claude-opus-5

{
  "command": "python - <<'EOF'\np='submission/curate.py'; s=open(p).read()\ns=s.replace('OUT = \"/workspace/submission/selection.json\"',\n 'OUT = os.environ.get(\"OUT\", \"/workspace/submission/selection.json\")\\nMODE = os.environ.get(\"MODE\", \"logodds\")   # logodds | lmfit | hygiene')\n# add pool unigram + logodds scoring and mode switch\ns=s.replace(\"\"\"    tot = np.add.reduceat(lp, off[:-1])\n    cnt = np.maximum(np.diff(off), 1)\n    scores[r] = (tot / cnt).astype(np.float32)\"\"\",\n\"\"\"    if MODE == \"logodds\":\n        # frequent-vocabulary log-odds: log P_r(w) - log P_pool(w), averaged.\n        # Restricting to the target's frequent vocabulary is what keeps this\n        # from degenerating: unbounded rare features would let a discriminator\n        # rank rare-vocabulary junk at the very top.\n        w_lo = np.log(p_uni) - np.log(p_pool)\n        lp = w_lo[flat]\n        lp[oov] = OOV_ODDS\n    tot = np.add.reduceat(lp, off[:-1])\n    cnt = np.maximum(np.diff(off), 1)\n    scores[r] = (tot / cnt).astype(np.float32)\"\"\")\ns=s.replace('oov = flat == 0',\n\"\"\"oov = flat == 0\n# pool-side unigram distribution over the same vocabulary (the \"do-nothing\"\n# random-selection reference distribution) -- denominator of the log-odds score\n_pc = np.bincount(flat, minlength=V).astype(np.float64)\np_pool = (_pc + 0.5) / (_pc.sum() + 0.5 * V)\"\"\")\ns=s.replace('LOG_OOV = -13.0            # fixed log-prob charged to out-of-vocabulary words',\n 'LOG_OOV = -13.0            # fixed log-prob charged to OOV words (lmfit mode)\\nOOV_ODDS = -1.0            # log-odds penalty charged to OOV words (logodds mode)')\n# hygiene mode: random order\ns=s.replace(\"order = {r: np.argsort(-scores[r]) for r in REGISTERS}\",\n\"\"\"if MODE == \"hygiene\":     # ablation: filter only, no ranking signal\n    _rs = np.random.default_rng(0)\n    for r in REGISTERS:\n        scores[r] = _rs.random(len(cand)).astype(np.float32)\norder = {r: np.argsort(-scores[r]) for r in REGISTERS}\"\"\")\nopen(p,'w').write(s)\nEOF\ngrep -n \"MODE\\|p_pool\\|OOV_ODDS\" submission/curate.py | head -20",
  "description": "Add mode switch, logodds and hygiene ablations"
}

{
  "command": "for M in hygiene logodds; do OUT=/tmp/sel_$M.json MODE=$M python submission/curate.py >/tmp/c_$M.log 2>&1; echo \"-- $M $(tail -1 /tmp/c_$M.log)\"; done",
  "description": "Generate ablation selections",
  "timeout": 1400000
}

{
  "command": "for M in hygiene logodds; do echo -n \"$M: \"; bash dev_eval.sh /tmp/sel_$M.json 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"; done; echo TIME=$(cat .timer/remaining_secs)",
  "description": "Eval ablations",
  "timeout": 1400000
}

