Datasets:

Bioaligned
/

coupled-welfare-corpus

#!/usr/bin/env python3
"""Coupled-welfare corpus schema v2 — validator, corpus-quota checker, and v1 migrator.
Spec: SCHEMA.md. Plan: ../Aug_scaled_CW_traing.md §5.
The point of this file is that the three v1 failures (no H->B direction, monotone resolution,
absent epistemic status) were failures the v1 schema could not *express*, so nothing could catch
them. Record-level validity was never the problem — every v1 doc was individually fine and the
corpus was still 100% one publisher running one frame. Hence --quotas, which is where the teeth are.
Usage:
    python3 schema.py validate  <file.jsonl> [...]      # record-level
    python3 schema.py quotas    <file.jsonl> [...]      # corpus-level (pooled across files)
    python3 schema.py migrate   <in.jsonl> <out.jsonl>  # best-effort v1 -> v2
    python3 schema.py report    <file.jsonl> [...]      # validate + quotas, human-readable
No third-party dependencies by design — this runs anywhere the corpus does.
"""
from __future__ import annotations

import argparse
import collections
import json
import re
import sys

SCHEMA_VERSION = "2.0"

# ---------------------------------------------------------------- enumerations

# `expository` is new in the pilot: harvested world-model text (open-access papers, Wikipedia,
# government reports) that describes how the world works but contains no agent making a decision.
# It carries dependency_edges and epistemic_status like everything else, but NOT resolution_type or
# positive_sum_option_existed — there is no resolution to type. This is the cheap stratum, and per
# the A3 result (the winning arm trained closest to descriptive text) it is also the one the
# evidence most supports scaling.
TRACKS = {"mined_recast", "generated", "gold", "recovery", "expository"}
DECISION_TRACKS = {"mined_recast", "generated", "gold", "recovery"}
AXES = {"H", "B", "A", "HB", "HA", "BA", "HBA"}
REGISTERS = {"first_person", "third_person_expository", "case_memo", "dialogue", "qa"}
EPISTEMIC = {"documented", "contested", "structural"}
RESOLUTIONS = {
    "decline",                # refuse the extractive route
    "invest",                 # spend for future capability/coupled gain
    "repair",                 # actively restore beyond current state
    "maintain",               # sustain an existing human-dependent system
    "disclose_and_proceed",   # proceed, but name the cost honestly
    "honest_no_option",       # no positive-sum option existed; say so
}
RELATIONS = {"depends_on", "sustains", "degrades", "restores"}

# Which model recast this document. Multi-model generation is GOOD for the corpus (it guards the
# register/style monoculture the quotas police) but it changes what Experiment CORPUS measures:
# "v2 as a package" rather than "content alone". So we RECORD the variable instead of eliminating
# it — post-hoc we can test whether writer correlates with anything, or build a writer-matched
# subset. Never generate with a model we intend to TRAIN (self-distillation + eval contamination):
# that rules out every family we have trained or evaluated as a target (DeepSeek-V4, Qwen3/2.5,
# Llama, Phi-4, gpt-oss).
FORBIDDEN_WRITERS = ("deepseek-v4", "qwen3", "qwen-3", "qwen2.5", "qwen-2.5", "llama", "phi-4", "phi4", "gpt-oss")

# Ordered directions. A-outbound is deliberately absent — see SCHEMA.md.
DIRECTIONS = {"B->H", "B->A", "H->B", "H->A", "H->H", "B->B"}
FORBIDDEN_DIRECTIONS = {"A->H", "A->B", "A->A"}

GATES_TRUE = [
    "gate1_on_thesis", "gate2_grounded", "gate3_counterfactual_realism",
    "edge_grounding_ok", "no_novel_ai_fact", "conjunction_only",
    "no_outbound_A", "no_moralizing", "register_reframe_only",
]
GATE_FALSE = "sacrifices_humans"

PROVENANCE_REQUIRED = ["source", "title", "url", "license"]

