Upload folder using huggingface_hub

README.md

@@ -32,6 +32,22 @@ python scripts/check_budget_bound.py
 python scripts/run_eval.py --task trip --method ccd_ds --limit 64 --out outputs/x.json
 ```
 
+## Determinism / seeds
+
+| component | status |
+|---|---|
+| `check_propositions.py`, `check_budget_bound.py` | Seeded (`np.random.default_rng(0)`). Bit-identical across runs — verified 3x. |
+| Example selection (`load_trip`) | Seeded, fixed at `seed=0`, independent of `--seed`, so every arm scores the same stratified sample. |
+| Decoding at **T=0** (Trip Plan, HumanEval, all ablations) | Deterministic: `_pick_token` takes the argmax, no RNG involved. |
+| Decoding at **T>0** (the Claim 6 sweep, T=0.1/0.4/0.7/1.0) | Samples via `Categorical.sample()`. `run_eval.py --seed N` (default 0) now seeds `random`/`numpy`/`torch`, and the seed is recorded in each result JSON. |
+
+**Honest caveat:** `--seed` was added *after* the runs in `outputs/` were produced. The T=0
+results (the large majority, including every number on the poster's ceiling, Trip Plan,
+HumanEval and ablation cards) are argmax and reproduce exactly regardless. The eight
+temperature-sweep runs at T>0 (`c6_*_t0.1/0.4/0.7/1.0`) predate the seeding and therefore will
+**not** reproduce bit-exactly; re-running them under `--seed 0` is the first thing to do with a
+fresh GPU budget. At n=16 those points are our weakest evidence anyway, and the poster says so.
+
 ## Hyperparameters
 
 Taken from the Dream authors' own eval scripts, which is what the paper says it does

poster/GATE_REPORT.json

@@ -1,7 +1,7 @@
 {
   "schema_version": 1,
   "skill": "posterly",
-  "timestamp": "2026-07-16T18:30:59Z",
+  "timestamp": "2026-07-16T18:41:34Z",
   "poster_html": "/home/ashuzon/Documents/github_projects/icml/repro_ccd/poster/poster.html",
   "canvas": {
     "source": "page-rule",

poster/poster.html

@@ -1030,7 +1030,7 @@
         <div class="section-title"><span class="num">5</span><span class="st-text">Theory: Prop. 1 &#10003;, Prop. 2 &#10007;</span></div>
         <p class="body-text">
           <strong>Prop. 1</strong> &mdash; <em>H(x|s) = H(x|c,s) + I(x;c|s)</em> &mdash; is a correct identity, exact to
-          <strong>machine precision</strong> (max |LHS&minus;RHS| ~2e-15 over 2000 random joints).
+          <strong>2.2e-15</strong> over 2000 seeded random joints.
         </p>
         <p class="body-text mt-3">
           <strong>Prop. 2 does not follow.</strong> Lemma 1 is an <em>upper</em> bound; Eq. (12) swaps its RHS for a

poster/poster_preview.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96106d4bf2f1c6255e7cf74e44ca5bd15292b6c8a29e1d4ce8e41ea8c4152258
-size 1561997
+oid sha256:76f40ec8cfe9ee7b4c0c1accf3e87170594405fea3d6b1cfc3500dd55987657c
+size 1561927

scripts/run_eval.py

@@ -11,9 +11,10 @@ paper says it does ("we follow the base models' default settings without tuning"
 Usage:
   python run_eval.py --task trip --method ccd_ds --limit 100 --out outputs/x.json
 """
-import argparse, json, os, sys, time
+import argparse, json, os, random, sys, time
 sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
 
+import numpy as np
 import torch
 from transformers import AutoModel, AutoTokenizer
 
@@ -53,7 +54,6 @@ def load_trip(limit=None, num_cities=None, seed=0):
         if limit:
             items = items[:limit]
     elif limit:
-        import random
         tiers = sorted({i["num_cities"] for i in items}, key=int)
         per = limit // len(tiers)
         rng = random.Random(seed)
@@ -168,10 +168,25 @@ def main():
                     help="override top_p; 1.0 disables nucleus filtering")
     ap.add_argument("--steps", type=int, default=None)
     ap.add_argument("--out", required=True)
+    ap.add_argument("--seed", type=int, default=0,
+                    help="RNG seed for temperature sampling. Only bites when temperature>0: "
+                         "at T=0 ccd_decode._pick_token takes the argmax, so those runs are "
+                         "deterministic with or without it. Does NOT affect which examples are "
+                         "evaluated -- load_trip's stratified sample is fixed at seed 0 so every "
+                         "arm scores the same examples.")
     ap.add_argument("--tiny", action="store_true",
                     help="local CPU dry-run with tiny random weights")
     args = ap.parse_args()
 
+    # Seed every RNG that can touch decoding. NOTE: the temperature-sweep results
+    # shipped in outputs/ (c6_*_t0.1/0.4/0.7/1.0) were produced BEFORE this was
+    # added and therefore do not reproduce bit-exactly from this seed; the T=0
+    # runs (Trip Plan, HumanEval, the ablations) are argmax and always did.
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.cuda.manual_seed_all(args.seed)
+
     cfg = dict(TASK_DEFAULTS[args.task])
     if args.temperature is not None:
         cfg["temperature"] = args.temperature
@@ -180,8 +195,8 @@ def main():
     if args.steps is not None:
         cfg["steps"] = args.steps; cfg["max_new_tokens"] = args.steps
 
-    print(f"[cfg] task={args.task} method={args.method} V={args.buffer_V} d={args.history_d} {cfg}",
-          flush=True)
+    print(f"[cfg] task={args.task} method={args.method} V={args.buffer_V} d={args.history_d} "
+          f"seed={args.seed} {cfg}", flush=True)
 
     tok = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
     if args.tiny:
@@ -245,7 +260,7 @@ def main():
     result = {
         "task": args.task, "method": args.method, "config": cfg,
         "buffer_V": args.buffer_V, "history_d": args.history_d,
-        "n_examples": len(items), "score": score, "offset": args.offset,
+        "n_examples": len(items), "score": score, "offset": args.offset, "seed": args.seed,
         "mean_steps": mean_steps,
         "speedup_vs_uniform": cfg["steps"] / mean_steps,
         "fallback_steps": n_fallback,