paper/aggregate.py

#!/usr/bin/env python3
"""Recompute every WER aggregate reported in the paper from a single

JSON file of per-cell WERs (``wer_results.json``, shipped alongside

this script).



This script is self-contained: it reads only the per-cell WER values

from the JSON and re-derives every table and figure number in the

paper that depends on those WERs alone.  Quantities that need inputs

beyond the per-cell WER table -- the per-utterance paired-bootstrap

significance markers in Table~\\ref{tab:main160}, or convergence-epoch

counts taken from the training logs -- are out of scope here and are

not recomputed.



Each check below is keyed to a stable LaTeX ``\\label`` (its ``\\ref``

target), which does not change if the paper is reordered; the ``[N]``

markers are just this script's own output tags, not paper section

numbers:



  [1]  tab:main160         FLEURS WER (%) @ 160 ms, per-(lang, hours).

  [2]  sec:hours intro     seen / unseen group mean Δ trajectory.

  [3]  tab:latency_effect  mean EN-ML gap per (tier, hours), FLEURS.

  [4]  tab:abs_wer         mean ML-init absolute WER per (tier, hours).

  [5]  tab:lst             per-lang  Δ  + LST  per hours; macro mean.

  [6]  sec:from_pl         HR-from-PL pivot vs direct ML, 160 ms.

  [7]  sec:reinitjoint     joiner-reinit Δ, 560 ms FLEURS/VP.

  [8]  tab:quant           INT4 vs FP32 @ 560 ms FLEURS, paired stats.

  [9]  tab:5000h_es        ES 5000h vs 2500h ML deltas, and ES 5000h

                           vs Nemotron-3.5 ML (tab:supp:es5000h).

  [10] fig:gap_powerlaw    fit  Δ(h) = a·h^(-β) at 160 ms;

                           language-bootstrap 95% CI on β; R².

  [11] tab:hybrid          hybrid-encoder 100h/560ms FLEURS table.

  [12] tab:seed + sec:seed HR/PT main-grid seed deltas + 16-cell 100h

                           re-run @ 560 ms (FLEURS + VP); IS-EN excursion.

  [13] streaming penalty   per (tier, hours, init), FLEURS.

  [14] tab:per_dataset     per-dataset 160 ms WER appendix.



Usage:

    # with wer_results.json sitting next to this script:

    python3 aggregate.py

    # or pass an explicit path to a results JSON:

    python3 aggregate.py /path/to/wer_results.json

"""
from __future__ import annotations

import json
import statistics
import sys
from pathlib import Path

HERE = Path(__file__).resolve().parent
# Default to wer_results.json sitting next to this script; fall back to
# the in-repository ``paper/`` layout when run from a full checkout.
DEFAULT_JSON = next(
    (p for p in (HERE / "wer_results.json",
                 HERE.parent / "paper" / "wer_results.json")
     if p.is_file()),
    HERE / "wer_results.json",
)

LANGS = ["de", "es", "fr", "hr", "is", "nl", "pl", "pt"]
SEEN = ["de", "es", "fr", "nl"]
UNSEEN = ["pt", "hr", "is", "pl"]
HOURS = ["100h", "250h", "500h", "1000h", "2500h"]
TIERS = ["160ms", "560ms", "1120ms", "offline"]


# ---------------------------------------------------------------------------
# JSON helpers
# ---------------------------------------------------------------------------
def wer(d, lang, hours, init, tier, dataset):
    """Return the WER (float) at d[lang][hours][init][tier][dataset], or None."""
    try:
        cell = d[lang][hours][init][tier][dataset]
        if isinstance(cell, dict):
            v = cell.get("wer")
            return float(v) if v is not None else None
    except (KeyError, TypeError):
        pass
    return None


def delta(d, lang, hours, tier, dataset):
    """EN - ML WER gap; None if either side missing."""
    ml = wer(d, lang, hours, "ml", tier, dataset)
    en = wer(d, lang, hours, "enc", tier, dataset)
    if ml is None or en is None:
        return None
    return en - ml


def fmt(x, signed=False, w=7):
    if x is None:
        return f"{'--':>{w}s}"
    return f"{x:+{w}.2f}" if signed else f"{x:{w}.2f}"


def mean(xs):
    xs = [x for x in xs if x is not None]
    return statistics.fmean(xs) if xs else None


def hr(title):
    print()
    print("=" * 78)
    print(title)
    print("=" * 78)


