🦋 i04 dragonfly: chameleon, faster, more careful, remembering its best

📦 archived 2026-09-25, v2.17.0 · smoke test only · its ingredients were tested one at a time in echidna, its belief and caution returned in hippogriff

Question. Chameleon, but fast enough to run, exploring by a guided random walk with momentum, cautious where it is unsure ("worst case until proven otherwise"), and remembering the best it has seen.

Answer. About 5× faster, and it learned faster than chameleon early on, then degraded; two of its parts turned out not to work as designed. Paused in favour of testing its ingredients one at a time (echidna); its belief-and-caution ideas returned in hippogriff.

Design (agents/dragonfly.py), what differs from chameleon

• History as leaky integrators: exponential traces of \([z_t, z_t - z_{t-1}, a_{t-1}, r_{t-1}]\) at decays 0.5, 0.9, 0.99 (momenta of the latent). A linear recurrence, trained with associative_scan, parallel over time, instead of backpropagation through a GRU.
• Exploration is an Ornstein-Uhlenbeck walk per latent dimension, every step: \(\xi_d \leftarrow \rho_d \xi_d + \sqrt{1-\rho_d^2}\,\sigma_d \eta\), \(\rho_d = e^{-1/\tau_d}\). Its spread is \(\sigma_d\) whatever \(\rho_d\) (constant entropy) while it carries \(-\tfrac12\log(1-\rho_d^2)\) nats of momentum from step to step; \(\tau_d\) is how many steps a heading persists.
• A learned guide sets \(\tau_d, \sigma_d\) and ε from the history, the measured entropy of each latent dimension, a pessimistic value, and p, the current performance relative to a golden cache.
• Pessimism: Q and V are ensembles (4 heads on bootstrap masks); the search scores candidates by alignment plus the ensemble's lower bound, so likely-bad candidates are vetoed.
• Golden cache: the best experiences by reward and by return, at most one per latent ball (canonical), with self-imitation of their actions.

Smoke test (330k steps)

Time per update	10.7 s on a quiet machine (chameleon: 47 s, loaded)
Episodes ending early	49 % → 1 % by 140k steps, then back to 12-22 %
Mean reward per step	−9.4 → +0.41 (best) → −1.8 at the end
Candidates vetoed by the lower bound	15-40 %
Deterministic test episode, last evals	overshoots at once (return −20)

Outcome

• The cache position p saturated at 0. The cache holds only golden examples, so even its worst entry is excellent and every ordinary state falls below it; the guide never got a signal from p.
• The guide never moved: τ, σ, ε stayed at their initial midpoints (32.5, 0.5, 0.5), so the question of how τ relates to p or to the measured entropy was never answered.
• The veto was the part that visibly helped early on, which is why pessimism returned in echidna (tested alone) and hippogriff (as a safety check on a belief).

Records: results/i04_dragonfly/ (smoke-test progress and summary).