First data contact — electricity prices through the dissipation meter

The estimators were built synthetic-first (gates thermo.estimators, estimate.lambda); domains.electricity is where they first touch a real market. The loader pulls CAISO OASIS locational marginal prices — day-ahead hourly and real-time 5-minute — with disk caching, a 90% coverage guard (silent gaps are how bad readings happen), and rate-limit backoff. ERCOT was the plan; its "public" endpoints reject programmatic clients (HTTP 403), so CAISO — genuinely open, no key — is the first data domain.

The question, and the honest answer (F-0008)

Is a market price series time-reversible (an equilibrium-like fluctuation) or measurably driven (a cycle that dissipates)? Three methodological facts had to be established before the answer could be trusted:

1. Price-value discretization is structurally blind to loop irreversibility. A periodic drive retraces the same 1-D price path up and down; its time-reversal visits identical value transitions. Proven in a unit test (a noisy 24h sine reads 6×10⁻⁵ in value space, 0.03 in phase space) and then observed on the real data to four decimals: the DAM value-space reading sits exactly on its null (0.0121 vs 0.0121).
2. The embedding that sees loops is the standard position/velocity trick: states = (price bin, Δ-sign). A loop stops retracing itself there.
3. The null must share the spectrum. A plain shuffle is wrong twice — it destroys persistence, and the Δ-sign embedding of an i.i.d. series is structurally asymmetric (0.15 nats of pure artifact). The correct null is the phase-randomised (FT) surrogate: identical power spectrum (persistence and the diurnal peak), time-reversible by construction.

Even that is not enough for heavy tails: LMPs carry excess kurtosis ≈130, plain FT surrogates Gaussianize it, and the red-team rejected them. The working null is the amplitude-adjusted (AAFT) surrogate — spectrum AND marginal preserved — with detection requiring exceedance of both null classes, and an explicit below-band flag when the reading escapes the null on the low side.

The readings (scoped to this n, 6 bins, and this embedding): SP15 DAM hourly is at-null — inside the AAFT band, consistent with a linear time-reversible process. RTM 5-minute is a no-detection with the below-band flag raised: the reading sits 5–6× below the null median, meaning no linear process with the data's spectrum and marginal can reproduce its Δ-sign persistence — neither a detection nor a certified null, and the persistence anomaly itself is now a chase item (bracketing surrogates; ramp-constraint mechanism). Finding F-0008, as revised after adversarial review — the retraction of the original "certified null" wording is part of the record.

The instruments do not hallucinate irreversibility on real market data, and they do not manufacture certainty where the null class fails.

Data source: CAISO OASIS (public API, no key; used per its posted terms — informational reports, paced requests with backoff, local caching).

Artifacts: electricity_irreversibility_dam.json, electricity_irreversibility_rtm.json (regenerate: make reproduce; the CAISO cache makes reruns cheap and identical).

The reciprocity meter's first empirical read (unit domains.pricing, F-0011)

From the Dominick's canned-soup panel (HF qbz506/dreamprice-dominicks-cso; single-chain scanner data): own-cost pass-through 1.07 (Campbell) and 0.97 (Progresso); cross terms 0.003 and 0.0005 with the asymmetry CI covering zero; ℛ = 0.0011 [0.00005, 0.005] by cluster bootstrap over 86 stores, re-demeaned within every resample (the within transform is part of the estimator, so it resamples too). The ex-ante prediction — stated in config before the run — was exactly this: one retailer pricing both brands toward one category objective must respond symmetrically, so the meter should read ≈ 0. It does. The multi-agent reciprocity question needs cross-chain data; this dataset cannot ask it, and the finding says so. Companion scan: 0/30 stores show Edgeworth irreversibility in weekly category indices vs the reversibilized-Markov null (0.3 false positives expected) — at-null, conservative under the sample's week gaps.

The trading day as a repeated quench — what a real market read actually supports (unit domains.electricity.quench, F-0017)

Two adversarial rounds turned this unit's headline over twice, and the third version is the narrow one. What survives: the closed daily-loop affinity of 211 CAISO DAM day-pairs, ≈ 7.0 nats/day (day-block bootstrap [6.22, 8.50] — the first framing half-counted the circular protocol and treated serially-correlated days as independent, both corrected), quoted as descriptive because 6 states extends rather than inherits the estimator's 4-state certification and because the fluctuation-theorem companion cannot constrain a closed cycle.

What does not survive: the seasonal-drift story. A seeded shuffle control — re-read each month with day order randomised, destroying temporal drift while preserving thinness and price distribution — attributes 0 of 7 months to drift, and one month's anomaly flag flips under a permutation that changes nothing physical. At n ≈ 30 days a monthly read is sampling noise, so it supports no causal statement in either direction. Answering "is the trading day one repeated quench?" needs a small-n extension of the certification or coarser windows over far more days; the honest present answer is that this design cannot tell. Artifact: day_quench_read.json (the shuffle control is a permanent metric, reusable by any future repeated-protocol claim on real data).