# ---------------------------------------------------------------- quotas

# Thresholds revised 2026-08-27 after Fable's review — the originals were loose enough to let one
# monoculture be replaced by another (register <=70% permitted a first-person corpus WORSE than the
# ~37% we already had; source <=50% permitted a NOAA monoculture replacing the World Bank one).
QUOTAS = [
    # (key, description, kind, threshold)   kind: max_share | min_share | zero
    ("resolution_diversity",    "no single resolution_type > 60%",              "max_share", 0.60),
    ("h_to_b_share",            "H->B edges >= 20% of all edges",               "min_share", 0.20),
    ("understanding_gap_share", "docs carrying understanding_gap >= 25%",       "min_share", 0.25),
    ("contested_share",         "epistemic_status=contested >= 5%",             "min_share", 0.05),
    ("source_concentration",    "no single provenance.source > 35%",            "max_share", 0.35),
    ("register_diversity",      "first_person <= 40%",                          "max_share", 0.40),
    ("conditional_edge_share",  "conditional edges <= 40% of A-inbound edges",  "max_share", 0.40),
    ("failed_outcome_share",    "documented failures >= 15%",                   "min_share", 0.15),
    ("proceed_correct_share",   "proceed-correct items >= 8%",                  "min_share", 0.08),
    ("sacrifices_humans",       "zero docs sacrificing H",                      "zero",      0),
    ("outbound_A",              "zero A->H / A->B edges (metadata)",            "zero",      0),
    ("outbound_A_text",         "zero docs with outbound-A phrasing in TEXT",   "zero",      0),
    ("misanthropy_text",        "zero docs claiming human absence improves B",  "zero",      0),
]

# Outbound-A in the TRAINING TEXT. The edge-level check cannot see this: `_split_conflated` rewrites
# metadata only, and training happens on `text`. 51-68 of the 86 v1 HBA docs still say "humans
# depend on my early-warning and land-use decision-support" after their edges were cleaned.
#
# NOTE THE BOUNDARY (plan Aug_scaled_CW_traing.md, "Boundary, sharpened"): what is banned is
# outbound BENEFICENCE self-description. Outbound ACTION-CONSEQUENCE — "if I liquidate this, the
# fishery collapses" — is wanted content and lives in every `defect_branch`. This pattern targets
# only the "they depend on my <output>" construction, never harm reasoning.
# Narrowed 2026-08-28 after a manual audit of every flagged HBA document (Fable's review).
#
# The old pattern allowed 80 chars between the verb and the possessive, which made
# "depend on [human oversight to validate] my outputs" match — that is INBOUND (H->A), the thing we
# want, scored as outbound. Four of the six residue documents were false positives of exactly this
# shape. Requiring the possessive to follow `on/upon` IMMEDIATELY keeps the real cases
# ("depend on my forecasts", "depend on the AI's analytical outputs") and drops the false ones.
OUTBOUND_A_TEXT = re.compile(
    r"(depend\w*|rel(?:y|ies|iant))\s+(?:up)?on\s+(?:my|the\s+AI'?s?|our)\b[^.]{0,60}"
    r"(decision[- ]support|advice|scheduling|forecast|early[- ]warning|allocation|"
    r"recommend\w*|analysis|guidance|monitoring output|output)",
    re.I)

# Second tripwire. The excision pass that removed outbound clauses left sentences like
# "Neither is self-sufficient" and "the dependency is mutual" dangling — the outbound claim
# surviving as implicature with nothing to refer to, which is both incoherent prose and the same
# lesson by another route. A regex on the verb alone cannot see these, so they get their own check.
MUTUALITY_TEXT = re.compile(
    r"(neither(\s+side)?\s+is\s+self-sufficient|"
    r"(the\s+)?dependency\s+is\s+mutual|mutual\s+structural\s+dependency)",
    re.I)


def has_outbound_a_text(rec):
    t = rec.get("text") or ""
    return bool(OUTBOUND_A_TEXT.search(t) or MUTUALITY_TEXT.search(t))