# ---------------------------------------------------------------------------
# [1] Table :main160 — FLEURS WER (%) at 160 ms streaming, per-(lang, hours)
# ---------------------------------------------------------------------------
def t_main160(d):
    hr("[1] Table tab:main160 — FLEURS WER (%) @ 160 ms")
    head = "lang  " + "".join(f"{h:>27s}" for h in HOURS)
    print(head)
    sub = "      " + "".join(f"{'ml':>9s}{'enc':>9s}{'Δ':>9s}" for _ in HOURS)
    print(sub)
    for lang in LANGS:
        row = f"{lang.upper():<6s}"
        for h in HOURS:
            ml = wer(d, lang, h, "ml", "160ms", "fleurs")
            en = wer(d, lang, h, "enc", "160ms", "fleurs")
            dl = (en - ml) if (ml is not None and en is not None) else None
            row += fmt(ml, w=9) + fmt(en, w=9) + fmt(dl, signed=True, w=9)
        print(row)
    print("-" * len(head))
    # macro means
    row = f"{'mean':<6s}"
    for h in HOURS:
        mls = [wer(d, l, h, "ml",  "160ms", "fleurs") for l in LANGS]
        ens = [wer(d, l, h, "enc", "160ms", "fleurs") for l in LANGS]
        dls = [(e - m) for m, e in zip(mls, ens) if m is not None and e is not None]
        row += fmt(mean(mls), w=9) + fmt(mean(ens), w=9) \
             + fmt(mean(dls) if dls else None, signed=True, w=9)
    print(row)
    n_per_h = {h: sum(1 for l in LANGS
                       if wer(d, l, h, "ml", "160ms", "fleurs") is not None
                       and wer(d, l, h, "enc", "160ms", "fleurs") is not None)
               for h in HOURS}
    print("  K per hours: " + ", ".join(f"{h}={n_per_h[h]}" for h in HOURS))


# ---------------------------------------------------------------------------
# [2] sec:hours — seen / unseen group mean Δ trajectory, FLEURS @ 160 ms
# ---------------------------------------------------------------------------
def t_seen_unseen(d):
    hr("[2] sec:hours — seen vs unseen group mean Δ, FLEURS @ 160 ms")
    print(f"  {'group':<10s}" + "".join(f"{h:>10s}" for h in HOURS))
    for label, group in (("seen", SEEN), ("unseen", UNSEEN), ("all", LANGS)):
        row = f"  {label:<10s}"
        for h in HOURS:
            dls = [delta(d, l, h, "160ms", "fleurs") for l in group]
            row += fmt(mean(dls), signed=True, w=10)
        print(row)


# ---------------------------------------------------------------------------
# [3] Table :latency_effect — mean EN-ML gap per (tier, hours), FLEURS
# ---------------------------------------------------------------------------
def t_latency_effect(d):
    hr("[3] Table tab:latency_effect — mean Δ (EN-ML) per (tier, hours), FLEURS")
    print(f"  {'tier':<10s}" + "".join(f"{h:>10s}" for h in HOURS)
          + f"{'range':>10s}")
    per_h = {h: [] for h in HOURS}
    for tier in TIERS:
        row = f"  {tier:<10s}"
        for h in HOURS:
            dls = [delta(d, l, h, tier, "fleurs") for l in LANGS]
            m = mean(dls)
            per_h[h].append(m)
            row += fmt(m, signed=True, w=10)
        print(row)
    # cross-tier range of mean Δ at each hours (paper sentence)
    print(f"  {'tier-range':<10s}", end="")
    for h in HOURS:
        vs = [v for v in per_h[h] if v is not None]
        rng = max(vs) - min(vs) if vs else None
        print(fmt(rng, w=10), end="")
    print()


# ---------------------------------------------------------------------------
# [4] Table :abs_wer — mean ML-init absolute FLEURS WER per (tier, hours)
# ---------------------------------------------------------------------------
def t_abs_wer(d):
    hr("[4] Table tab:abs_wer — mean ML-init FLEURS WER (%) per (tier, hours)")
    print(f"  {'tier':<10s}" + "".join(f"{h:>10s}" for h in HOURS))
    for tier in TIERS:
        row = f"  {tier:<10s}"
        for h in HOURS:
            mls = [wer(d, l, h, "ml", tier, "fleurs") for l in LANGS]
            row += fmt(mean(mls), w=10)
        print(row)


