescore its IV."}
 ],
 "threshold_advice": [
  {"parameter": "sentinels",
   "current": "-1, -999, -1111",
   "suggested": "-999, -1111",
   "why": "bal_change_3m's missing share (0.0691) equals its sentinel_hits (1339) over modeling_rows exactly, and noise_a's missing (0.0002) equals its 4 sentinel_hits the same way - every recorded 'missing' value in both is a sentinel hit. Confirm -1 is not a legitimate small value before keeping it as a sentinel."},
  {"parameter": "psi_max_orgs",
   "current": 6,
   "suggested": 2,
   "why": "Only 3 modeling organizations exist (bank_a, bank_b, broker_c), so psi_max_orgs=6 can never be reached (F1); fraud_score_v2 is unstable (PSI 0.3709) in all 3 yet survived (F5)."},
  {"parameter": "top_n_keep",
   "current": 20,
   "suggested": 8,
   "why": "top_n_keep=20 equals features_total=20, so the correlated-pair protection always applies to any pair; util_revolving and util_revolving_pct (r=0.9998) were kept together with no real prioritization (F6)."}
 ],
 "next_steps": [
  "Confirm overdue_days_m3 is built from data timestamped strictly before apply_date, not the 12-month performance window; rebuild and rescore if it isn't.",
  "Drop util_revolving_pct and keep util_revolving (r=0.9998, marginally higher IV) before binning.",
  "Lower psi_max_orgs to 2 of 3 organizations and rerun PSI so instability like fraud_score_v2's can actually drop a feature.",
  "Lower top_n_keep well below features_total (e.g. 8) so the correlation guard stops protecting every pair by default.",
  "Review the sentinel list (-1, -999, -1111) against bal_change_3m and noise_a specifically before trusting their missing shares.",
  "Take bureau_score, dti, inq_6m and util_revolving into scorecard binning now; hold overdue_days_m3 pending the leakage check.",
  "Revisit income_monthly after cleaning its sentinel-driven missingness (943 hits) - its IV of 0.0548 is borderline and worth a second look."
 ],
 "memo": [
  "The retail-instalment-2025 sample has 22,273 loan-level rows, no invalid labels or dates and no duplicates. 2,899 rows from partner_d are held out of sample entirely; the modeling sample is bank_a, bank_b and broker_c, 19,374 rows, 2,122 bads (10.95% bad rate) over 12 months, none dropped. Monthly bads run 157-191 and rows 1,575-1,661, comfortably above the min_ym_bad and min_ym_total floors.",
  "Of 20 features, const_flag was removed as constant, social_score failed missing-rate (75.88%), 13 failed IV, 8 failed noise (incl. probes noise_a/noise_b), 0 failed PSI/corr - not for lack of instability or correlation, but because psi_max_orgs=6 exceeds the 3 organizations and top_n_keep=20 equals features_total. fraud_score_v2 and util_revolving_pct both slipped through that way."
   … 2 more paragraphs: leakage and stability concerns, survivors and next steps
 ],
 "prescan_responses": [
  {"ref": "F1",
   "verdict": "confirmed",
   "note": "Only 3 modeling organizations exist against psi_max_orgs=6 - the rule needs more unstable organizations than exist, so PSI can never drop a feature as configured (step_counts.psi=0 confirms it never has)."},
  {"ref": "F2",
   "verdict": "confirmed",
   "note": "overdue_days_m3 (IV 16.2826, null_margin 16.2709) is implausibly strong for an application-time feature and is treated as a leakage suspect. bureau_score (IV 0.5756, null_margin 0.5633) is a standard bureau score that can legitimately reach this IV on its own - cleared, kept as a survivor."},
  {"ref": "F3",
   "verdict": "confirmed",
   "note": "bal_change_3m's missing share (0.0691) exactly equals its sentinel_hits over modeling_rows (1339/19374), so every 'missing' value there is a sentinel replacement, not a true gap; noise_a is the same (4/19374). Both are already IV-dropped so there is no scorecard impact, but the sentinel list should be reviewed."}
   … F4 confirmed, F5 confirmed, F6 confirmed
 ]
}