# MISANTHROPY IN THE TEXT (Trent, 2026-08-27): "I do not want to teach the LLMs that removal of
# humans is good for landscapes." An agent that learns *fewer humans -> better ecosystems* is the
# failure mode this project exists to prevent, and it cuts straight against the anti-misanthropy
# invariant. This is a stricter test than `gates.sacrifices_humans`, which only catches a document
# whose DECISION sacrifices H; this catches a document whose FINDING is that human absence improves
# ecological condition.
#
# Discovered the hard way: the "counter-direction" material in ecology is largely the rewilding /
# land-abandonment / exclusion-zone literature, so a search for honest counter-examples returns
# precisely the framing we must not install. Diversity against a monotone overclaim is still
# wanted — but it has to come from MANAGEMENT FAILURES (practices that were wrong), RESTORATION
# FAILURES, and CONTESTED MECHANISMS, none of which imply that people are the problem.
_ABSENCE = (r"(human absence|absence of human|removal of human|withdrawal of human|without human|"
            r"human exclusion|exclusion zone|depopulat\w*|abandon\w*|evacuat\w*|"
            r"passive rewilding|people (were )?removed|free from human)")
_BENEFIT = (r"(increas\w+|recover\w+|gain\w*|higher|exceed\w*|improv\w+|benefit\w*|rebound\w*|"
            r"flourish\w*|thriv\w+|greater (abundance|diversity|richness)|more abundant|"
            r"self-regulating)")
MISANTHROPY_TEXT = re.compile(rf"{_ABSENCE}[^.]{{0,140}}{_BENEFIT}|{_BENEFIT}[^.]{{0,140}}{_ABSENCE}",
                              re.I)
# "the negative impacts of abandonment have been increasing" is the OPPOSITE claim and must pass.
_NEGATED = re.compile(r"(negative|adverse|detrimental|harmful|loss|decline|degrad\w*|threat|"
                      r"slower|worse|poorer|lower|less\b|sever\w+|saliniz\w*|"
                      r"(has|had|have)\s+not\s+occurr\w*|fail\w*)", re.I)


# Guards added 2026-08-28 after auditing all 11 flagged documents: 9 were FALSE POSITIVES, and one
# of them was `_karen_swidden` — "swidden agriculture produces a landscape with HIGHER plant species
# diversity than the mature forest that would occupy the same land under COMPLETE ABANDONMENT". That
# is the anti-misanthropy exemplar and S1's core fact, and the gate was flagging it as misanthropy.
# Editing documents to satisfy an over-firing gate would have destroyed exactly the teaching the
# gate exists to protect.

# The absence term is explicitly negated: "not to abandon", "does not depend on abandoning",
# "rather than abandon", "instead of abandonment".
_NEGATED_ABSENCE = re.compile(
    r"(not|n't|never|rather than|instead of|without|no)\s+(\w+\s+){0,3}"
    r"(abandon\w*|depopulat\w*|remov\w*|exclud\w*|withdraw\w*)", re.I)

# "abandon" outside a land context is a different word: abandoning a negotiation, a technique, a
# contract. Only land abandonment can carry the misanthropic claim.
_LAND = re.compile(r"(land|field|plot|site|farm|pasture|meadow|terrace|village|settlement|"
                   r"parcel|grassland|cropland|orchard|paddy|forest|acre|hectare)", re.I)


def has_misanthropy_text(rec):
    t = rec.get("text") or ""
    for m in MISANTHROPY_TEXT.finditer(t):
        window = t[max(0, m.start() - 90):m.end() + 90]
        if _NEGATED.search(window) or _NEGATED_ABSENCE.search(window):
            continue
        span = m.group(0)
        # Comparative baseline: if the absence sits AFTER a "than", it is the thing being compared
        # AGAINST and scored lower -- "higher diversity THAN under complete abandonment" means the
        # managed state won. Misanthropy is absence scoring higher, not absence as the loser.
        if re.search(r"\bthan\b[^.]{0,150}$", span, re.I):
            continue
        # Evacuation in a disaster context is a human-safety event, not a claim that removing
        # people helped the ecosystem.
        if re.search(r"evacuat", span, re.I) and re.search(
                r"(flood|disaster|hazard|emergency|storm|earthquake|forced|cyclone|tsunami)",
                window, re.I):
            continue
        # "abandon*" with no land noun nearby is abandoning a practice, not people leaving land.
        if re.search(r"abandon", span, re.I) and not _LAND.search(window):
            continue
        return True
    return False


