[2.0] Add new Frontier-CS 2.0 problem: nanoslm_hybrid_arch_design

[sijial430]

Summary

• Problem id: nanoslm_hybrid_arch_design
• Task: Design a hybrid language-model architecture (attention + a
  linear-recurrent sequence mixer) that reaches the lowest held-out
  bits-per-byte (val_bpb) when trained from scratch under a fixed
  wall-clock budget of 15 minutes on one H100. The agent submits a single
  model.py; the hidden judge drops it into a locked dolma2-BPE train+eval
  harness, trains it for T = 900 s (the LR cosine spans a full 6-hour horizon,
  so each run is the prefix of a long schedule), and scores it as
  score = 100·2^(−γ·val_bpb), γ = log₂(100/70)/ref_bpb. Perfect fit (0 bpb) = 100, the measured reference
  solution = 70 exactly, failures = 0.
• Iteration loop: submissions are asynchronous. agents can keep improving their next candidate while a run is being scored (~20 min wall-clock per result).
• Grounding: a hybrid architecture (following Olmo Hybrid, arXiv:2604.03444) which is a 3:1 GDN:attention interleave at the ~190M olmo3_190M shape (d=768, L=12,
  H=12), using dolma2 tokenizer.

Measured calibration

Architecture	@ 15 min (current: bpb → score)	@ 30 min (bpb)	@ 6 h (bpb)
Locked baseline	1.58648 → 69.49 (327 steps)	1.33325 (655 steps)	0.98112 (7,897 steps)
Reference 3:1 hybrid	1.55484 → 70.00 (264 steps)	1.31208 (496 steps)	0.93795 (6,544 steps)

The 15-min column is the scored configuration; earlier budgets are historical
calibration.

Validation

Commands (from repo root):

uv run frontier list 2.0
uv run frontier show 2.0 nanoslm_hybrid_arch_design
PYTHONPYCACHEPREFIX=/private/tmp/frontier-cs-pycache \
  python3 -m py_compile 2.0/problems/nanoslm_hybrid_arch_design/evaluator.py
# torch-free CPU self-test (policy + scoring + fingerprint + real-load fixtures):
python3 2.0/problems/nanoslm_hybrid_arch_design/evaluator.py --selftest
# Harbor smoke trial (see CONTRIBUTING.md; a scored run takes ~20 min at the
# 15-min budget, so give the verifier headroom):
uv run frontier harbor trial 2.0 nanoslm_hybrid_arch_design \
  -a codex -m <model> --agent-timeout 600 --verifier-timeout 1800 --force-build --json

Judge-side hardening (dual Claude+Codex review): spawn+poll Modal dispatch with
a config-fingerprint + code-version staleness gate, logits-width guard,
warm-container env hygiene, isolation audit (no experiment data or held-out
stream reachable from the agent container).

Validation logs with codex (gpt-5.5 one-shot + gpt-5.6 iterative campaigns)
(access needs to be granted):
wandb: https://api.wandb.ai/links/sijial-ai2/u86xnx57
results (updated, 15-mins interval): https://drive.google.com/drive/folders/1Xcyu0hJSh2p7tmMFNbY5SGkfFQNcovCS?usp=drive_link
results (old, 30-mins interval): https://drive.google.com/drive/folders/157qqq--nHlSpkkur7TT-HkCHBjaAFQdH?usp=drive_link

[sijial430]

Updated (b57c492):

1. New score formula: score = 100·2^(−γ·val_bpb) with γ = log₂(100/70)/ref_bpb: a smooth tempered per-byte likelihood. Perfect fit (0 bpb) = 100, the measured reference solution = 70 exactly, failures = 0, no manual capping; val_bpb always reported alongside.
2. Shorter iteration interval: training budget cut to 15 minutes per run (was 30) so agents get scored feedback in ~20 min wall-clock; the LR cosine keeps its full 6-hour horizon, so every run is the prefix of a long schedule rather than a compressed anneal.
3. Agents can iterate while the judge works: submissions are asynchronous (submit returns a UUID immediately; 2-slot queue; cancel supported), and agents are encouraged to keep improving their next solution while a run is being scored, mirroring the vllm problem's workflow.

[momoway]

LGTM!