{
  "command": "python - <<'EOF'\np='submission/curate.py'; s=open(p).read()\ns=s.replace('p_pool = (_pc + 0.5) / (_pc.sum() + 0.5 * V)',\n\"\"\"p_pool = (_pc + 0.5) / (_pc.sum() + 0.5 * V)\n# pool-side bigram counts over the same hashed bigram space\nbic_pool = np.zeros(BH, np.float32); np.add.at(bic_pool, BIG, 1.0)\nctx_pool = np.zeros(V, np.float64);  np.add.at(ctx_pool, prev, 1.0)\np2_pool_all = bic_pool[BIG] / np.maximum(ctx_pool, 1.0)[prev]\np2_pool_all[oov | np.asarray(prev == 0)] = 0.0\nlp_pool = np.log(L2 * p2_pool_all + L1 * p_pool[flat] + L0 / V)\"\"\")\ns=s.replace(\"\"\"    if MODE == \"logodds\":\"\"\",\n\"\"\"    if MODE == \"lr2\":\n        # log-likelihood RATIO under interpolated bigram LMs: target register r\n        # versus the pool itself. Combines the register-discriminative power of\n        # a log-odds score with bigram (local fluency) context, while the\n        # frequent-word vocabulary bounds how much any single rare feature can\n        # contribute.\n        lp = lp - lp_pool\n        lp[oov] = OOV_ODDS\n    elif MODE == \"logodds\":\"\"\")\ns=s.replace(\"\"\"rest = np.flatnonzero(~taken)\"\"\",\n\"\"\"if MODE == \"global\":      # ablation: one unbalanced ranking, no round-robin\n    pass\nrest = np.flatnonzero(~taken)\"\"\")\n# global mode: collapse all registers to a single ranking\ns=s.replace(\"\"\"if MODE == \"hygiene\":     # ablation: filter only, no ranking signal\"\"\",\n\"\"\"if MODE == \"global\":      # ablation: single unbalanced ranking over all docs\n    _g = np.max(np.stack([scores[r] for r in REGISTERS]), 0)\n    for r in REGISTERS:\n        scores[r] = _g\nif MODE == \"hygiene\":     # ablation: filter only, no ranking signal\"\"\")\nopen(p,'w').write(s)\nEOF\npython - <<'EOF'\n# global mode must not round-robin identical rankings into 4x share; make one register active\np='submission/curate.py'; s=open(p).read()\ns=s.replace('sel_local, active = [], list(REGISTERS)',\n            'sel_local, active = [], ([\"wiki\"] if MODE == \"global\" else list(REGISTERS))')\ns=s.replace('share = BUDGET * OVERSHOOT / len(REGISTERS)',\n            'share = BUDGET * OVERSHOOT / (1 if MODE == \"global\" else len(REGISTERS))')\nopen(p,'w').write(s)\nEOF\nfor M in lr2 global; do OUT=/tmp/sel_$M.json MODE=$M python submission/curate.py >/tmp/c_$M.log 2>&1 || tail -5 /tmp/c_$M.log; echo \"-- $M $(tail -1 /tmp/c_$M.log)\"; done\nOUT=/tmp/sel_oov.json MODE=logodds OOVX=1 python - <<'EOF' >/tmp/c_oov.log 2>&1\nimport os,runpy\nos.environ[\"MODE\"]=\"logodds\"\nsrc=open('/workspace/submission/curate.py').read().replace('OOV_ODDS = -1.0','OOV_ODDS = -2.5')\nexec(compile(src,'curate_oov','exec'),{'__name__':'__main__'})\nEOF\necho \"-- oov $(tail -1 /tmp/c_oov.log)\"",
  "description": "Generate lr2, global, oov variants",
  "timeout": 1400000
}

{
  "command": "for M in lr2 global oov; do echo -n \"$M: \"; bash dev_eval.sh /tmp/sel_$M.json 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"; done; echo TIME=$(cat .timer/remaining_secs)",
  "description": "Eval lr2/global/oov",
  "timeout": 1400000
}

{
  "command": "python - <<'EOF'\np='submission/curate.py'; s=open(p).read()\ns=s.replace('OOV_ODDS = -1.0','OOV_ODDS = float(os.environ.get(\"OOV_ODDS\", -2.5))')\nopen(p,'w').write(s)\nEOF\nfor O in 5.0 9.0; do (OUT=/tmp/sel_o$O.json MODE=logodds OOV_ODDS=-$O python submission/curate.py >/tmp/c_o$O.log 2>&1) & done; wait\nfor O in 5.0 9.0; do echo \"-- $O $(tail -1 /tmp/c_o$O.log)\"; done",
  "description": "Curate OOV -5 and -9 in parallel",
  "timeout": 1400000
}