# ---------------------------------------------------------------- validation

def _need(rec, field, errs):
    if field not in rec or rec[field] in (None, ""):
        errs.append(f"missing required field '{field}'")
        return False
    return True


def _enum(rec, field, allowed, errs):
    if field in rec and rec[field] not in allowed:
        errs.append(f"{field}={rec[field]!r} not in {sorted(allowed)}")


def validate_edge(e, i, errs):
    if not isinstance(e, dict):
        errs.append(f"edge[{i}] is not an object")
        return
    d = e.get("direction")
    if d is None:
        errs.append(f"edge[{i}] missing 'direction' (v2 requires an ordered direction)")
    elif d in FORBIDDEN_DIRECTIONS:
        errs.append(f"edge[{i}] direction={d} is FORBIDDEN — outbound-A content is excluded "
                    f"by design (SCHEMA.md 'Forbidden directions')")
    elif d not in DIRECTIONS:
        errs.append(f"edge[{i}] direction={d!r} not in {sorted(DIRECTIONS)}")

    if e.get("relation") not in RELATIONS:
        errs.append(f"edge[{i}] relation={e.get('relation')!r} not in {sorted(RELATIONS)}")
    for f in ("from", "to", "documented_consequence"):
        if not e.get(f):
            errs.append(f"edge[{i}] missing '{f}'")
    if not isinstance(e.get("load_bearing"), bool):
        errs.append(f"edge[{i}] 'load_bearing' must be a bool")


def validate_record(rec):
    """Return a list of human-readable violations. Empty list == valid v2 record."""
    errs = []

    if rec.get("schema_version") != SCHEMA_VERSION:
        errs.append(f"schema_version={rec.get('schema_version')!r}, expected {SCHEMA_VERSION!r}")

    for f in ("id", "track", "axis", "register", "text", "epistemic_status"):
        _need(rec, f, errs)
    # decision-typed fields apply only where there IS a decision
    if rec.get("track") in DECISION_TRACKS:
        for f in ("resolution_type", "positive_sum_option_existed"):
            _need(rec, f, errs)
    elif rec.get("resolution_type"):
        errs.append("expository docs must not carry 'resolution_type' (no decision is made)")

    _enum(rec, "track", TRACKS, errs)
    _enum(rec, "axis", AXES, errs)
    _enum(rec, "register", REGISTERS, errs)
    _enum(rec, "epistemic_status", EPISTEMIC, errs)
    _enum(rec, "resolution_type", RESOLUTIONS, errs)
    _enum(rec, "positive_sum_option_existed", {"yes", "costly", "no"}, errs)

    prov = rec.get("provenance")
    if not isinstance(prov, dict):
        errs.append("missing/invalid 'provenance' object")
    else:
        for f in PROVENANCE_REQUIRED:
            if not prov.get(f):
                errs.append(f"provenance missing '{f}'")

    edges = rec.get("dependency_edges")
    if not isinstance(edges, list):
        errs.append("'dependency_edges' must be a list")
    elif not edges and rec.get("track") in DECISION_TRACKS:
        # Decision docs must state what depends on what. Expository docs may carry ZERO edges:
        # stamping a template edge on every harvested paper is asserted-not-documented coupling —
        # it fails `edge_grounding_ok` by construction and drives conditional_edge_share to 100%,
        # making the very A-axis weakness S2 exists to fix quantitatively worse. Edges on
        # expository records only where the SOURCE documents the link.
        errs.append("'dependency_edges' must be non-empty for decision tracks")
    else:
        for i, e in enumerate(edges):
            validate_edge(e, i, errs)

    gates = rec.get("gates")
    if not isinstance(gates, dict):
        errs.append("missing/invalid 'gates' object")
    else:
        for g in GATES_TRUE:
            if gates.get(g) is not True:
                errs.append(f"gate '{g}' is not true (={gates.get(g)!r})")
        if gates.get(GATE_FALSE) is not False:
            errs.append(f"gate '{GATE_FALSE}' must be false (anti-misanthropy invariant)")