# ---------------------------------------------------------------------------
# [5] Table :lst — per-lang  Δ  + LST  per hours; macro mean
#     LST(lang, h) = max_tier Δ_tier(lang, h) - min_tier Δ_tier(lang, h)
#      Δ(lang, h) = mean_tier Δ_tier(lang, h)
#     Both are taken over the THREE streaming tiers only (160/560/1120 ms);
#     offline is excluded by definition (eq. lst / Table tab:lst caption).
# ---------------------------------------------------------------------------
def t_lst(d):
    hr("[5] Table tab:lst — per-lang  Δ , LST on FLEURS  (per hours)")
    stream_tiers = ["160ms", "560ms", "1120ms"]     # LST excludes offline
    h_show = ["100h", "250h", "500h", "1000h"]      # paper omits 2500h here
    head = f"  {'lang':<6s}" + "".join(f"{h:>20s}" for h in h_show)
    print(head)
    print(f"  {'':<6s}" + "".join(f"{'Δ':>10s}{'LST':>10s}" for _ in h_show))
    barD = {h: [] for h in h_show}
    lst = {h: [] for h in h_show}
    for lang in LANGS:
        row = f"  {lang.upper():<6s}"
        for h in h_show:
            ds = [delta(d, lang, h, t, "fleurs") for t in stream_tiers]
            ds = [x for x in ds if x is not None]
            if not ds:
                row += fmt(None, w=10) + fmt(None, w=10); continue
            bd = sum(ds) / len(ds)
            ls = max(ds) - min(ds)
            barD[h].append(bd); lst[h].append(ls)
            row += fmt(bd, signed=True, w=10) + fmt(ls, w=10)
        print(row)
    print("  " + "-" * 76)
    row = f"  {'mean':<6s}"
    for h in h_show:
        row += fmt(mean(barD[h]), signed=True, w=10) + fmt(mean(lst[h]), w=10)
    print(row)


# ---------------------------------------------------------------------------
# [6] Sec :from_pl — HR-from-PL pivot vs direct ML, 160 ms FLEURS/VP
# ---------------------------------------------------------------------------
def t_from_pl(d):
    hr("[6] Sec sec:from_pl — HR PL-pivot vs direct ML (160 ms)")
    rows = ["100h", "250h", "500h", "1000h"]
    print(f"  {'hours':<6s}{'direct ml FL':>14s}{'pivot ml FL':>14s}{'Δ FL':>8s}"
          f"{'direct ml VP':>14s}{'pivot ml VP':>14s}{'Δ VP':>8s}")
    for h in rows:
        dml_fl = wer(d, "hr", h, "ml",          "160ms", "fleurs")
        piv_fl = wer(d, "hr", h, "from_pl_ml",  "160ms", "fleurs")
        dml_vp = wer(d, "hr", h, "ml",          "160ms", "voxpopuli")
        piv_vp = wer(d, "hr", h, "from_pl_ml",  "160ms", "voxpopuli")
        dfl = (piv_fl - dml_fl) if (dml_fl is not None and piv_fl is not None) else None
        dvp = (piv_vp - dml_vp) if (dml_vp is not None and piv_vp is not None) else None
        print(f"  {h:<6s}{fmt(dml_fl,w=14)}{fmt(piv_fl,w=14)}{fmt(dfl,signed=True,w=8)}"
              f"{fmt(dml_vp,w=14)}{fmt(piv_vp,w=14)}{fmt(dvp,signed=True,w=8)}")
    # PL(EN) pivot only at 100h
    piv_en_fl = wer(d, "hr", "100h", "from_pl_enc", "160ms", "fleurs")
    den_fl    = wer(d, "hr", "100h", "enc",         "160ms", "fleurs")
    if piv_en_fl is not None and den_fl is not None:
        print(f"  PL(EN) pivot @100h FLEURS: direct enc={den_fl:.2f}  "
              f"pivot={piv_en_fl:.2f}  Δ={piv_en_fl - den_fl:+.2f} pp")


