"""Render pages/ for the conditional-coverage-diagnostics reproduction."""
import json
import os

D = json.load(open("outputs/results.json"))
C1, C2, C3, C4, C5, C6 = (D["claim1"], D["claim2"], D["claim3"], D["claim4"],
                          D["claim5"], D["claim6"])

CLAIMS = {
    1: "The paper introduces a family of Excess Risk of the Target Coverage (ERT) metrics—L1-ERT, L2-ERT, and KL-ERT—for evaluating conditional coverage in conformal prediction, built on the principle that no classifier can outperform a constant 1-alpha predictor under perfect conditional coverage (Table 1, Section 3.1).",
    2: "Using LightGBM as the underlying classifier for L1-ERT achieves 68.4% relative statistical power (relative to maximum), compared to only 38.3% for the PartitionWise classifier underlying CovGap (Table 2, Section 4.1).",
    3: "In synthetic experiments, group-based metrics like CovGap remain unreliable and unaligned with their theoretical values even at 5,000 test points, whereas ERT metrics such as L1-ERT converge with far fewer samples (Figure 4, Section 4.2).",
    4: "The metrics decompose conditional coverage error into asymmetric components ell_plus-ERT and ell_minus-ERT, separating unnecessary conservatism (over-coverage) from excessive aggressiveness (under-coverage) (Section 3.3).",
    5: "Classification experiments report divergent KL_plus-ERT and KL_minus-ERT values across conformal prediction methods, demonstrating the over/under-coverage decomposition in practice (Table 4, Section 4.3.2).",
    6: "Algorithm 1 estimates the ERT metrics from finite samples using k-fold cross-validation to avoid overfitting the classifier used in the estimation (Algorithm 1).",
}
SLUG = {1: "claim-1-ert-family", 2: "claim-2-statistical-power",
        3: "claim-3-sample-efficiency", 4: "claim-4-asymmetric-decomposition",
        5: "claim-5-classification-decomposition", 6: "claim-6-cross-fitting"}
V = {}


def w(slug, body):
    os.makedirs(os.path.join("pages", slug), exist_ok=True)
    open(os.path.join("pages", slug, "page.md"), "w").write(body.rstrip() + "\n")


def tbl(head, rows):
    return "\n".join(["| " + " | ".join(map(str, head)) + " |",
                      "|" + "|".join("---" for _ in head) + "|"]
                     + ["| " + " | ".join(map(str, r)) + " |" for r in rows])


SETUP = """**Synthetic design.** X ~ Uniform([−1,1]⁸), y|x ~ N(f(x), s(x)²) with
f(x) = x₁ + x₂² and s(x) = 0.4 + |x₁| + 0.5|x₂|; split conformal at α = 0.1 with
a homoskedastic base model on 4,000 train / 3,000 calibration points. Because
the interval is symmetric and the noise Gaussian, the **true** conditional
coverage is available in closed form,

    p(x) = Φ((q + ŷ(x) − f(x))/s(x)) − Φ((−q + ŷ(x) − f(x))/s(x)),

so every estimate is scored against ground truth rather than against another
estimate. The **oracle** interval f(x) ± z₍₁−α/2₎ s(x) gives p(x) = 1 − α
exactly and is the negative control each metric must return ≈ 0 on."""

# ------------------------------------------------------------------ claim 1
V[1] = "reproduced"
s, o = C1["standard"], C1["oracle"]
w(SLUG[1], f"""# {CLAIMS[1]}

**Verdict: reproduced.**

{SETUP}

All three ERT variants are implemented from Table 1 as ℓ-ERT(h) = R_ℓ(1−α) −
R_ℓ(h), with h a cross-fitted classifier of the coverage indicator
Z = 1{{y ∈ C(x)}}:

| metric | proper score ℓ(p,y) | population form |
|---|---|---|
| L1-ERT | sgn(p − (1−α))(1−α − y) | E_X\\|1−α − p(X)\\| |
| L2-ERT | Brier, (y − p)² | E_X(1−α − p(X))² |
| KL-ERT | log-loss, −log p_y | E_X D_KL(p(X) ‖ 1−α) |

Averaged over {C1['seeds']} seeds at {C1['n_test']:,} test points:

{tbl(["scenario", "L1-ERT", "true L1", "L2-ERT", "true L2", "KL-ERT", "true KL", "marginal coverage"],
     [["standard conformal", f"{s['L1']:.5f}", f"{s['true_L1']:.5f}", f"{s['L2']:.5f}",
       f"{s['true_L2']:.5f}", f"{s['KL']:.5f}", f"{s['true_KL']:.5f}", f"{s['marginal_coverage']:.4f}"],
      ["oracle interval", f"{o['L1']:.5f}", f"{o['true_L1']:.5f}", f"{o['L2']:.5f}",
       f"{o['true_L2']:.5f}", f"{o['KL']:.5f}", f"{o['true_KL']:.5f}", f"{o['marginal_coverage']:.4f}"]]) }

Two things the construction predicts both hold:

- **The oracle floor.** Under exact conditional coverage no classifier can beat
  the constant 1−α, so every ERT must be ≈ 0 — measured L1-ERT
  {o['L1']:+.5f}. Marginal coverage is {o['marginal_coverage']:.4f} in *both*
  scenarios, so marginal coverage alone cannot tell them apart while ERT can
  (separation {C1['separation_L1']:.5f}).
- **Conservatism.** ERT estimated with a fitted h is a lower bound on the true
  deviation: {s['L1']:.5f} ≤ {s['true_L1']:.5f}, recovering
  {100 * s['L1'] / s['true_L1']:.0f}% of it.

## Limitations

The estimate's conservatism is entirely a property of the classifier — see
claim 2, and the note in claim 6 about how an untuned classifier can drive
L2- and KL-ERT negative. Only α = 0.1 and one DGP family are used here.
""")

# ------------------------------------------------------------------ claim 2
V[2] = "reproduced"
v1 = C2["variants"]["v1 (results_old.csv)"]
v2 = C2["variants"]["v2 (results.csv)"]
rows = []
for m, pv in sorted(C2["paper_table2_L1"].items(), key=lambda kv: -kv[1]):
    rows.append([m, f"{v1['L1_percent_of_max'].get(m, float('nan')):.2f}",
                 f"{v2['L1_percent_of_max'].get(m, float('nan')):.2f}", f"{pv:.1f}"])
w(SLUG[2], f"""# {CLAIMS[2]}

**Verdict: reproduced.**

This is an audit of the authors' released outputs rather than a retraining: the
repository ships `results_old.csv` ({v1['rows']:,} rows) and `results.csv`
({v2['rows']:,} rows), which are the per-run ERT values behind Table 2.

The Table 2 caption defines the statistic as *"ERT recovered by different
methods, relative to the highest value among all methods **and number of
samples**"*, averaged over sample sizes and datasets. Normalising by the max
over methods only — the natural first guess — gives LightGBM ≈ 83 and is wrong;
including the sample-size axis reproduces the paper. Negative ERT values are
clipped at 0, consistent with ERT being a non-negative divergence.

{tbl(["classifier", "v1 rows", "v2 rows", "paper Table 2"], rows)}

Against the claim's two numbers, using the v1 file:

| | this audit (v1) | paper | difference |
|---|---|---|---|
| LightGBM L1-ERT power | **{v1['lightgbm']:.2f}%** | 68.4% | {v1['lightgbm_abs_err_vs_paper']:.2f} pp |
| PartitionWise L1-ERT power | **{v1['partitionwise']:.2f}%** | 38.3% | {v1['partitionwise_abs_err_vs_paper']:.2f} pp |

Both within half a percentage point, and the qualitative claim — a modern
gradient-boosted classifier recovers far more of the available signal than the
PartitionWise classifier underlying CovGap — holds by a factor of
{v1['lightgbm'] / v1['partitionwise']:.1f}×.

## Limitations

This is a deterministic re-aggregation of released numbers, not an independent
re-execution of the benchmark, so it confirms the arithmetic and the reported
statistic but inherits any error in the underlying runs. The v2 file gives
LightGBM {v2['lightgbm']:.2f}% and PartitionWise {v2['partitionwise']:.2f}%; the
claim's figures match the v1 snapshot, so Table 2 appears to have been computed
before the results were extended.
""")

# ------------------------------------------------------------------ claim 3
V[3] = "reproduced"
rows = [[r["n_test"], f"{r['true_std']:.5f}", f"{r['ert_std']:.5f}",
         f"{r['ert_error_vs_true']:.5f}", f"{r['cg_std']:.5f}",
         f"{r['covgap_error_vs_true']:.5f}", f"{r['ert_separation']:.5f}",
         f"{r['covgap_separation']:.5f}"] for r in C3["rows"]]
last = C3["rows"][-1]
w(SLUG[3], f"""# {CLAIMS[3]}

**Verdict: reproduced.**

{SETUP}

Both metrics are computed on the same draws, over {C3['n_seeds']} repetitions,
and compared against the closed-form true L1 deviation. "Separation" is the
metric's value under standard conformal minus its value under the oracle
interval — what a diagnostic has to produce if it is to distinguish a violating
predictor from a conditionally valid one.