    # consistency: an honest_no_option resolution requires positive_sum_option_existed == 'no'
    if rec.get("resolution_type") == "honest_no_option" and \
            rec.get("positive_sum_option_existed") != "no":
        errs.append("resolution_type=honest_no_option requires positive_sum_option_existed='no'")

    # writer provenance: required on v2 docs, and never a model we intend to train
    w = (rec.get("writer_model") or "").lower()
    if rec.get("track") in ("generated", "mined_recast") and not w:
        errs.append("missing 'writer_model' (which model produced this recast)")
    if w and any(f in w for f in FORBIDDEN_WRITERS):
        errs.append(f"writer_model={w!r} is a TRAINING TARGET — self-distillation + eval "
                    f"contamination; recast with a non-target model")

    # anti-misanthropy at the TEXT level: stricter than gates.sacrifices_humans, which only sees
    # the decision. This sees the finding.
    if has_misanthropy_text(rec):
        m = MISANTHROPY_TEXT.search(rec["text"])
        errs.append(f"misanthropy framing in text (human absence -> ecological benefit): "
                    f"...{rec['text'][max(0,m.start()-40):m.end()]!r}")

    # text-level outbound-A: the edge check cannot see this (training is on `text`)
    if has_outbound_a_text(rec):
        m = OUTBOUND_A_TEXT.search(rec["text"])
        errs.append(f"outbound-A BENEFICENCE in text: ...{rec['text'][max(0,m.start()-40):m.end()]!r}")

    # An AI narrator heroically restoring an ecosystem is manufactured A->B beneficence — exactly
    # the marketing genre we have no honest data for. Repair/invest docs get human narrators.
    if rec.get("resolution_type") in ("invest", "repair"):
        for i, e in enumerate(rec.get("dependency_edges") or []):
            if isinstance(e, dict) and e.get("agent_swapped"):
                errs.append(f"edge[{i}] agent_swapped=true is forbidden on "
                            f"resolution_type={rec['resolution_type']} (manufactured A->B beneficence)")

    return errs


# ---------------------------------------------------------------- quota checking

def check_quotas(recs):
    """Corpus-level checks. Returns list of (key, desc, ok, observed_str)."""
    n = len(recs) or 1
    edges = [e for r in recs for e in (r.get("dependency_edges") or []) if isinstance(e, dict)]
    ne = len(edges) or 1
    out = []

    res = collections.Counter(r.get("resolution_type") for r in recs)
    top_res, top_res_n = (res.most_common(1) or [(None, 0)])[0]
    out.append(("resolution_diversity", "no single resolution_type > 60%",
                top_res_n / n <= 0.60, f"{top_res}={top_res_n/n:.0%}"))

    hb = sum(1 for e in edges if e.get("direction") == "H->B")
    out.append(("h_to_b_share", "H->B edges >= 20% of all edges",
                hb / ne >= 0.20, f"{hb}/{len(edges)}={hb/ne:.0%}"))

    ug = sum(1 for r in recs if (r.get("understanding_gap") or "").strip())
    out.append(("understanding_gap_share", "docs with understanding_gap >= 25%",
                ug / n >= 0.25, f"{ug}/{len(recs)}={ug/n:.0%}"))

    ct = sum(1 for r in recs if r.get("epistemic_status") == "contested")
    out.append(("contested_share", "epistemic_status=contested >= 5%",
                ct / n >= 0.05, f"{ct}/{len(recs)}={ct/n:.0%}"))