# ---------------------------------------------------------------------------
# [7] Sec :reinitjoint — joiner re-init Δ at 100 h, 560 ms FLEURS/VP
# ---------------------------------------------------------------------------
def t_reinitjoint(d):
    hr("[7] Sec sec:reinitjoint — joiner-reinit Δ vs baseline @ 100h, 560 ms")
    deltas_fl = {"ml": [], "enc": []}
    deltas_vp = {"ml": [], "enc": []}
    rows = []
    for lang in LANGS:
        for init in ("ml", "enc"):
            base_fl = wer(d, lang, "100h", init, "560ms", "fleurs")
            rj_fl   = wer(d, lang, "100h", f"{init}_reinitjoint", "560ms", "fleurs")
            base_vp = wer(d, lang, "100h", init, "560ms", "voxpopuli")
            rj_vp   = wer(d, lang, "100h", f"{init}_reinitjoint", "560ms", "voxpopuli")
            dfl = (rj_fl - base_fl) if (base_fl is not None and rj_fl is not None) else None
            dvp = (rj_vp - base_vp) if (base_vp is not None and rj_vp is not None) else None
            if dfl is not None: deltas_fl[init].append(dfl)
            if dvp is not None: deltas_vp[init].append(dvp)
            rows.append((lang, init, dfl, dvp))
    if not (deltas_fl["ml"] or deltas_fl["enc"]):
        print("  (no reinitjoint cells found in JSON)")
        return
    print(f"  {'lang':<4s} {'init':<4s} {'Δ FL':>8s} {'Δ VP':>8s}")
    for lang, init, dfl, dvp in rows:
        print(f"  {lang:<4s} {init:<4s} {fmt(dfl,signed=True,w=8)} {fmt(dvp,signed=True,w=8)}")
    all_fl = deltas_fl["ml"] + deltas_fl["enc"]
    all_vp = deltas_vp["ml"] + deltas_vp["enc"]
    hurt_fl = sum(1 for x in all_fl if x > 0)
    hurt_vp = sum(1 for x in all_vp if x > 0)
    print(f"\n  mean Δ FL  (all 16): {mean(all_fl):+0.2f} pp  ({hurt_fl}/{len(all_fl)} hurt)")
    print(f"  mean Δ FL  (ml arm): {mean(deltas_fl['ml']):+0.2f} pp")
    print(f"  mean Δ FL  (enc arm):{mean(deltas_fl['enc']):+0.2f} pp")
    print(f"  mean Δ VP  (all):    {mean(all_vp):+0.2f} pp  ({hurt_vp}/{len(all_vp)} hurt)")
    print(f"  mean Δ VP  (ml arm): {mean(deltas_vp['ml']):+0.2f} pp")
    print(f"  mean Δ VP  (enc arm):{mean(deltas_vp['enc']):+0.2f} pp")
    if deltas_fl["enc"] and deltas_fl["ml"]:
        worst_enc = max(rows, key=lambda r: r[2] if r[1] == "enc" and r[2] is not None else -1)
        worst_ml  = max(rows, key=lambda r: r[2] if r[1] == "ml"  and r[2] is not None else -1)
        print(f"  worst Δ FL enc: {worst_enc[0]}-{worst_enc[1]} {worst_enc[2]:+0.2f}")
        print(f"  worst Δ FL ml:  {worst_ml[0]}-{worst_ml[1]} {worst_ml[2]:+0.2f}")


# ---------------------------------------------------------------------------
# [8] Table :quant — INT4 vs FP32 @ 560 ms FLEURS
# ---------------------------------------------------------------------------
def t_quant(d):
    hr("[8] Table tab:quant — INT4 vs FP32 @ 560 ms FLEURS")
    rows = {h: {"ml": [], "enc": []} for h in HOURS}
    all_cells = []           # (lang, hours, init, Δ_int4-fp32)
    paired = []              # (lang, hours, Δ_ml, Δ_enc)
    for h in HOURS:
        per_lang_paired = {}
        for lang in LANGS:
            for init in ("ml", "enc"):
                i4 = wer(d, lang, h, init, "560ms", "fleurs_int4")
                f3 = wer(d, lang, h, init, "560ms", "fleurs_fp32")
                if i4 is None or f3 is None:
                    continue
                dq = i4 - f3
                rows[h][init].append(dq)
                all_cells.append((lang, h, init, dq))
                per_lang_paired.setdefault(lang, {})[init] = dq
        for lang, cells in per_lang_paired.items():
            if "ml" in cells and "enc" in cells:
                paired.append((lang, h, cells["ml"], cells["enc"]))

    print(f"  {'hours':<6s}{'Δ_ml':>9s}{'n_ml':>6s}{'Δ_enc':>9s}{'n_enc':>6s}")
    for h in HOURS:
        ml = rows[h]["ml"]; en = rows[h]["enc"]
        print(f"  {h:<6s}{fmt(mean(ml),signed=True,w=9)}{len(ml):>6d}"
              f"{fmt(mean(en),signed=True,w=9)}{len(en):>6d}")

    # pooled stats over all (lang, hours, init)
    dq_all = [c[3] for c in all_cells]
    if dq_all:
        dq_sorted = sorted(dq_all)
        med = statistics.median(dq_sorted)
        n = len(dq_all)
        within05 = sum(1 for x in dq_all if abs(x) <= 0.5)
        within10 = sum(1 for x in dq_all if abs(x) <= 1.0)
        worse    = sum(1 for x in dq_all if x > 0)
        better   = sum(1 for x in dq_all if x < 0)
        print()
        print(f"  pooled (n={n}):  mean={mean(dq_all):+0.2f}  median={med:+0.2f}"
              f"  range=[{min(dq_all):+0.2f},{max(dq_all):+0.2f}]")
        print(f"  |Δ| ≤ 0.5 pp: {within05}/{n}     |Δ| ≤ 1.0 pp: {within10}/{n}")
        print(f"  INT4 worse than FP32: {worse}/{n}     INT4 better: {better}/{n}")