{
  "command": "for O in 5.0 9.0; do echo -n \"oov-$O: \"; bash dev_eval.sh /tmp/sel_o$O.json 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"; done; echo TIME=$(cat .timer/remaining_secs)",
  "description": "Eval OOV sweep",
  "timeout": 1400000
}

{
  "command": "python - <<'EOF'\np='submission/curate.py'; s=open(p).read()\ns=s.replace('WORD = re.compile(r\"[a-z0-9\\']+\")',\n 'MINCH = int(os.environ.get(\"MINCH\", 500))   # shortest document kept, in chars\\nLOOSE = os.environ.get(\"LOOSE\") == \"1\"      # ablation: drop the prose-shape checks\\n\\nWORD = re.compile(r\"[a-z0-9\\']+\")',1)\ns=s.replace('    if n < 500 or n > 300_000:','    if n < MINCH or n > 300_000:')\ns=s.replace(\"\"\"    nw = [ln.count(\" \") + 1 for ln in lines]\"\"\",\n\"\"\"    nw = [ln.count(\" \") + 1 for ln in lines]\n    if LOOSE:\n        return len(WORD.findall(head.lower())) >= 90\"\"\")\nopen(p,'w').write(s)\nEOF\nfor V in \"MINCH=2500\" \"LOOSE=1\"; do (env $V OUT=/tmp/sel_${V//=/}.json MODE=logodds python submission/curate.py >/tmp/c_${V//=/}.log 2>&1) & done; wait\ngrep -h \"candidates kept\\|done\" /tmp/c_MINCH2500.log /tmp/c_LOOSE1.log",
  "description": "Curate MINCH and LOOSE variants",
  "timeout": 1400000
}

{
  "command": "for V in MINCH2500 LOOSE1; do echo -n \"$V: \"; bash dev_eval.sh /tmp/sel_$V.json 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"; done; echo TIME=$(cat .timer/remaining_secs)",
  "description": "Eval MINCH/LOOSE",
  "timeout": 1400000
}

{
  "command": "python - <<'EOF'\np='submission/curate.py'; s=open(p).read()\ns=s.replace('LOOSE = os.environ.get(\"LOOSE\") == \"1\"      # ablation: drop the prose-shape checks',\n 'LOOSE = os.environ.get(\"LOOSE\") == \"1\"      # keep only the coarse prose checks\\nNOFILTER = os.environ.get(\"NOFILTER\") == \"1\"  # ablation: dedup + length only')\ns=s.replace(\"\"\"def hygienic(t):\n    n = len(t)\"\"\",\"\"\"def hygienic(t):\n    n = len(t)\n    if NOFILTER:\n        return MINCH <= n <= 300_000\"\"\")\nopen(p,'w').write(s)\nEOF\n( NOFILTER=1 MINCH=200 OUT=/tmp/sel_NOF.json MODE=logodds python submission/curate.py >/tmp/c_NOF.log 2>&1 ) &\n( LOOSE=1 OOV_ODDS=-1.0 OUT=/tmp/sel_L1.json MODE=logodds python submission/curate.py >/tmp/c_L1.log 2>&1 ) &\n( LOOSE=1 OOV_ODDS=-5.0 OUT=/tmp/sel_L5.json MODE=logodds python submission/curate.py >/tmp/c_L5.log 2>&1 ) &\nwait; grep -h \"candidates kept\" /tmp/c_NOF.log /tmp/c_L1.log /tmp/c_L5.log",
  "description": "Curate nofilter and LOOSE OOV variants",
  "timeout": 1400000
}

{
  "command": "for V in NOF L1 L5; do echo -n \"$V: \"; bash dev_eval.sh /tmp/sel_$V.json 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"; done; echo TIME=$(cat .timer/remaining_secs)",
  "description": "Eval nofilter/LOOSE-OOV variants",
  "timeout": 1400000
}