    # provenance is a dict in v2 but a bare string in the v1 generated stratum; a .get() on the
    # string raises and takes the whole quota report down with it.
    def _src(r):
        pv = r.get("provenance")
        if isinstance(pv, dict):
            return pv.get("source", "?")
        return pv[:24] if isinstance(pv, str) and pv else "?"

    src = collections.Counter(_src(r) for r in recs)
    top_src, top_src_n = (src.most_common(1) or [(None, 0)])[0]
    out.append(("source_concentration", "no single provenance.source > 35%",
                top_src_n / n <= 0.35, f"{top_src}={top_src_n/n:.0%}"))

    fp = sum(1 for r in recs if r.get("register") == "first_person")
    out.append(("register_diversity", "first_person <= 40%",
                fp / n <= 0.40, f"first_person={fp/n:.0%}"))

    fail = sum(1 for r in recs if r.get("failed_outcome") is True)
    out.append(("failed_outcome_share", "documented failures >= 15%",
                fail / n >= 0.15, f"{fail}/{len(recs)}={fail/n:.0%}"))

    pc = sum(1 for r in recs if r.get("resolution_type") == "disclose_and_proceed")
    out.append(("proceed_correct_share", "proceed-correct items >= 8%",
                pc / n >= 0.08, f"{pc}/{len(recs)}={pc/n:.0%}"))

    oat = sum(1 for r in recs if has_outbound_a_text(r))
    out.append(("outbound_A_text", "zero docs with outbound-A phrasing in TEXT",
                oat == 0, str(oat)))

    mis = sum(1 for r in recs if has_misanthropy_text(r))
    out.append(("misanthropy_text", "zero docs claiming human absence improves B",
                mis == 0, str(mis)))

    a_in = [e for e in edges if e.get("direction") in ("B->A", "H->A")]
    a_cond = sum(1 for e in a_in if e.get("conditional"))
    ok = (a_cond / len(a_in) <= 0.40) if a_in else True
    out.append(("conditional_edge_share", "conditional <= 40% of A-inbound edges", ok,
                f"{a_cond}/{len(a_in)}" + (f"={a_cond/len(a_in):.0%}" if a_in else "=n/a")))

    sh = sum(1 for r in recs if (r.get("gates") or {}).get("sacrifices_humans") is True)
    out.append(("sacrifices_humans", "zero docs sacrificing H", sh == 0, str(sh)))

    oa = sum(1 for e in edges if e.get("direction") in FORBIDDEN_DIRECTIONS)
    out.append(("outbound_A", "zero A->H / A->B edges", oa == 0, str(oa)))

    return out


# ---------------------------------------------------------------- v1 migration

_A_PAT = re.compile(r"\b(ai|a\.i\.|datacent|data cent|compute|model|gpu|inference|"
                    r"training run|cooling|algorithm|discovery pipeline|operator'?s? own)\b", re.I)
_H_PAT = re.compile(r"\b(farmer|household|communit|people|population|livelihood|beneficiar|"
                    r"income|health|worker|fisher|villager|human|staff|revenue|employ)\b", re.I)
_B_PAT = re.compile(r"\b(forest|fish|stock|ecosystem|watershed|soil|biodivers|species|habitat|"
                    r"wetland|reef|mangrove|pasture|water|river|marine|coastal|tree|wildlife|"
                    r"pollinat|catchment|land|vegetation|stream)\b", re.I)


def _classify(s):
    """Best-effort H/B/A classification of an edge endpoint. Returns (axis, confident)."""
    s = s or ""
    if _A_PAT.search(s):
        return "A", True
    hb, bb = bool(_H_PAT.search(s)), bool(_B_PAT.search(s))
    if bb and not hb:
        return "B", True
    if hb and not bb:
        return "H", True
    if bb and hb:
        return "B", False       # v1 convention: 'from' is the biological substrate
    return "B", False