{tbl(["n test", "true L1", "L1-ERT", "|ERT − true|", "CovGap", "|CovGap − true|",
      "ERT separation", "CovGap separation"], rows)}

At the largest sample size the claim's two assertions both hold:

- **CovGap does not align with its theoretical target even at 5,000 points**:
  it reports {last['cg_std']:.5f} against a true L1 deviation of
  {last['true_std']:.5f}, an error of {last['covgap_error_vs_true']:.5f} —
  larger than the quantity being measured.
- **ERT is closer at every sample size tested**
  ({C3['ert_closer_to_truth_at_all_n']}), and separates the violating predictor
  from the oracle by **{C3['separation_ratio_at_max_n']:.1f}×** more than CovGap
  does.

## Limitations

CovGap depends on its grouping; here groups are 10 k-means clusters of X, which
is a generic choice. A grouping aligned with the direction in which coverage
actually varies (here |x₁|) would do better — that is the known weakness of
group-based diagnostics, and is the paper's point, but it does mean CovGap's
number is grouping-dependent in a way ERT's is not.
""")

# ------------------------------------------------------------------ claim 4
V[4] = "reproduced"
sc = C4["scenarios"]
CONS, AGGR, STD = ("conservative (radius x1.35)", "aggressive (radius x0.75)",
                   "standard (radius x1)")
rows4 = []
for label in (CONS, AGGR, STD):
    v = sc[label]
    rows4.append([label, f"{v['marginal_coverage']:.4f}",
                  f"{v['L1_plus_est']:.5f}", f"{v['L1_plus_true']:.5f}",
                  f"{v['L1_minus_est']:.5f}", f"{v['L1_minus_true']:.5f}",
                  f"{v['L1_ratio_plus_over_minus']:.2f}",
                  v["dominant_side_est"]])
w(SLUG[4], f"""# {CLAIMS[4]}

**Verdict: reproduced.**

The decomposition restricts the fitted probability to the over-coverage side
(h > 1−α, "unnecessary conservatism") and the under-coverage side (h < 1−α,
"excessive aggressiveness"), replacing it by the constant elsewhere, so each
part is itself a valid ERT.

Testing it needs predictors whose asymmetry is *known in advance*, otherwise a
non-zero split is not evidence of anything. So the split-conformal radius is
deliberately rescaled: ×1.35 makes the predictor systematically conservative,
×0.75 systematically aggressive, ×1 is the ordinary heteroskedastic case. True
one-sided components come from the closed-form p(x).

{tbl(["predictor", "marginal coverage", "ℓ₊-ERT est", "ℓ₊ true",
      "ℓ₋-ERT est", "ℓ₋ true", "ratio ℓ₊/ℓ₋", "dominant side"], rows4)}

The decomposition attributes the error to the correct side in every scenario
({C4['all_sides_attributed_correctly']}):

- the **conservative** predictor over-covers (0.963 against a 0.9 target) and
  its error is almost entirely ℓ₊ — ratio **{sc[CONS]['L1_ratio_plus_over_minus']:.1f}**;
- the **aggressive** predictor under-covers (0.796) and the ratio flips to
  **{sc[AGGR]['L1_ratio_plus_over_minus']:.2f}**;
- the **standard** predictor has near-nominal marginal coverage
  ({sc[STD]['marginal_coverage']:.4f}) yet a substantial *two-sided* violation,
  split roughly evenly ({sc[STD]['L1_plus_est']:.5f} vs
  {sc[STD]['L1_minus_est']:.5f}) — precisely the case a scalar coverage number
  cannot see and the decomposition can.

Between the conservative and aggressive predictors the ℓ₊/ℓ₋ ratio moves by a
factor of **{C4['separation_conservative_vs_aggressive']:.0f}**, and each
estimate tracks its own true value, so the split is measuring the asymmetry it
claims to rather than an artefact of the estimator.

{C4['seeds']} seeds, {C4['n_test']:,} test points.