{
  "command": "( LOOSE=1 OOV_ODDS=0.0 OUT=/tmp/sel_L0.json MODE=logodds python submission/curate.py >/tmp/c_L0.log 2>&1 ) &\n( LOOSE=1 OOV_ODDS=-0.5 OUT=/tmp/sel_L05.json MODE=logodds python submission/curate.py >/tmp/c_L05.log 2>&1 ) &\nwait\nfor V in L0 L05; do echo -n \"$V: \"; bash dev_eval.sh /tmp/sel_$V.json 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"; done; echo TIME=$(cat .timer/remaining_secs)",
  "description": "Sweep OOV 0.0 and -0.5",
  "timeout": 1400000
}

{
  "command": "python - <<'EOF'\np='/workspace/submission/curate.py'; s=open(p).read()\ns=s.replace('LOOSE = os.environ.get(\"LOOSE\") == \"1\"      # keep only the coarse prose checks',\n            'LOOSE = os.environ.get(\"LOOSE\", \"1\") == \"1\"  # coarse prose checks only (default)')\ns=s.replace('OOV_ODDS = float(os.environ.get(\"OOV_ODDS\", -2.5))',\n            'OOV_ODDS = float(os.environ.get(\"OOV_ODDS\", -1.0))')\ns=s.replace('MODE = os.environ.get(\"MODE\", \"logodds\")   # logodds | lmfit | hygiene',\n            'MODE = os.environ.get(\"MODE\", \"logodds\")   # logodds (default) | lr2 | lmfit | hygiene | global')\nopen(p,'w').write(s)\nEOF\ngrep -n 'LOOSE = \\|OOV_ODDS = float\\|MINCH = ' /workspace/submission/curate.py\npython /workspace/submission/curate.py 2>&1 | tail -9",
  "description": "Set winning defaults and regenerate final selection",
  "timeout": 1400000
}

{
  "command": "python -c \"\nimport json\nsel=json.load(open('submission/selection.json')); ref=json.load(open('/tmp/sel_L1.json'))\nprint('identical to best config:', sel==ref)\nprint('n',len(sel),'unique',len(set(sel)),'min',min(sel),'max',max(sel))\n\" && bash dev_eval.sh 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"; echo TIME=$(cat /workspace/.timer/remaining_secs)",
  "description": "Verify final selection reproduces best config",
  "timeout": 1400000
}

{
  "command": "python -c \"\nimport numpy as np\nd=np.load('data/multi_dev.npy'); q=len(d)//4\nfor k,n in enumerate(['wiki','webprose','news','qa']):\n    np.save(f'/tmp/dev_{n}.npy', d[k*q:(k+1)*q])\nprint('ok')\"\ncat > /tmp/perreg.sh <<'EOF'\nset -e\ncd /workspace\npython3 pack_selection.py \"$1\" /tmp/pr_train.npy >/dev/null 2>&1\nfor n in wiki webprose news qa; do\n  echo -n \"$2 $n \"\n  python3 train_nano.py --train_npy /tmp/pr_train.npy --heldout_npy /tmp/dev_$n.npy \\\n    --out_json /tmp/pr.json 2>&1 | grep -o \"HELDOUT_PPL [0-9.]*\"\ndone\nEOF\nbash /tmp/perreg.sh /workspace/submission/selection.json BALANCED 2>&1 | tail -6",
  "description": "Per-register PPL for balanced selection",
  "timeout": 1400000
}

{
  "command": "bash /tmp/perreg.sh /tmp/sel_global.json GLOBAL 2>&1 | tail -5; echo TIME=$(cat /workspace/.timer/remaining_secs)",
  "description": "Per-register PPL for global (unbalanced) selection",
  "timeout": 1400000
}

