# Registered controlled feature-availability test v1

## Question

The second-source study found that missing 60-day volatility triggered the registered 1x fallback on 716/2,674 Binance daily decisions since 2019, versus 60/2,674 reference decisions. Since 2023 there were zero such Binance decisions versus sixty reference decisions. Test the historical effect of those different available feature observations without changing price values or execution observations. Do not tune a gap rule or choose a profitable mask.

## Unchanged inputs and accounts

Reuse the exact ten models, two native price sources, fourteen windows, costs 6/30 bps, execution delays one/three four-hour bars, borrowing APRs 5/10/20%, daily decision cadence, initial allocation, 0.25-1.50 exposure range and hypothetical liquidation/insolvency accounting from VOLATILITY_VENUE_PROTOCOL.md. USD reference and USDT Binance remain separate native quote accounts without FX conversion. All old inputs, helpers, cases and traces remain immutable. No change to price sanitization, borrow terms, cash yield, missing-fill expiry or liquidation observations.

## Feature-only intervention

Create two paired source histories, one for the historical account windows and one for the recent replay. Historical pair: frozen canonical BTCUSD rows through April 28, 2026 at 00:00 UTC, and Binance rows through that same closing bound. Recent pair: frozen recent-market-cache BTCUSD warmup through September 8, 2026 at 00:00 UTC and Binance through that bound. Retain full native warmup before account slicing.

Shared coverage starts at the later of the two already-frozen first opening timestamps (August 17, 2017 at 04:00 UTC). The end bounds above are fixed before evaluation. Within that common coverage, a four-hour price observation is eligible for volatility calculation only when both sources contain that exact opening timestamp. The mask depends on timestamp availability, not price values, returns, volatility levels or account outcomes. Native warmup observations before shared coverage are retained; all evaluated accounts start more than sixty days after shared coverage begins. Neither source is extended beyond its registered end.

For each source, calculate the unchanged sample log-return volatility on its own prices at eligible timestamps only. An excluded observation creates a gap in that feature input, clearing the contiguous return window under the original helper. Map feature values back onto every original native account row; an excluded row has missing volatility. At each original native midnight decision, missing volatility sets target exposure to the original 1x fallback, including when that day's observation was withheld only from the feature calculation. Between midnight decisions carry the desired target exactly as before. Initial allocation uses the fully warmed target at the account's first row.

Preserve every original native row for accounting, ordinary trade fills, price marks, interest elapsed time and low-based liquidation checks. This deliberately withholds information only from the volatility calculation, even if a price remains available for execution. It is an artificial availability intervention, not a realistic publication-availability claim or a proposed dual-exchange trading strategy. A native missing execution row remains missing. The shared feature mask does not standardize execution grids, quote instruments or price levels.

## Matrix and references

Run 3,360 new shared-mask account cases: two sources x fourteen windows x ten models x two costs x two delays x three APRs. Pair each with its exact native case from experiments/volatility-venue-v1/report.json. The 3,360 native cases are reused, not counted as new backtests. Verify all native report/trace hashes from prior provenance. For the four fixed/hold controls, masking features must leave all 1,344 new curves exactly equal to their native references.

Reconstruct the no-mask target path from the new feature-to-native mapping and require exact agreement with the frozen signals helper on both historical and both recent source inputs. At shared observed timestamps from January 2018 onward, verify equal missing/nonmissing status for each 14/30/60-day feature across each pair. Price-derived volatility values need not agree. Verify future price and future mask mutation leave earlier features, desired targets and account prefixes unchanged; scale prices and confirm exposure/account invariance.

## Evidence and interpretation

Retain all new feature/target/account traces, original case identifiers and complete paired results, including insolvency, forced sales, costs, interest, annual failures and drawdowns. Report native versus masked return and drawdown for every model/source/window/assumption. Include final log-wealth differences where both accounts remain positive; leave them unavailable otherwise. For all six volatility candidates report wins against own-source BTC and against BTC plus both fixed 1.25x/1.50x controls in the four core windows, with native and masked counts side by side. Counts are descriptive, not independent trials or p-values.

Primary settings remain 10% APR, 6 bps and four-hour delay. Quantify whether the 2023-onward cross-source gap narrows under shared feature availability and whether higher-return claims survive. A within-source return change is the effect of this specified historical information intervention under the simulator assumptions; it is not a forecasting edge, a live outcome or proof that masking data is desirable. Any new gap-handling rule needs its own later preregistration. No production or current forward-registry changes or promotion.