Rules frozen before data, results published regardless — the earlier registrations and their outcomes, public like the current page.
CECadence's validation rules are pre-registered: each rule is frozen in a dated, committed protocol before the data it will be judged against is fetched, and the results are published whatever they say. This page is the public record of that discipline. The full protocol text — rules, parameterisations and adjudication procedures — is part of the licensed diligence pack: request it at hello@compoundingenergy.com.
Every pre-registration below is a git commit in the product repository, made before the corresponding data fetch; a continuous-integration guard fails the build if held-out data appears without its pre-registration. On request we walk diligence teams through the commit history — timestamps and content hashes cannot be edited after the fact. And the record includes failures and rejected model changes: a vendor curating its results would not publish these.
This is the archive half of the public record: the 24 registrations frozen before 2026-08-31 (42 days before the most recent registration, 2026-10-12), each with its outcome, in the order the record lists them. The current record carries the rest; together the two pages are the whole record, rendered from one committed source and split by registration date, never by hand. Nothing here is edited when it moves — a row reads on this page exactly as it read on the current one.
| Pre-registered rule | Frozen (commit · date) | Outcome |
|---|---|---|
| 2025 held-out validation year — every rule frozen before any 2025 backtest input was fetched | 8667ce5 · 2026-06-12 |
Measured FAIL published unmodified: annual mean +12.0% on the default basis vs the 5% target; the P95 scarcity tail fails in both profile modes |
| Scarcity-knee re-shape — adoption rule fixed before fitting | 623f126 · 2026-06-14 |
NOT ADOPTED under its own frozen rule (an honest negative) |
| Richer curve family (regime-scaled knee) | 838d3c2 · 2026-06-15 |
NOT ADOPTED — the pre-registered generalisation guard vetoed a clean-looking in-sample fit |
| Numerical-correctness audit — engine fixes adjudicated against the held-out record | 5a2f458 · 2026-06-15 |
Adjudicated held-out-safe; evidence regenerated, rules unchanged |
| Piecewise scarcity-tail rule | 8edaf0c · 2026-06-16 |
Fitted and NOT ADOPTED; the exercise localised the tail failure |
| Inner-knee re-shape | c9aca57 · 2026-06-17 |
NOT ADOPTED — the most informative negative in the record |
| Standing quarterly out-of-sample revenue grading (every new index quarter grades the frozen revenue parameterisation) | 7a60d5f · 2026-07-05 |
Standing; two out-of-sample quarters on the record — 2026Q2 (computed 2026-07-20) and 2026Q3 (computed 2026-10-08, graded before the advance and recalibration could use it) — both informational per the frozen convention; the graded test begins at four quarters and, once started, is never restarted |
| Held-out 2026H1 price window — 13 season-stratified weeks, frozen checks | 7a60d5f · 2026-07-05 |
EXECUTED (7678204): synthetic window mean 1.4% PASS (90% CI 0.2–6.0% spans the 5% target — a fragile pass, disclosed), P50 23.0% FAIL, P95 55.6% FAIL, published unmodified. The realised-basis leg was NOT OBTAINABLE at execution (profile coverage below the frozen gate; the gap published verbatim), completed in 2026-07 under the same rules once coverage crossed it: realised mean 7.8% FAIL vs 5%, P50 26.3% FAIL, P95 54.5% FAIL |
| Six revenue-stack adoption rules — zonal-BM honesty (thin-zone shrinkage, two-leg structure with no-flattery checks), capacity-market walk refit, frequency-response demand trajectories, a dispatch-efficiency instrument and a variance-refit data-gate — every rule frozen before any fitting code existed | 4ea6671 · 2026-07-06 |
EXECUTED: three ADOPTED under their frozen gates (zonal shrinkage and the two-leg BM structure — both lower or hold every Scottish-boundary value — and the capacity-market walk refit); three NOT ADOPTED (demand trajectories failed their identification clamp; the dispatch-efficiency multiplier stays data-gated; the variance refit missed its fit gate by 0.005) — published either way |
| Fifth curve re-shape (regime-identified) — fitted at five windows, both executed held-out windows included | 600060b · 2026-07-19 |
NOT ADOPTED (measured same day, b071171): the candidates that fix the chronic hot tail break the crisis-year and precision-mean legs, so the frozen no-regression gate refused every one; the regime-width hypothesis was refuted |
| Dispatch-efficiency instrument re-adjudicated — the six-rules instrument re-run when newly pinned data made its measured leg computable | de5d7d0 · 2026-07-24 |
NOT ADOPTED (stability-gated): the gate's computability part now passes; its cross-quarter stability bound still fails — the instrument stays published and is applied to no revenue leg |
| 2026Q2 window-advance target recalibration — the recalibration rule frozen before the new window's measured target existed | b224ac9 · 2026-07-20 |
APPLIED 2026-07-21 (the next day, per the frozen no-same-day rule): all twelve measured target cells moved, each reproduced verbatim from the committed caches before application; two graded verdicts flipped, published as flips, not absorbed |