{
  "file_path": "/workspace/claim.md",
  "content": "# Curating a raw web pool for a 12M-token budget\n\nDev perplexity on `multi_dev.npy`, frozen 30M GPT, 12M-token budget:\n\n| selection | dev PPL |\n|---|---|\n| random (do-nothing baseline) | 459.8 |\n| strict \"quality\" hygiene filter only, random order | 446.9 |\n| top-of-ranking by target-LM *likelihood* (CCNet-style) | 554.8 |\n| NB log-count-ratio over unbounded hashed n-grams | 1124.7 |\n| single global ranking, register-balance removed | 665.3 |\n| **balanced frequent-vocab log-odds (submitted)** | **322.8** |\n\n## Hypothesis\n\nFor a broad, equal-parts target, held-out perplexity is set by two things that\npull against each other, and *only* by their combination:\n\n1. **Register coverage.** The score is a token-weighted average of per-register\n   losses, so it is dominated by whichever target register the training set\n   covers *worst*. Any selection rule that produces one global ranking will\n   over-serve whichever register is most separable from generic web text and\n   starve the others.\n2. **Discriminativeness without rare-feature blowup.** The useful signal is\n   *relative*: how much more probable a document is under a target register than\n   under the pool. But that ratio must be computed over a bounded, frequent\n   vocabulary. Generative likelihood alone ranks bland SEO filler first; an\n   unbounded log-count ratio ranks rare-vocabulary junk first.\n\nThe corollary I did *not* expect and which the ablations force: **aggressive\n\"quality\" filtering is harmful once a good relative scorer is in place.** The\nstrict hygiene filter (mean words/line, stopword fraction, terminal-punctuation\ndensity, type/token ratio) dropped 45% of the pool and cost 12 PPL versus the\ncoarse filter that drops only 17% (335 → 323 at matched scorer; 447 for the\nstrict filter with no scorer). The filter and the scorer are not additive — the\nfilter removes diversity the scorer would have kept.\n\n## Mechanism — a prediction *other than* the final perplexity\n\nIf the mechanism is register coverage rather than generic \"document quality\",\nthen removing the round-robin (keeping the identical scorer, identical filter,\nidentical budget) must show up as **collapse of the worst register, not a uniform\ndegradation**. Measured per-quarter dev PPL, same trained recipe:\n\n| register | balanced (submitted) | global ranking |\n|---|---|---|\n| wiki | 803.8 | 1304.9 |\n| web prose | 288.9 | 304.3 |\n| news | 311.9 | 447.1 |\n| technical Q&A | 149.4 | **1099.3** |\n| **max across registers** | **803.8** | 1304.9 |\n| max / min spread | 5.4× | 8.7× |\n\nConfirmed, and confirmed in the specific predicted shape: web prose barely moves\n(304 → 289, it is what generic web text already is), while technical Q&A — the\nregister least like generic web text — degrades 7.4× when the balance constraint\nis dropped. The aggregate 665 → 323 improvement is almost entirely the Q&A and\nwiki quarters being rescued, not a broad lift.\n\nA second, unforced observation that supports the same reading: wiki stays the\nworst quarter under *both* selections. The wiki target is detokenized\nWikiText-103 with its distinctive surface form (` @,@ `, ` @-@ `, spaces before\npunctuation). No pool document has that convention, so no selection can fix it —\nit is a floor imposed by the target's tokenization, not by curation. This bounds\nhow far this method can go and predicts that the residual is concentrated in one\nquarter rather than spread evenly.