The opponent brain of Parry
(HF × Gradio Build Small Hackathon): a real-time 1v1 footsies duel where a 1.5B
model runs in the player's browser, emitting one grammar-constrained intent
letter (S F P L R W) every ~160ms, conditioned on a natural-language
plan-string written by its own Analyst pass — the plan it shows you on screen
is the plan that steers it, and you can edit that plan live in the judge panel
and watch its behavior flip.
This checkpoint runs inside the submitted app: converted to
GGUF Q4_K_M, it
executes in-browser through llama.cpp's WASM runtime (wllama) as the Analyst —
reading your behavior log and writing the plans. The judge-override demo
additionally serves it through a classifier-free-guidance gateway (vLLM, two
parallel calls, uncond + γ·(cond − uncond) over the action letters).
Training
Behavior cloning (LoRA r=32 α=64, completion-only loss) on plan-conditioned duel
logs from a deterministic integer engine: on-trajectory expert play +
counterfactual same-state×6-plans expansion (the conditioning signal) +
Analyst-role rows. This final generation adds 20% plan-dropout so the model
learns an unconditional branch — enabling CFG at inference. Trained on Modal
(A100); no added tokens, so weights-only conversion reaches MLC/WebLLM and GGUF.
The six-generation arc (honest numbers)
Pre-registered gate before any training: plan-intervention ΔP ≥ 0.30
(mean pairwise total-variation across held-out states × held-out plan
paraphrases), ≥2× base, direction-correctness ≥0.70.
gen
examples
held-out agreement
ΔP (γ=1)
what we learned
pilot
20k
0.795
—
wiring works
bc_v1
98k
0.771
0.071
plan-DEAF (base: 0.129) — great state player, ignores the plan
bc_v2
153k
0.695
0.159
counterfactual data helps; no-plan ablation 0.643 proves the plan now carries signal
bc_v3
188k
0.735
0.224
disagreement-upweighting — best pure-BC steering
bc_v4
129k
0.739
0.078
engine-v4 retrain silently lost the steering — regressions are real
bc_v5
173k
0.724
0.123
plan-flavored expert + tiered contrast weights
bc_v6
173k
0.727
0.112 → 0.292 @ γ=4
plan-dropout + classifier-free guidance
The pre-registered ΔP≥0.30 gate never passed at γ=1 — and the eval explains
why: the expert itself is plan-invariant in 40% of states, putting the
identifiability ceiling at ΔP 0.330. No objective can recover steering the
teacher doesn't exhibit.
Classifier-free guidance closes the gap
Two parallel decodes (with-plan / plan-stripped), letter-logprobs combined as
uncond + γ·(cond − uncond):
γ
ΔP
argmax-flip (disagreeing states)
rush→turtle P(STRIKE)
1.0
0.112
0.179
0.30→0.07
1.5
0.150
0.215
0.29→0.05
2.0
0.187
0.260
0.29→0.03
3.0
0.249
0.326
0.28→0.03
4.0
0.292
0.385
0.28→0.03
γ=4 reaches 88% of the expert ceiling (0.292 / 0.330). In the app, guidance
stays at γ=1 during normal play (higher γ trades fight-IQ for obedience) and
jumps to γ=3 when a judge edits the plan — obedience is exactly what that demo
measures.
Raw reports for every generation — metrics, intervention tests, the CFG sweep —
are in evals/ (18 JSON files, nothing curated away).