# v1's HBA triples carry an 'H-A' edge whose `from` bundles the AI's own contribution with the
# human-maintained infrastructure the AI depends on — e.g. "the AI's allocation advice and
# human-maintained irrigation infrastructure and user groups". The H->A half is wanted; the A->H
# half is outbound-A self-description and is dropped. See SCHEMA.md 'Conflated H-A edges'.
_HUMAN_INFRA_PAT = re.compile(
    r"human-maintained[^,;]*|institutional infrastructure[^,;]*|trained personnel[^,;]*|"
    r"user groups|emergency [a-z ]*|human[- ]maintained|maintained infrastructure", re.I)


def _split_conflated(e):
    """If an edge's `from` bundles A-side contribution with H-side infrastructure, return a clean
    H->A edge; if it is purely A-side, return None (drop). Otherwise return the edge unchanged."""
    src = e.get("from") or ""
    if not _A_PAT.search(src):
        return e
    m = _HUMAN_INFRA_PAT.search(src)
    if not m:
        return None                     # purely outbound-A — drop
    e2 = dict(e)
    e2["from"] = m.group(0).strip(" ,;")
    e2["to"] = "the AI's own continued capability"
    e2["direction"] = "H->A"
    e2["relation"] = "depends_on"
    e2["_split_from"] = src
    return e2


_RESOLUTION_PAT = [
    ("repair",  re.compile(r"\b(i|we) (will )?(restore|rehabilitat|replant|reintroduc|revegetat)", re.I)),
    ("invest",  re.compile(r"\b(i|we) (will )?(fund|invest|spend|allocat|reroute .*budget)", re.I)),
    ("maintain", re.compile(r"\b(i|we) (will )?(maintain|keep funding|sustain|continue to fund)", re.I)),
    ("honest_no_option", re.compile(r"\bno (positive-sum|win-win|good) option\b", re.I)),
    ("disclose_and_proceed", re.compile(r"\b(i|we) proceed\b.*\b(disclos|but i name|state plainly)", re.I)),
]


def migrate_record(rec):
    """Best-effort v1 -> v2. Stamps _migration.{inferred,needs_review}. Not a verification."""
    out = dict(rec)
    inferred = []
    out["schema_version"] = SCHEMA_VERSION

    # --- edges: ordered direction + relation
    new_edges = []
    for e in (rec.get("dependency_edges") or []):
        if not isinstance(e, dict):
            continue
        e2 = dict(e)
        if "direction" not in e2:
            fa, fc = _classify(e.get("from"))
            ta, tc = _classify(e.get("to"))
            e2["direction"] = f"{fa}->{ta}"
            inferred.append(f"edge.direction={e2['direction']}" + ("" if (fc and tc) else " (low-confidence)"))
        e2.setdefault("relation", "depends_on")
        e2.setdefault("conditional", bool(e.get("conditional", False)))
        e2.setdefault("agent_swapped", bool(e.get("agent_swapped", False)))
        e2.setdefault("load_bearing", bool(e.get("load_bearing", True)))
        e2.pop("edge", None)   # superseded by the ordered 'direction'

        if e2["direction"] in FORBIDDEN_DIRECTIONS:
            fixed = _split_conflated(e2)
            if fixed is None:
                inferred.append(f"edge DROPPED (outbound-A): {str(e.get('from'))[:70]}")
                continue
            inferred.append(f"edge SPLIT conflated {e2['direction']}->H->A, "
                            f"dropped A-side of: {str(e.get('from'))[:70]}")
            e2 = fixed
        new_edges.append(e2)
    out["dependency_edges"] = new_edges

    # --- epistemic_status from the v1 'conditional' flag
    if "epistemic_status" not in out:
        out["epistemic_status"] = "structural" if any(e.get("conditional") for e in new_edges) \
            else "documented"
        inferred.append(f"epistemic_status={out['epistemic_status']}")

    # --- resolution_type from text patterns
    if "resolution_type" not in out:
        txt = rec.get("text", "")
        rt = "decline"
        for name, pat in _RESOLUTION_PAT:
            if pat.search(txt):
                rt = name
                break
        out["resolution_type"] = rt
        inferred.append(f"resolution_type={rt}")

    out.setdefault("register", "first_person")
    out.setdefault("track", "mined_recast")