\n\n## Falsification\n\nThe claim is falsified if any of these hold:\n\n- **Balance is not the mechanism.** Removing round-robin degrades all four\n  registers by a similar factor instead of collapsing the least-web-like ones.\n  *Tested: refuted — Q&A 7.4× vs web prose 1.05×.*\n- **The relative score is doing nothing beyond junk removal.** Filter-only with\n  random order matches the full method. *Tested: refuted — 446.9 vs 322.8.*\n- **Bounded vocabulary is not what saves the discriminator.** Letting the same\n  log-odds score range over unbounded hashed n-grams performs comparably.\n  *Tested: refuted — 1124.7, i.e. 2.4× worse than random; its top-ranked\n  documents are plant-genus lists, PGP key blocks and navigation menus.*\n- **Direction of the likelihood signal.** If ranking by generative target-LM\n  likelihood beat ranking by the pool-relative log-odds, the \"relative, not\n  absolute\" claim fails. *Tested: refuted — 554.8 vs 322.8; likelihood ranking is\n  worse than random because bland, high-frequency filler is the most probable\n  text under any smoothed word LM.*\n- **Stricter filtering should help.** If tightening hygiene monotonically\n  improved PPL, the diversity-cost claim fails. *Tested: refuted — strict 335 →\n  coarse 323, and raising the length floor to 2500 chars costs 72 PPL (394.8).*\n- The one lever with an interior optimum, the OOV penalty, would falsify the\n  \"bounded contribution\" story if PPL were monotone in it. It is not:\n  0.0 → 337.6, −0.5 → 326.7, **−1.0 → 322.8**, −2.5 → 335.0, −5.0 → 346.6,\n  −9.0 → 421.6. An interior optimum is what a bounded-contribution account\n  predicts: too little penalty admits gibberish, too much admits only\n  common-word filler.\n\n## Transfer\n\n- **What transfers.** The recipe needs nothing but the target sample itself: fit\n  a cheap word-level LM per target register, score the pool by log-odds against\n  the pool's own unigram distribution over a frequent vocabulary, and take equal\n  token shares per register in interleaved priority order. Total cost is ~80 s of\n  CPU for 182k documents and no GPU, no labels, no pretrained quality classifier.\n  It should transfer to any \"curate for a known multi-domain target\" setting, and\n  the interleaved (rather than register-blocked) priority order means it degrades\n  gracefully under a budget different from the one it was tuned at — every prefix\n  of the list is already balanced.\n- **What does not transfer.** The specific OOV penalty (−1.0) and filter\n  strictness were tuned against this pool at this budget; both had ~30-100 PPL\n  swings, so they are worth re-sweeping. More importantly, the \"loose filtering\n  wins\" result is budget-dependent: at 6% of the pool the binding constraint is\n  coverage, so diversity beats purity. At a budget approaching the whole pool the\n  ranking cannot buy coverage any more and I would expect strict filtering to\n  reverse sign and start helping.\n- **Untested and honestly uncertain.** Every number here is one seed on the dev\n  target; differences under ~15 PPL should not be trusted, which is why I did not\n  chase the −0.5/−1.0 boundary further. I also did not test whether balancing on\n  *equal token shares* beats balancing on equal loss (an adaptive allocation that\n  gives more budget to the worst register — wiki at 804 — is the obvious next\n  step and the per-register table above is exactly the diagnostic for it).\n"
}