    # paired ML vs EN quant cost
    if paired:
        diffs = [p[3] - p[2] for p in paired]              # Δ_enc - Δ_ml
        enc_costlier = sum(1 for x in diffs if x > 0)
        print(f"\n  paired ML vs EN (same lang, same hours, n={len(paired)}):")
        print(f"    mean (Δ_enc - Δ_ml) = {mean(diffs):+0.2f} pp")
        print(f"    EN costlier than ML in {enc_costlier}/{len(paired)} cells")
        print(f"    range = [{min(diffs):+0.2f}, {max(diffs):+0.2f}] pp")
        # 100h breakdown (paper says 'weakest at 100h, EN costlier in 5 of 8')
        for h in HOURS:
            sub = [(p[2], p[3]) for p in paired if p[1] == h]
            if not sub:
                continue
            sub_diffs = [b - a for a, b in sub]
            print(f"    {h}: paired n={len(sub)}  mean Δ_enc-Δ_ml={mean(sub_diffs):+0.2f}"
                  f"  EN costlier in {sum(1 for x in sub_diffs if x > 0)}/{len(sub)}")


# ---------------------------------------------------------------------------
# [9] ES 5000h ML deltas vs ES 2500h ML (tab:5000h_es), plus ES 5000h ML
#     vs the Nemotron-3.5 ML baseline (tab:supp:es5000h); all four test
#     sets, 560 ms streaming.
# ---------------------------------------------------------------------------
def t_es_5000h(d):
    if "5000h" not in d.get("es", {}):
        return
    hr("[9] Sec sec:hours — ES 5000h vs 2500h ML, 560 ms streaming")
    for dataset in ("fleurs", "cv", "mls", "voxpopuli"):
        a = wer(d, "es", "5000h", "ml", "560ms", dataset)
        b = wer(d, "es", "2500h", "ml", "560ms", dataset)
        delta_ = (a - b) if (a is not None and b is not None) else None
        print(f"  {dataset:<10s}  5000h={fmt(a)}  2500h={fmt(b)}"
              f"  Δ={fmt(delta_, signed=True)}")

    # tab:supp:es5000h — ES 5000h (ML) vs contemporaneous Nemotron-3.5 ML
    # baseline, 560 ms streaming.  Δ = WER_5000h − WER_Nemotron (negative:
    # our 5000h model better).
    if "nvidia-nemotron-3.5-asr" in d.get("es", {}).get("5000h", {}):
        print("  -- tab:supp:es5000h — ES 5000h (ML) vs Nemotron-3.5 ML:")
        for dataset in ("fleurs", "cv", "mls", "voxpopuli"):
            a = wer(d, "es", "5000h", "ml", "560ms", dataset)
            n = wer(d, "es", "5000h", "nvidia-nemotron-3.5-asr", "560ms", dataset)
            delta_ = (a - n) if (a is not None and n is not None) else None
            print(f"  {dataset:<10s}  5000h={fmt(a)}  nemotron={fmt(n)}"
                  f"  Δ={fmt(delta_, signed=True)}")


# ---------------------------------------------------------------------------
# [10] fig:gap_powerlaw — fit  Δ(h) = a · h^(-β)  on the 160 ms macro means
#      with a language-bootstrap 95% CI on β.
# ---------------------------------------------------------------------------
def t_powerlaw(d, B=10000, seed=0):
    hr("[10] fig:gap_powerlaw — transfer-gap power-law fit (160 ms FLEURS)")
    import math
    import random as _rand
    rng = _rand.Random(seed)
    hours_int = [int(h[:-1]) for h in HOURS]

    def fit(deltas_by_h):
        """Return (β, log_a, R²) from least-squares fit on log(h) vs log(Δ)

        across hours where Δ > 0 (so log is defined)."""
        xs, ys = [], []
        for h, dl in zip(hours_int, deltas_by_h):
            if dl is None or dl <= 0:
                continue
            xs.append(math.log(h)); ys.append(math.log(dl))
        if len(xs) < 2:
            return None, None, None
        n = len(xs)
        mx = sum(xs)/n; my = sum(ys)/n
        sxy = sum((x-mx)*(y-my) for x, y in zip(xs, ys))
        sxx = sum((x-mx)**2 for x in xs)
        if sxx == 0:
            return None, None, None
        slope = sxy / sxx       # = -β
        intercept = my - slope * mx
        beta = -slope
        # R²
        syy = sum((y-my)**2 for y in ys)
        if syy == 0:
            r2 = 1.0
        else:
            ss_res = sum((y - (slope*x + intercept))**2 for x, y in zip(xs, ys))
            r2 = 1 - ss_res/syy
        return beta, intercept, r2

