A catalogue where the entries have verdicts.

Opening range, initial balance, overnight range, prior-day H/L, Crabel stretch, Asia/London ranges, the pre-market range and the opening drive. 25 of our registry entries. Two laws established INSIDE this room: high-win/small-target dies on MAE geometry, and confluence has the wrong sign. Neither law says anything about the rooms below.

Sweep → FVG / IFVG / breaker / CISD / order block. 15 entries, one beta survivor (gold reclaim); every fade variant loses out-of-sample. §63 added the fourth law: test the fill, not just the signal. §73b tested drive → FVG with controls: the gap scores what any limit at the same depth in the leg scores. §81 tested the first real breaker block: it adds nothing over a bare swing break and never beats a coin-flip direction.

Using a second index to gate a trade in the first. The AND-gate lives here too but is filed under session-level-break because its signal is a level break. Confirmation has the WRONG sign and the mechanism is known: confirmation is a proxy for 'the complex has already moved', so the fill is worst where the signal looks strongest.

The overnight gap as a statistic and as a trade. Both of our entries are stat-only: size predicts the fill, nothing predicts the direction, and the tradeable version is ~breakeven.

Varying ONLY the exit. Tested 2026-09-12: 17 of 18 longer-hold cells fail; the one survivor holds nothing overnight (17:00 Globex close instead of 16:00 cash close). Any import that holds a position past the session close lands here, and a $2,000 trailing drawdown is the binding constraint on all of them — 44.2% of overnight gaps exceed a whole $2,000 buffer on one NQ.

Short weeks, turn of month, OPEX, roll week, day-of-week × session, half days. Orthogonal information — it does not depend on a price pattern, so it cannot fail for either of our two laws. Our two entries are TIME-OF-DAY folklore (the London fix, the 10am reversal, an hourly mean-reversion clock) and both are rejected. The CALENDAR grid — week shape, turn of month, OPEX, events — has never been run.

Running several rules together. Coverage is `tested` with ZERO registry rows — §62 is a study over results we already owned, not a strategy with its own expectancy. The answer was that combining our survivors COSTS size rather than buying it, and that the diversification benefit which did appear was luck (P(null <= observed drawdown) = 0.51 for the pair that matters). Almost nothing in the import lands here, because a public strategy catalogue is a list of single rules by construction.

Two legs, traded as a spread rather than used as a filter. One tested entry (NQ/ES spread mean reversion, rejected), which is a real test of the shape but on exactly one pair. Every 'arbitrage / hedging' import lands here, and most of them are not really spreads: 46 of 69 state no market at all and the titles are ordinary indicator strategies carrying a hedging flag. Treat the count as an upper bound on how much genuine relative-value work is in the import.

NR7/NR4, inside days, consecutive narrow ranges, realised-vol percentile, squeeze → expansion. Tested 2026-09-12 as the one corner of the geometry our two laws do not forbid — low win rate, tight stop, very large target. 60 cells on NQ/ES/GC, all printed, nothing promoted. The claimed 20–30% win rate REPRODUCES (26–29% with a ¼-IB stop and an 8R target); the ~8R payoff does not, because an unstopped intraday position reaches 8R on 0.0–5.0% of qualifying days before any cost. An 'expanding market' label computable at 09:30 is indistinguishable from a random relabelling of the calendar (P = 0.213 on §58's screen statistic) and changes sign by instrument. The NR premise is half true: relative travel does improve on compressed days, and the stop that expresses it shrinks with the same range, which cancels it. Any import whose edge is a volatility state rather than a price pattern lands here.

Delta, imbalance, absorption — from the Databento order-flow file already sitting unopened in the cache. Distinct from the `volume-flow` family: OBV and a volume histogram are computed from the same OHLCV bars we already have, while this needs a different Databento product. ➜ §95 (2026-10-20): the 4-day file opened — impact real, prediction nil — and a side-convention bug fixed; a real family test needs 12–18 months of trades/tbbo (content/orderflow-plan-2026-10.md).

Rates, DXY, oil, bonds as a REGIME LABEL rather than a confluence filter (confluence has the wrong sign). ➜ FIRST TEST 2026-10-27 (§101): the dollar (6E) at the entry minute as a label on the §57b gold reclaim — it does not sort the trades (agrees leads at 1–5 min, trails at 15; a random day's dollar tape makes spreads that size 28% of the time; no skip filter beats its shuffle). Coverage is `token`: one label on one rule. Gold is a real-rates instrument and we have still never looked at a rate.

FOMC / CPI / NFP as the trade rather than as a stand-down. We have an econ calendar and have only ever used it to stay out. NEVER TESTED.

