strataq

Thermodynamic instruments for strategic systems. Quantal-response equilibria, potential/harmonic game decomposition, entropy-regularised response, and non-equilibrium strategic dynamics — in JAX, with a plain-numpy toolkit on top.

Some strategic systems settle (landscapes); some circulate forever (whirlpools). From the outside they can look identical. strataq is a set of calibrated meters that tell them apart — from a payoff matrix if you have one, from measured responses or a raw time series if you don't.

import strataq.toolkit as tk

Three questions, one call each

Is my system a landscape or a whirlpool? — from any measured cross-response matrix (e.g. cost pass-through between two firms):

read = tk.reciprocity_read([[1.07, 0.003], [0.0005, 0.97]])
read.r  # 0.0011 — this is the actual Dominick's grocery reading
read.verdict  # "reciprocal (landscape-like): ..."
read.warnings  # the honesty notes travel with the number

Is my time series irreversibly driven? — the instrument that found the diurnal loop in day-ahead electricity prices:

verdict = tk.irreversibility_test(weekly_prices)  # any scalar series, ≥ a few hundred points
verdict.detected, verdict.p_value  # vs a reversible null with matched persistence

How payoff-sensitive are my agents? — a Bayesian posterior over the logit rationality λ from observed choice frequencies:

est = tk.estimate_rationality([u1, u2], counts=[[412, 95, 493], [301, 402, 297]])
est.mean, (est.ci_low, est.ci_high)  # calibrated 95% interval
est.warnings  # flags flat likelihoods instead of quoting noise

And for a game you can write down, the full dashboard:

tk.game_thermo([u1, u2], lam=1.5)  # harmonic fraction α, reciprocity ℛ, dissipation

The full library

Under the facade: three QRE solvers (damped fixed-point, mirror descent, arclength branch tracer, pygambit-validated), the Hodge potential/harmonic decomposition, equilibrium response matrices χ = (I − SB)⁻¹S, exact Glauber dynamics with currents and entropy production, trajectory estimators (KLD, certified TUR bounds), Hatano–Sasa quench thermodynamics, surrogate nulls, Bayesian estimation with an EFE experiment-selection loop, and domain plugins (Colonel Blotto, traffic networks, electricity markets, retail pricing).

Every claim the library makes is backed by a gated unit with an adversarial review on record, and every number regenerates from fixed seeds (make reproduce in the SAGE monorepo). The interactive companion lives at sage-labs.vercel.app.

Install

pip install strataq        # once published; until then (pip ≥ 21.1):
pip install "strataq @ git+https://github.com/SharathSPhD/sage.git#subdirectory=packages/strataq"

Python ≥ 3.11. CPU JAX by default; float64 is enabled by the library.

import strataq — never import sage (SageMath owns that name).

License

Apache-2.0. Dominick's-derived artifacts carry CC-BY-NC-4.0 (see strataq.domains.pricing).