Cookbook — the instruments on your data

Five-minute recipes. Everything here uses strataq.toolkit (plain lists in, verdicts out) or the public HTTP API; the numbers come from the same gated, red-teamed machinery as the research findings, and every result carries its honesty warnings. Install:

pip install strataq

No Python? Every recipe has a curl twin against the live API (https://sage-labs.vercel.app/api proxies it), and a point-and-click twin at sage-labs.vercel.app/tools.

Is my market / ecosystem a landscape or a whirlpool?

You have a measured cross-response matrix — how each agent's action moves when another agent's incentives shift. Cost pass-through between firms is the canonical example; any small-shift response estimate works.

import strataq.toolkit as tk

# the actual Dominick's grocery estimates (finding F-0011):
read = tk.reciprocity_read(
    [[1.0697, 0.0028], [0.0005, 0.9685]],
    chi_se=[[0.02, 0.001], [0.001, 0.02]],  # your regression's standard errors
)
print(read.r)  # 0.0011
print(read.verdict)  # reciprocal (landscape-like) — 95% CI [...]
print(read.warnings)  # the fine print travels with the number

curl -X POST https://sage-labs.vercel.app/api/v1/toolkit/reciprocity \
  -H 'Content-Type: application/json' \
  -d '{"chi": [[1.0697, 0.0028], [0.0005, 0.9685]]}'

Read the warnings

Without chi_se the verdict is a point read of a noisy estimate and says so. Only the zero test of ℛ is λ-free; compare magnitudes only at matched conditions.

Is my time series irreversibly driven?

Any scalar series ordered in time — prices, order flow, load, sentiment. The test phase-embeds it (value bins alone are provably blind to loop irreversibility) and compares against reversible surrogates with matched persistence. This is the instrument that found the day-ahead electricity market's diurnal loop (F-0009) and certified retail category prices at-null (F-0011's companion scan).

verdict = tk.irreversibility_test(weekly_prices, n_bins=3, n_surrogates=200)
verdict.detected  # True: no reversible chain with these pair statistics does this
verdict.p_value

Power: ≥ 80% detection at n ≥ 300 on a known driven series; n ≈ 100 is underpowered and the API will tell you so. NaNs and constant series raise — the instruments never compute through bad input silently.

How payoff-sensitive are my agents?

You know the game's payoffs and observed choice frequencies; you want the logit rationality λ with a defensible interval:

est = tk.estimate_rationality(
    [u_row_player, u_col_player],  # one matrix per player
    counts=[[412, 95, 493], [301, 402, 297]],
)
est.mean, (est.ci_low, est.ci_high)  # calibrated 95% credible interval
est.warnings

Two honesty guards you will meet: the scale fold (λ is per payoff unit — rescaling payoffs by s rescales λ by 1/s; only the product is ever identified) and the flat-likelihood warning (on symmetric games the QRE can be uniform at every λ; the toolkit warns instead of quoting noise).

Dashboard for a game you can write down

tk.game_thermo([u1, u2], lam=1.5)
# -> alpha (harmonic fraction), R (reciprocity defect), epr (dissipation), verdict

Rock–paper–scissors reads whirlpool (α ≈ 1, positive dissipation); a coordination game reads landscape (α ≈ 0, zero dissipation to machine precision); real systems land in between — the calibration bracket is road-network 0 / Blotto 0.12 / RPS 0.69.

Going deeper

The facade is a thin layer: strataq.estimate.bayes exposes the posterior machinery and the EFE experiment-selection loop (run_campaign), strataq.thermo the exact dissipation meters, quench protocols and surrogate nulls, strataq.finite the solvers and the Hodge decomposition. Every unit's gate, artifacts and adversarial-review record are in the progress dashboard.