N consecutive closes, ±2σ from a moving average, distance from VWAP in σ, daily RSI extremes. Every fade we have rejected was at a LEVEL; this is a different population and 'fading loses' has not been shown here — which makes it the most valuable never-tested room on the agenda, whatever the import's counts say. ➜ FIRST TESTS 2026-10-20 (§83 AM TBR σ-band fade, §92 daily RSI(2)): both rejected as trades. ➜ §100 (2026-10-27): the vendors' percentile 'probable bands' calibrate only after dividing by ATR20 (i.e. they are ATR × a constant); fading the 20%/10% band and targeting the 50% band are both rejected (0 of 96 cells).

The last 30 minutes as its own trade; MOC imbalance. Gold's post-close window carried +950 of its +2,122-point 3-year move. NEVER TESTED.

SMA/EMA crosses, Supertrend, Ichimoku, Parabolic SAR, ADX/DMI trend filters. THE LARGEST FAMILY IN THE IMPORT BY A WIDE MARGIN, and coverage is `token`, not `tested`: our two entries are a 2026 regime expression (2×200EMA + VWAP, rejected because it expresses one year's trend) and a vendor black box (WillyAlgo). Neither is a pre-registered test of 'does a moving-average cross carry an edge on NQ/ES/GC'. Counting them as coverage would be the flattering arithmetic we tear other people apart for.

RSI / Stochastic / MACD / CCI / Williams %R / ROC used as a directional trigger. Distinct from `statistical-extremes`, which fades the same oscillators at an extreme rather than following them. NEVER TESTED — we have not run a single oscillator strategy in the whole registry, and the import is full of them.

Bollinger, Keltner, Donchian. Sits between two agenda rooms — a squeeze is `volatility-expansion`, a band touch fade is `statistical-extremes` — and the dump's tags do not say which, so it keeps its own family rather than being silently assigned to one. NEVER TESTED.

VWAP and its bands, as a trend filter or a reversion target. Tested 2026-09-15 (§73) as the reversion target — fade the first ±2σ / ±1σ band touch back to VWAP, stop 0.5σ beyond — and it loses on all three instruments because the stop the band implies sits inside the cost. The same section measured VWAP, the session VWAP and every band as LEVELS: at first touch VWAP and its ±2σ bands are NO EDGE AT THIS LINE on every instrument. The TREND-FILTER half of the family (long above VWAP, short below) is still untested; it appears in `ema-vwap-scalp` only as half of a compound rule.

Volume histograms, OBV, MFI — computed from the OHLCV bars we already hold, unlike the `order-flow` family which needs a different data product. Coverage is `token` ON PURPOSE: §73 tested the volume PROFILE member (prior-day POC as a magnet — it does not beat its own reflection — and POC/VAH/VAL as levels, NO EDGE on nearly every row), but OBV, MFI and volume-oscillator rules, which is what the import's strategies in this family actually are, have never been run. Counting one profile test as coverage of the family would be the flattering arithmetic this page exists to refuse.

Fib retracements/extensions, floor pivots, standard-deviation projections. Coverage is `token`: our one entry is filed `untestable` because it was taught without a mechanical spec, so it produced no verdict about the family. The import's versions DO carry specs, which makes this family testable in a way our own entry never was.

Averaging into a losing position on a ladder, with or without doubling. NEVER TESTED — and worth saying plainly on the page: these are position-sizing schemes, not edges, and a $2,000 trailing drawdown is the exact account that a martingale is designed to blow up.

Scheduled accumulation and portfolio rebalancing. Not an intraday futures shape at all. NEVER TESTED.

Quoting both sides for the spread. NEVER TESTED, and honestly out of reach: it needs a fill model we do not have and a venue relationship a retail trader does not have.

Sub-minute holding periods. NEVER TESTED as a family. Our cost model (NQ 1.5pt round trip) would dominate the result, which is itself the finding worth publishing about it.

Anything advertising a model. NEVER TESTED. The overfitting problem here is the one the whole EdgeTape method exists to catch, so this family is more interesting as content than as a lead.

Re-barring the price series before signalling. NEVER TESTED — and flagged, not neutral: both transforms smooth using information that is not final until the bar closes, which is the same class of defect as §59c's sub-minute lookahead. Any test of this family needs `scripts/test-no-lookahead.py` pointed at it first.

No usable type or indicator tag and no decisive title phrase. NOT A FAMILY — a holding pen, and it is excluded from the never-tested ranking for that reason: 'we have never tested the unclassified family' is not a statement about anything. An entry only leaves it by someone reading the source and writing a rule statement. Its size is a measure of the import's tagging, not of a gap in our research.