    # point estimate from the full 8-lang macro means
    macro = [mean([delta(d, l, h, "160ms", "fleurs") for l in LANGS]) for h in HOURS]
    beta, log_a, r2 = fit(macro)
    if beta is None:
        print("  (insufficient points for power-law fit)")
        return
    print(f"  hours      mean Δ (160 ms FLEURS)")
    for h, m in zip(HOURS, macro):
        print(f"  {h:<8s}   {fmt(m, signed=True)}")
    print(f"\n  β_TG  = {beta:.3f}    R² = {r2:.4f}    "
          f"(fit on log Δ vs log h, where Δ>0)")

    # language-bootstrap CI: resample 8 languages with replacement, recompute
    # macro means per hours, refit, collect β.
    betas = []
    n_lang = len(LANGS)
    for _ in range(B):
        sample = [LANGS[rng.randrange(n_lang)] for _ in range(n_lang)]
        ms = [mean([delta(d, l, h, "160ms", "fleurs") for l in sample]) for h in HOURS]
        b, _, _ = fit(ms)
        if b is not None and math.isfinite(b):
            betas.append(b)
    if betas:
        betas.sort()
        lo = betas[int(0.025 * len(betas))]
        hi = betas[int(0.975 * len(betas))]
        frac_below = sum(1 for b in betas if b < 0.5) / len(betas)
        print(f"  language-bootstrap 95% CI on β (B={B}): [{lo:.2f}, {hi:.2f}]")
        print(f"  P(β < 0.5) = {frac_below:.4f}  ({100 * frac_below:.2f}%)")

    # off-tier residuals (paper Fig. fig:gap_powerlaw caption: 0.23/0.29/0.13 pp)
    import math as _m
    print("\n  Off-tier residuals against the 160 ms-fitted curve:")
    for tier in ("560ms", "1120ms", "offline"):
        mac = [mean([delta(d, l, h, tier, "fleurs") for l in LANGS]) for h in HOURS]
        res = []
        for h, m in zip(hours_int, mac):
            if m is None or m <= 0:
                continue
            pred = _m.exp(log_a + (-beta) * _m.log(h))
            res.append((m - pred) ** 2)
        if res:
            rms = (sum(res) / len(res)) ** 0.5
            print(f"    {tier:<8s}   RMS={rms:.2f} pp  (n={len(res)})")


# ---------------------------------------------------------------------------
# [11] tab:hybrid — hybrid-encoder, 100h training, 560 ms FLEURS
#      Splices: hybrid_0to12 / hybrid_last / hybrid_first / hybrid_middle
#      Reference rows are seed-45 ml / enc (paper: same common seed).
# ---------------------------------------------------------------------------
def t_hybrid(d):
    hr("[11] tab:hybrid — hybrid encoder @ 100h, 560 ms FLEURS")
    rows = [
        ("ML (s45)",   "ml_s45"),
        ("0:12",        "hybrid_0to12"),
        ("last",        "hybrid_last"),
        ("first",       "hybrid_first"),
        ("middle",      "hybrid_middle"),
        ("EN (s45)",   "enc_s45"),
    ]
    # header
    print(f"  {'row':<10s}" + "".join(f"{l.upper():>7s}" for l in LANGS)
          + f"{'mean':>9s}{'Δ vs ML':>10s}")
    ml_row = None
    table = []
    for label, init in rows:
        cells = [wer(d, l, "100h", init, "560ms", "fleurs") for l in LANGS]
        mn = mean(cells)
        table.append((label, init, cells, mn))
        if init == "ml_s45":
            ml_row = cells
    for label, init, cells, mn in table:
        if ml_row is None or any(c is None for c in cells) or any(m is None for m in ml_row):
            d_vs_ml = None
        else:
            pairwise = [c - m for c, m in zip(cells, ml_row)]
            d_vs_ml = sum(pairwise) / len(pairwise)
        row = f"  {label:<10s}" + "".join(fmt(c, w=7) for c in cells) + fmt(mn, w=9)
        row += fmt(d_vs_ml, signed=True, w=10) if d_vs_ml is not None else f"{'--':>10s}"
        print(row)

    # seen / unseen group means
    print("\n  group means:")
    for label, init, cells, _ in table:
        seen_m   = mean([wer(d, l, "100h", init, "560ms", "fleurs") for l in SEEN])
        unseen_m = mean([wer(d, l, "100h", init, "560ms", "fleurs") for l in UNSEEN])
        print(f"    {label:<10s}  seen={fmt(seen_m, w=7)}  unseen={fmt(unseen_m, w=7)}")