| 2026Q3 index re-basing — frozen rules, free data only, dual-published, never retroactive | 60c5c84 · 2026-07-23 |
APPLIED 2026-10-09 at the 2026Q3 window advance: the index moved to 2025Q4–2026Q3 on the re-based basis (a behavioural membership screen, wider battery recognition, measured duration classes, a wholesale floor netted of energy already paid through balancing actions, capacity payments in delivery-year money, part-quarter counting for new units, response and reserve shown separately), the previous basis published in full beside it; the earlier quarters' first restatement on their own basis, beside the record, never over it |
| Model recalibration to the re-based index — frozen before any target exists to fit | be5e6a0 · 2026-07-23 |
EXECUTED 2026-10-09, inside the window advance's release: the balancing and frequency-response lines re-fitted to the re-based index (frequency response to its response services only), capacity prices in delivery-year money. Graded as they landed: balancing 2h passes, 1h and the thin 4h class (two units) fail; frequency response is in-sample by construction, badged so; the cycles check stays REJECTED — published either way |
| The living index — provisional series, strategy mix, per-asset publication | a42f055 · 2026-07-24 |
Registered before any series existed; executed — the provisional surfaces are live, additive and ungraded (the settled quarterly remains the record); re-based with the index from October 2025, refreshed daily since 2026-10-08 |
| Battery ageing on the bankable candidate track | 2e83bb4 · 2026-07-24 |
Registered before any implementation code existed; executed — the multi-year candidate track ages the battery (conservative by construction, user-overridable) |
| FR structure — the auction-shaped response book | 32e1a01 · 2026-07-27 |
Registered; executes at/after the model recalibration above — deliberately not yet in code |
| CM leg — cleared where cleared, expected where not | df6ce85 · 2026-07-28 |
EXECUTED 2026-07-29: delivery years whose T-4 auctions have cleared now price at the final clearing prices (auction results are facts, not forecasts); beyond the cleared horizon the central line is the deterministic expectation of the frozen bounded walk, not a seeded realisation |
| Candidate CM agreement lock — a capacity agreement is priced once, at the auction that awarded it | 3c052b5 · 2026-08-28 |
EXECUTED — ADOPTED 2026-08-29: a new-build candidate that clears a capacity auction holds that agreement's price for its full term instead of re-pricing each year against a moving forecast |
| Class balancing-mechanism value at the fleet's real zone mix — capacity-weighted, not single-zone | 13d9d84 · 2026-08-28 |
EXECUTED — ADOPTED 2026-08-29: a class is aggregated across the zones its fleet occupies, weighted by capacity; the single-zone proxy is retired |
| The competing fleet ages too — the bankable-track ageing law applied to every cohort | 6a789b8 · 2026-08-29 |
EXECUTED — ADOPTED 2026-08-30 under a graded evaluation: no graded cell changed verdict and pooled accuracy was unchanged to five decimal places — adopted for consistency with the candidate track, not because it improved the record |
| One charge for cycle wear, not two | bf5b928 · 2026-08-29 |
EXECUTED — ADOPTED 2026-08-30: cycle-wear economics are carried once, by the registered fade law, rather than charged a second time in candidate finance |
| Standing loss at cell physics — self-discharge re-registered to the modern-cell band | e7c8a7c · 2026-08-30 |
EXECUTED — NOT ADOPTED 2026-08-31 (fbec34c): the more physically accurate value missed its own frozen accuracy gate by a hair, so the frozen rule stands and the code was reverted — an honest negative: the graded record preferred the less realistic physics |
| The fingerprint attests what the evidence reads | c53096e · 2026-08-30 |
EXECUTED — ADOPTED 2026-08-31: the evidence fingerprint enumerates the graded inputs it covers rather than excluding a list, so a newly added input is covered by default, not silently uncovered |
Five re-shape attempts have been pre-registered, fitted and rejected under their own frozen adoption rules. The persistent P95 scarcity-tail failure is a data limit — one crisis year in the historical record — not a knob to be quietly tuned; the registered unlock is a second genuine high-scarcity window. The full diagnosis is in the held-out note; the graded verdicts render live on the methodology tab beside the measured CE GB BESS Index.
The complete document of record: frozen rule text and parameterisations, exact run plans, fuel-multiplier derivation rules, adjacency disclosures (what the model could have "seen"), the publication rule, and the appended measured results of every section whose convention lands them there. Every executed rule's outcome is on the record either way — as an appended protocol result, or in the dated commit and changelog record the table above indexes. Signed-in users (free account included) can read the validation dossier and limitations register; email hello@compoundingenergy.com for the protocol walk-through.