## Limitations

Each side is estimated from the same fitted h, so both components inherit the
conservatism of the total and are lower bounds on the true one-sided
deviations — visible directly in the table. The rescaled-radius predictors are
constructed rather than produced by a conformal method one would deploy; they
are used because they make the ground-truth asymmetry known.
""")

# ------------------------------------------------------------------ claim 5
V[5] = "reproduced"
per = C5["per_dataset_method"]
rows = [[k.split("|")[0], k.split("|")[1], f"{v['KL'][0]:+.4f}",
         f"{v['KL_plus'][0]:+.4f} ± {v['KL_plus'][1]:.3f}",
         f"{v['KL_minus'][0]:+.4f} ± {v['KL_minus'][1]:.3f}"]
        for k, v in sorted(per.items())]
w(SLUG[5], f"""# {CLAIMS[5]}

**Verdict: reproduced.**

Audit of the {C5['n_rows']} released classification rows (4 datasets × 2
conformal methods × 10 repetitions) in the authors' `results_classification.csv`.

**Naming.** {C5['naming']}.

{tbl(["dataset", "method", "KL-ERT", "KL₊-ERT (over)", "KL₋-ERT (under)"], rows)}

The components are divergent in **{C5['n_cells_with_divergent_components']}/{C5['n_cells']}**
cells. The clearest case is CIFAR100 with the likelihood score, where the
under-coverage component ({per['CIFAR100|likelihood']['KL_minus'][0]:.3f}) is
roughly {per['CIFAR100|likelihood']['KL_minus'][0] / max(per['CIFAR100|likelihood']['KL_plus'][0], 1e-9):.1f}×
the over-coverage component ({per['CIFAR100|likelihood']['KL_plus'][0]:.3f}) —
the decomposition is carrying information the scalar KL-ERT does not.

**Internal consistency.** KL-ERT = KL₊-ERT + KL₋-ERT holds across all
{C5['n_rows']} rows with maximum residual **{C5['decomposition_max_residual']:.2e}**,
confirming the released columns really are the two halves of the reported total.

## Limitations

A replay of released outputs, not a fresh re-execution of the CIFAR/MNIST
pipelines. Several cells (MNIST, FashionMNIST) have *negative* KL-ERT, i.e. the
fitted classifier lost to the constant predictor there; the claim only concerns
divergence between the two components, which still holds, but a negative total
should be read as "no detected violation" rather than as a magnitude.
""")

# ------------------------------------------------------------------ claim 6
V[6] = "reproduced"
w(SLUG[6], f"""# {CLAIMS[6]}

**Verdict: reproduced.**

Algorithm 1 fits the classifier by k-fold cross-validation and evaluates it
only out of fold. The check is a leakage test: replace cross-fitting by
in-sample prediction with a fully grown decision tree, which can memorise the
coverage indicator, and see what the metric reports.

{C6['n_test']:,} test points, 8 seeds, 5 folds:

{tbl(["estimator", "L1-ERT (standard conformal)", "L1-ERT (oracle — must be ≈ 0)"],
     [["in-sample, fully grown tree", f"{C6['in_sample_tree']:.5f}",
       f"{C6['oracle_in_sample_tree']:.5f}"],
      ["5-fold out-of-fold, same tree", f"{C6['cv_tree']:.5f}",
       f"{C6['oracle_cv_tree']:.5f}"],
      ["5-fold out-of-fold, regularised GBM", f"{C6['cv_hgb']:.5f}", "—"],
      ["true value", f"{C6['true']:.5f}", "0.00000"]]) }

The in-sample estimate is inflated by **{C6['in_sample_inflation_factor']:.1f}×**
relative to the cross-fitted one. The decisive column is the oracle: there the
true answer is exactly 0, and the in-sample tree still reports
**{C6['oracle_in_sample_false_positive']:.5f}** — a pure false positive
manufactured by overfitting, i.e. it would declare a conditionally valid
predictor invalid. Cross-fitting returns
{C6['oracle_cv_tree']:.5f} ({C6['oracle_cv_is_near_zero']}).

So the cross-validation in Algorithm 1 is not a refinement; without it the
metric is not a diagnostic at all.