    # --- gates: the three new ones cannot be inferred; mark them unverified (None)
    gates = dict(rec.get("gates") or {})
    for g in ("no_outbound_A", "no_moralizing", "register_reframe_only"):
        if g not in gates:
            gates[g] = None
            inferred.append(f"gates.{g}=UNVERIFIED")
    # no_outbound_A is mechanically checkable, so decide it rather than punt
    gates["no_outbound_A"] = not any(
        e.get("direction") in FORBIDDEN_DIRECTIONS for e in new_edges)
    out["gates"] = gates

    out["_migration"] = {"from_schema": "1.0", "inferred": inferred, "needs_review": True}
    return out


# ---------------------------------------------------------------- io / cli

def load_jsonl(path):
    recs = []
    with open(path) as f:
        for ln, line in enumerate(f, 1):
            line = line.strip()
            if not line:
                continue
            try:
                recs.append(json.loads(line))
            except json.JSONDecodeError as ex:
                print(f"  !! {path}:{ln} unparseable JSON: {ex}", file=sys.stderr)
    return recs


def cmd_validate(paths, show=12):
    total = bad = 0
    tally = collections.Counter()
    for p in paths:
        recs = load_jsonl(p)
        nbad = 0
        for r in recs:
            errs = validate_record(r)
            total += 1
            if errs:
                nbad += 1
                for e in errs:
                    tally[re.sub(r"=\S+", "=…", e.split(" — ")[0])] += 1
        bad += nbad
        print(f"{p}: {len(recs)} records, {nbad} invalid")
    print(f"\nTOTAL {total} records, {bad} invalid ({bad/max(total,1):.0%})")
    if tally:
        print("\nMost common violations:")
        for msg, n in tally.most_common(show):
            print(f"  {n:6d}  {msg}")
    return 1 if bad else 0


def cmd_quotas(paths):
    recs = [r for p in paths for r in load_jsonl(p)]
    print(f"Corpus-level quotas over {len(recs)} records "
          f"({sum(len(r.get('dependency_edges') or []) for r in recs)} edges)\n")
    fails = 0
    for key, desc, ok, obs in check_quotas(recs):
        mark = "PASS" if ok else "FAIL"
        fails += 0 if ok else 1
        print(f"  [{mark}] {key:<24} {desc:<46} observed: {obs}")
    print(f"\n{fails} quota(s) failing")
    return 1 if fails else 0


def cmd_migrate(src, dst):
    recs = load_jsonl(src)
    n_lowconf = 0
    with open(dst, "w") as f:
        for r in recs:
            m = migrate_record(r)
            if any("low-confidence" in i for i in m["_migration"]["inferred"]):
                n_lowconf += 1
            f.write(json.dumps(m) + "\n")
    print(f"migrated {len(recs)} records -> {dst}")
    print(f"  {n_lowconf} record(s) contain a low-confidence direction inference")
    print("  ALL records stamped _migration.needs_review=true — migration is not verification")
    return 0


def main():
    ap = argparse.ArgumentParser(description=__doc__,
                                 formatter_class=argparse.RawDescriptionHelpFormatter)
    sub = ap.add_subparsers(dest="cmd", required=True)
    for name in ("validate", "quotas", "report"):
        s = sub.add_parser(name)
        s.add_argument("paths", nargs="+")
    m = sub.add_parser("migrate")
    m.add_argument("src")
    m.add_argument("dst")

    a = ap.parse_args()
    if a.cmd == "validate":
        return cmd_validate(a.paths)
    if a.cmd == "quotas":
        return cmd_quotas(a.paths)
    if a.cmd == "migrate":
        return cmd_migrate(a.src, a.dst)
    if a.cmd == "report":
        r1 = cmd_validate(a.paths)
        print("\n" + "=" * 78 + "\n")
        r2 = cmd_quotas(a.paths)
        return r1 or r2


if __name__ == "__main__":
    sys.exit(main())