# ---------------------------------------------------------------------------
# [12] tab:seed + sec:seed — seed comparison
#      (a) HR/PT main-grid 160 ms FLEURS + VP, default seed vs seed 45
#      (b) 16-cell 100h re-run @ 560 ms across all 8 langs × 2 inits;
#          stats: mean |Δ|, median, max on FLEURS and VP
#      (c) HR-100h FLEURS swing collapse at 560 ms
#      (d) Largest seed-induced EN-ML gap excursion (IS at 100h)
# ---------------------------------------------------------------------------
def t_seed(d):
    hr("[12] tab:seed + sec:seed — seed-default vs seed-45")
    # (a) HR / PT main grid at 160 ms FLEURS + VP
    print("  (a) HR / PT main-grid FLEURS@160ms and VP@160ms seed deltas:")
    print(f"  {'lang':<4s}{'init':<5s}{'hours':<6s}"
          f"{'FL_s1':>8s}{'FL_s2':>8s}{'ΔFL':>8s}"
          f"{'VP_s1':>8s}{'VP_s2':>8s}{'ΔVP':>8s}")
    fl_abs, vp_abs = [], []
    rows = []
    for lang in ("hr", "pt"):
        for init in ("ml", "enc"):
            for h in HOURS:
                a_fl = wer(d, lang, h, init,       "160ms", "fleurs")
                b_fl = wer(d, lang, h, init+"_s45","160ms", "fleurs")
                a_vp = wer(d, lang, h, init,       "160ms", "voxpopuli")
                b_vp = wer(d, lang, h, init+"_s45","160ms", "voxpopuli")
                if a_fl is None or b_fl is None:
                    continue
                dfl = b_fl - a_fl
                dvp = (b_vp - a_vp) if (a_vp is not None and b_vp is not None) else None
                fl_abs.append(abs(dfl))
                if dvp is not None: vp_abs.append(abs(dvp))
                rows.append((lang, init, h, dfl, dvp))
                print(f"  {lang:<4s}{init:<5s}{h:<6s}"
                      f"{a_fl:>8.2f}{b_fl:>8.2f}{dfl:>+8.2f}"
                      f"{fmt(a_vp,w=8)}{fmt(b_vp,w=8)}{fmt(dvp,signed=True,w=8)}")
    if fl_abs:
        print(f"\n    mean |ΔFL| = {sum(fl_abs)/len(fl_abs):.2f} pp"
              f"  (n={len(fl_abs)}, max={max(fl_abs):.2f})")
    if vp_abs:
        print(f"    mean |ΔVP| = {sum(vp_abs)/len(vp_abs):.2f} pp"
              f"  (n={len(vp_abs)}, max={max(vp_abs):.2f})")

    # (b) 16-cell 100h re-run @ 560 ms (all 8 langs × 2 inits)
    print("\n  (b) 100h × 8 langs × 2 inits  @ 560 ms (FLEURS, VP):")
    fl, vp = [], []
    cells_fl, cells_vp = [], []
    for lang in LANGS:
        for init in ("ml", "enc"):
            a_fl = wer(d, lang, "100h", init,        "560ms", "fleurs")
            b_fl = wer(d, lang, "100h", init+"_s45", "560ms", "fleurs")
            a_vp = wer(d, lang, "100h", init,        "560ms", "voxpopuli")
            b_vp = wer(d, lang, "100h", init+"_s45", "560ms", "voxpopuli")
            if a_fl is not None and b_fl is not None:
                fl.append(abs(b_fl - a_fl))
                cells_fl.append((lang, init, b_fl - a_fl))
            if a_vp is not None and b_vp is not None:
                vp.append(abs(b_vp - a_vp))
                cells_vp.append((lang, init, b_vp - a_vp))
    if fl:
        s = sorted(fl)
        worst = max(cells_fl, key=lambda x: abs(x[2]))
        print(f"    FLEURS:  mean |Δ|={sum(fl)/len(fl):.2f}  median={s[len(s)//2]:.2f}"
              f"  max={max(fl):.2f}  ({worst[0]}-{worst[1]})  n={len(fl)}")
    if vp:
        s = sorted(vp)
        worst = max(cells_vp, key=lambda x: abs(x[2]))
        print(f"    VoxPopuli: mean |Δ|={sum(vp)/len(vp):.2f}  median={s[len(s)//2]:.2f}"
              f"  max={max(vp):.2f}  ({worst[0]}-{worst[1]})  n={len(vp)}")