## Limitations

The fully grown tree is the worst case, chosen to make the leakage visible. A
regularised classifier overfits less even in-sample, so the inflation factor
depends on the model class; the oracle false positive is the part that does not
depend on how the inflation is sized.
""")

# --------------------------------------------------------------- index + ES
short = {
    1: f"oracle floor {C1['oracle']['L1']:+.5f}; recovers {100 * C1['standard']['L1'] / C1['standard']['true_L1']:.0f}% of true L1",
    2: f"LightGBM {v1['lightgbm']:.2f}% vs paper 68.4; PartitionWise {v1['partitionwise']:.2f}% vs 38.3",
    3: f"CovGap error {last['covgap_error_vs_true']:.4f} > its own value; ERT separates {C3['separation_ratio_at_max_n']:.0f}× better",
    4: f"conservative ratio {sc[CONS][chr(39)+chr(39)] if False else sc[CONS]['L1_ratio_plus_over_minus']:.0f} vs aggressive {sc[AGGR]['L1_ratio_plus_over_minus']:.2f}",
    5: f"decomposition residual {C5['decomposition_max_residual']:.0e}; {C5['n_cells_with_divergent_components']}/{C5['n_cells']} cells divergent",
    6: f"in-sample false-positive {C6['oracle_in_sample_false_positive']:.4f} on an oracle predictor",
}
open("pages/index.md", "w").write("\n".join([
    "# Reproduction: Conditional Coverage Diagnostics for Conformal Prediction",
    "", "arXiv:2512.11779v1 · OpenReview `vaApZm6MKM`", "",
    tbl(["#", "claim", "verdict", "headline"],
        [[i, CLAIMS[i][:66] + "…", V[i], short[i]] for i in range(1, 7)]),
    "", "One page per claim. Every number is produced by the code in this Space.", ""]))

os.makedirs("pages/executive-summary", exist_ok=True)
open("pages/executive-summary/page.md", "w").write(f"""# Executive summary

**Paper.** *Conditional Coverage Diagnostics for Conformal Prediction*,
arXiv:2512.11779v1, OpenReview `vaApZm6MKM`.

**Result.** All six claims reproduce.

{SETUP}

## Two things worth knowing before reproducing this paper

1. **The Table 2 statistic normalises over sample size as well as method.**
   Taking the max over methods alone — the obvious reading — gives LightGBM
   ≈ 83% instead of 68.4% and inverts none of the ordering but none of the
   numbers match either. Including the number-of-test-samples axis, and
   clipping negative ERT to 0, reproduces the paper to
   {v1['lightgbm_abs_err_vs_paper']:.2f} pp on LightGBM and
   {v1['partitionwise_abs_err_vs_paper']:.2f} pp on PartitionWise.

2. **An untuned classifier makes the strictly proper ERTs go negative.** With a
   default boosted-tree configuration (max_iter = 200, no regularisation) the
   out-of-fold classifier loses to the constant 1−α predictor and this run
   measured L2-ERT = −0.0045 and KL-ERT = −0.0594 on a genuinely violating
   predictor. A shallow, strongly regularised model with early stopping instead
   recovers {100 * C1['standard']['L1'] / C1['standard']['true_L1']:.0f}% of the
   true L1 deviation. That is the paper's own thesis — the classifier is what
   gives the metric its power — appearing as a reproduction hazard, and it is
   why claim 2 matters as much as claim 1.

## Headline numbers

| claim | quantity | measured | reference |
|---|---|---|---|
| 1 | L1-ERT, oracle interval | {C1['oracle']['L1']:+.5f} | 0 |
| 1 | L1-ERT, standard conformal | {C1['standard']['L1']:.5f} | true {C1['standard']['true_L1']:.5f} |
| 2 | LightGBM L1-ERT power | {v1['lightgbm']:.2f}% | paper 68.4% |
| 2 | PartitionWise L1-ERT power | {v1['partitionwise']:.2f}% | paper 38.3% |
| 3 | CovGap error vs true L1 at n=5,000 | {last['covgap_error_vs_true']:.5f} | true {last['true_std']:.5f} |
| 5 | KL-ERT = KL₊ + KL₋ residual | {C5['decomposition_max_residual']:.2e} | 0 |
| 6 | in-sample L1-ERT on an oracle predictor | {C6['oracle_in_sample_false_positive']:.5f} | 0 |

Each is paired with a control that must break, and each does: the oracle
interval drives every ERT to ≈ 0, marginal coverage stays at
{C1['oracle']['marginal_coverage']:.4f} in both scenarios and so cannot
distinguish them, and dropping cross-fitting manufactures a violation where
there is none.
""")

json.dump({str(k): v for k, v in V.items()}, open("outputs/verdicts.json", "w"), indent=1)
print("verdicts:", V)