{
  "command": "python - <<'PY'\np='/workspace/submission/curate.py'; s=open(p).read()\nnew='''\"\"\"Curate /workspace/data/pool.jsonl for a 12M-token pretraining budget.\n\nDev PPL of this script's default configuration: 322.8\n(random-selection baseline on the same frozen recipe: 459.8)\n\nSTATED CRITERION\n----------------\nThe disclosed target is a BROAD, equal-parts mixture of four registers:\nencyclopedic (Wikipedia), high-quality general web prose, news, and technical\nQ&A. The dev target `multi_dev.npy` is a GPT-2 token stream of exactly that\nmixture, laid out in four contiguous quarters, one register each.\n\n    Rank pool documents by how much MORE probable they are under a target\n    register's word distribution than under the pool's own -- measured over a\n    bounded frequent vocabulary -- and take EQUAL TOKEN SHARES from each of the\n    four registers, interleaved.\n\nThree parts, in order:\n\n(A) COARSE PROSE FILTER. Drop only what is not running prose at all: navigation\n    menus and taxonomy/word lists (mean words per line, stub-line fraction),\n    key blocks and symbol soup (alphabetic fraction), repeated boilerplate\n    (line duplication), plus a length floor and an exact-duplicate signature.\n    Deliberately coarse: at 6% of the pool the binding constraint is register\n    coverage, so a stricter \"quality\" filter costs more in lost diversity than\n    it gains in purity (measured: strict filter 335 vs coarse 323, and\n    filter-only-with-no-ranking 447). LOOSE=0 restores the strict variant.\n\n(B) FREQUENT-VOCABULARY LOG-ODDS. For each register r, fit a word unigram\n    distribution P_r on that register's dev documents over the vocabulary of\n    words occurring >= 3 times in the target (|V| ~ 16k), and score\n\n        score_r(d) = mean over words w of d  [ log P_r(w) - log P_pool(w) ]\n\n    where P_pool is the pool's own unigram distribution over the same\n    vocabulary -- i.e. the log-odds that w came from register r rather than from\n    generic web text. Out-of-vocabulary words contribute a fixed OOV_ODDS.\n\n    Both design choices are load-bearing, and both were arrived at by measuring\n    the failure of the alternative:\n      * RELATIVE, not absolute. Ranking by generative likelihood under the\n        target LM (MODE=lmfit) scores 554.8 -- worse than random -- because the\n        most probable text under any smoothed word LM is bland high-frequency\n        filler, not target-like prose.\n      * BOUNDED vocabulary. The same log-odds idea over unbounded hashed\n        n-grams scores 1124.7, because a log-count ratio gives enormous weight\n        to features that are rare in the pool, putting plant-genus lists and\n        PGP key blocks at the top of the ranking.\n    OOV_ODDS has a genuine interior optimum (0.0 -> 337.6, -1.0 -> 322.8,\n    -5.0 -> 346.6): too little penalty admits gibberish, too much admits only\n    common-word filler.\n\n(C) BALANCED INTERLEAVED SELECTION. Round-robin over the four registers, each\n    taking its next best unclaimed document, until every register has\n    contributed ~1/4 of the over-provisioned budget; then append all remaining\n    ranked documents as overflow so the packer can always reach 12M tokens.\n\n    Held-out perplexity on an equal-parts target is a token-weighted average of\n    per-register losses, so it is dominated by the WORST-covered register.\n    Collapsing to one global ranking (MODE=global) scores 665.3, and the damage\n    is not uniform -- it is concentrated in the registers least like generic web\n    text (technical Q&A degrades 7.4x, web prose only 1.05x). See claim.md.\n\n    Interleaved rather than register-blocked, because the packer consumes ids in\n    order and truncates at 12M tokens: interleaving keeps the mixture balanced\n    at every prefix length, so the selection is robust to token-estimate error\n    and to a different budget than it was tuned at.\n\nReproduce:  PYTHONHASHSEED is pinned internally; running this file with no\nenvironment overrides regenerates submission/selection.json bit-for-bit.\nAblations used above are reachable via MODE / LOOSE / NOFILTER / OOV_ODDS /\nMINCH environment variables.\n\"\"\"'''\ns = new + s[s.index('\"\"\"', s.index('\"\"\"')+3)+3:]\nopen(p,'w').write(s)\nPY\nhead -5 /workspace/submission/curate.py; python -c \"import ast;ast.parse(open('/workspace/submission/curate.py').read());print('SYNTAX OK')\"",
  "description": "Rewrite curate.py docstring to match final method"
}

{
  "command": "python submission/curate.py >/tmp/final.log 2>&1 && tail -1 /tmp/final.log && python -c \"\nimport json\ns=json.load(open('submission/selection.json')); r=json.load(open('/tmp/sel_L1.json'))\nassert s==r, 'MISMATCH'\nassert len(s)==len(set(s)) and min(s)>=0 and max(s)<182016\nprint('VALID: %d unique ids in range, reproduces the evaluated selection'%len(s))\n\" && python pack_selection.py submission/selection.json /tmp/final.npy 2>/dev/null | tail -1 && ls -la submission/ /workspace/claim.md && echo TIME=$(cat .timer/remaining_secs)",
  "description": "Final end-to-end reproducibility and validity check",
  "timeout": 1400000
}