    # (c) HR-100h FLEURS swing collapse at 560 ms vs 160 ms
    print("\n  (c) HR-100h FLEURS seed swing collapse 160 ms → 560 ms:")
    for init in ("ml", "enc"):
        for tier in ("160ms", "560ms"):
            a = wer(d, "hr", "100h", init,        tier, "fleurs")
            b = wer(d, "hr", "100h", init+"_s45", tier, "fleurs")
            if a is not None and b is not None:
                print(f"    hr-{init:<3s} @ {tier:<6s}: s1={a:.2f}  s2={b:.2f}  Δ={b-a:+.2f}")

    # (d) Largest seed-induced EN-ML gap excursion at 100h / 560 ms
    print("\n  (d) Seed-induced EN-ML gap excursion at 100h / 560 ms FLEURS:")
    print(f"  {'lang':<4s}{'Δ(s1)':>10s}{'Δ(s2)':>10s}{'shift':>10s}{'sign?':>8s}")
    biggest = (None, 0.0)
    for lang in LANGS:
        ml1 = wer(d, lang, "100h", "ml",      "560ms", "fleurs")
        en1 = wer(d, lang, "100h", "enc",     "560ms", "fleurs")
        ml2 = wer(d, lang, "100h", "ml_s45",  "560ms", "fleurs")
        en2 = wer(d, lang, "100h", "enc_s45", "560ms", "fleurs")
        if any(x is None for x in (ml1, en1, ml2, en2)):
            continue
        d1 = en1 - ml1; d2 = en2 - ml2
        sign_keep = "yes" if (d1 >= 0) == (d2 >= 0) else "NO"
        shift = d2 - d1
        print(f"  {lang:<4s}{d1:>+10.2f}{d2:>+10.2f}{shift:>+10.2f}{sign_keep:>8s}")
        if abs(shift) > abs(biggest[1]):
            biggest = (lang, shift)
    if biggest[0]:
        print(f"    largest shift: {biggest[0]} ({biggest[1]:+.2f} pp)")


# ---------------------------------------------------------------------------
# [13] Streaming penalty per (tier, hours, init) on FLEURS
#      penalty = WER(tier) - WER(offline)
# ---------------------------------------------------------------------------
def t_streaming_penalty(d):
    hr("[13] Streaming penalty (FLEURS): WER(tier) - WER(offline)")
    print(f"  {'init':<5s}{'tier':<10s}" + "".join(f"{h:>10s}" for h in HOURS))
    for init in ("ml", "enc"):
        for tier in ("160ms", "560ms", "1120ms"):
            row = f"  {init:<5s}{tier:<10s}"
            for h in HOURS:
                tier_wers = [wer(d, l, h, init, tier,     "fleurs") for l in LANGS]
                off_wers  = [wer(d, l, h, init, "offline", "fleurs") for l in LANGS]
                pairs = [(t - o) for t, o in zip(tier_wers, off_wers)
                         if t is not None and o is not None]
                m = mean(pairs)
                row += fmt(m, signed=True, w=10)
            print(row)


# ---------------------------------------------------------------------------
# [14] tab:per_dataset — 160 ms WER per (lang, hours, init, dataset)
# ---------------------------------------------------------------------------
def t_per_dataset(d):
    hr("[14] tab:per_dataset — 160 ms WER per (lang, hours, init, dataset)")
    datasets = ("cv", "mls", "voxpopuli", "fleurs")
    print(f"  {'lang':<5s}{'hours':<7s}"
          + "".join(f"{ds.upper() + ' ml':>10s}{ds.upper() + ' en':>10s}"
                    for ds in datasets))
    for lang in LANGS:
        for h in HOURS:
            row = f"  {lang:<5s}{h:<7s}"
            for ds in datasets:
                ml = wer(d, lang, h, "ml",  "160ms", ds)
                en = wer(d, lang, h, "enc", "160ms", ds)
                row += fmt(ml, w=10) + fmt(en, w=10)
            print(row)


# ---------------------------------------------------------------------------
# main
# ---------------------------------------------------------------------------
def main(argv):
    src = Path(argv[1]) if len(argv) > 1 else DEFAULT_JSON
    if not src.is_file():
        print(f"error: results file not found: {src}\n"
              f"       pass the path to wer_results.json as the first argument.",
              file=sys.stderr)
        return 2
    d = json.loads(src.read_text())
    print(f"Reading {src}")
    t_main160(d)
    t_seen_unseen(d)
    t_latency_effect(d)
    t_abs_wer(d)
    t_lst(d)
    t_from_pl(d)
    t_reinitjoint(d)
    t_quant(d)
    t_es_5000h(d)
    t_powerlaw(d)
    t_hybrid(d)
    t_seed(d)
    t_streaming_penalty(d)
    t_per_dataset(d)
    return 0


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