score_v4: What Six Months of Live Decisions Changed
2,069 live snapshots (Feb–Oct 2026) · 465 recorded engine decisions (Oct–Nov 2026) · 8-year reconstruction (2018–2026) · BTC + ETH · score_v3.2 → score_v4
RiskState Research · 16 November 2026
Abstract
score_v3.2 was frozen from 19 May to 19 November 2026 so that it could be judged on data it had never seen. This paper reports what that record showed and which changes it justified. The composite score has no detectable relation to forward returns (block-permutation p > 0.56 at every horizon, family-wise p = 0.98). Several score layers track the previous week of price (Spearman ρ up to +0.74), which made score-driven sizing pro-cyclical. The candidate we pre-registered in June, sizing by realized volatility, did not beat a constant position at equal exposure over eight years and was not adopted. A trend gate did: on BTC, 2018–2026, computed with the production module, return per unit of maximum drawdown is 0.61 against 0.37 for a constant position at the same mean exposure. The rest of score_v4 is correctness: a DeFi health factor that binds as a hard limit, a rules engine that counts each condition once, and sizing arithmetic that returns zero on any non-finite input, checked against a C port on 200,000 inputs and by an external automated-reasoning review.
Plain-language summary
We stopped changing the engine for six months and let it make decisions in public. Then we checked those decisions against what the market did.
The scores did not predict returns. That is the honest result, and score_v4 is built around it: the engine no longer leans on a forecast. It limits exposure with rules that hold up without one. Hold less when price is below its own recent average. Never allow more than a borrowing position can bear. Count each warning once. Return zero when a number cannot be trusted.
The one new sizing rule, the trend gate, improved return for the same average risk over eight years of BTC history. It did not make the worst drawdown shallower, and we say so below.
1. Background
Two earlier papers set this one up. The first tested whether the composite score separates good weeks from bad ones and found the effect weakened out of sample. The second tested the sizing output itself and found that the level of exposure did most of the protective work, while the dynamic part did not beat a constant cap at seven days. It pre-registered realized-volatility-first sizing as the score_v4 candidate.
During the freeze the candidates ran in shadow: computed next to every live decision, stored, never served. This paper closes that cycle.
2. What the live record showed
2.1 No detectable return signal
On 2,069 snapshots (17 February – 13 October 2026), with every label computed from exchange candles at both entry and exit, the composite has no detectable relation to forward returns at 24 hours, 72 hours, 7 days or 30 days. Block-permutation p-values exceed 0.56 at every horizon; the family-wise p-value across horizons is 0.98. The pre-specified quintile contrast is +5.6 pp with a 20.5% probability of being zero or negative: not confirmed.
Earlier headline figures from this program (a +21.7 pp contrast in April, a pooled ρ = +0.16 in June) did not survive more data and corrected labels. They are withdrawn.
2.2 Score layers echo trailing price
For each layer we measured the rank correlation with the trailing 7-day return, per data cohort, on the 13 October export.
| Layer | ρ with trailing 7-day return |
|---|---|
| Trend strength | +0.74 |
| Structural score | +0.43 to +0.50 |
| Risk-permission score | +0.31 to +0.50 |
| Composite | +0.43 on browser-recorded rows, −0.03 on five weeks of clean server rows |
A layer that describes the move already made reads strong after a rally and weak after a fall. Sizing that follows it is largest just after the move. On the recorded decisions, size driven by the composite or by the combined score was negatively related to the subsequent 7-day adverse move (ρ −0.38 and −0.43): pro-cyclical.
2.3 Evidence limits
The clean daily cohort contains about 18 independent 7-day windows, all in one calm regime with no panic or euphoria period. Nothing in §2 is a statement about other regimes. It is the reason §3 leans on an eight-year reconstruction.
3. Candidates, and what happened to each
| Candidate | Test | Outcome |
|---|---|---|
| Realized-volatility-first sizing (pre-registered June 2026) | 8-year reconstruction, exposure-matched | Not adopted. Sharpe 0.094 against 0.098 for a constant position at equal exposure; 5% expected shortfall −6.0% against −5.8%. Right-signed on the 2026 record, not significant. |
| De-echoed structural score | Live shadow record | Not adopted. Still tracks trailing price. |
| Size from the combined score | Live shadow record | Not adopted. Pro-cyclical; its shadow field is removed from the response. |
| Trend gate | 8-year reconstruction with the production module | Adopted. §4. |
| Rules counted once per family | Live shadow record, three variants | Adopted. §5. |
| DeFi cap as a hard limit | Property check on the arithmetic | Adopted. §6. |
Three of the six were rejected, including the one we had announced as the favourite. A pre-registration that can only confirm is not a test.
4. The trend gate
Rule. While price is below the mean of its last 50 daily closes, the permitted size is multiplied by 0.5. Missing daily data counts as below.
Method. BTC, 5 March 2018 – 23 May 2026, 3,002 daily records. The size series is produced by the module that serves the API, with the rules and DeFi caps held at 1.0 (they cannot be reconstructed historically). Each benchmark is held to the same mean exposure as the series it is compared with.
| CAGR ÷ max drawdown | |
|---|---|
score_v4 chain with the trend gate | 0.61 |
score_v4 chain without the gate, same exposure | 0.39 |
| Constant position, same mean exposure | 0.37 |
score_v3.2 chain | 0.34 |
Annualised return is 19.5% against 12.3% for the constant position. The gate changes state about 20 times a year.
What the gate does not do. The full-period maximum drawdown is −31.8% against −33.3% for the constant position: about the same. By calendar year:
| Year | Days below average | Return, gate | Return, constant | Max DD, gate | Max DD, constant |
|---|---|---|---|---|---|
| 2018 | 73% | −26.3% | −23.6% | −29.5% | −27.4% |
| 2019 | 46% | +51.6% | +33.5% | −17.7% | −19.7% |
| 2020 | 33% | +89.6% | +72.2% | −17.9% | −20.4% |
| 2021 | 41% | +36.6% | +23.1% | −14.5% | −17.8% |
| 2022 | 78% | −23.3% | −18.0% | −24.2% | −19.7% |
| 2023 | 32% | +42.3% | +44.1% | −6.8% | −7.3% |
| 2024 | 34% | +40.6% | +33.8% | −9.5% | −8.7% |
| 2025 | 53% | −1.1% | −0.1% | −11.2% | −10.8% |
| 2026 (to May) | 48% | −1.8% | −3.3% | −10.7% | −12.4% |
The gate has the smaller drawdown in five of nine years and the larger one in four, including both long bear years. In 2018 and 2022 price sat below the average most of the time, and the brief returns above it failed. The gate earns its place through return in advancing years, not through a shallower worst case.
Against score_v3.2 the drawdown is smaller in all nine years, but that comparison mostly reflects lower mean exposure (0.31 against 0.42) and we do not use it as evidence for the gate.
5. The rules engine counted one condition several times
The audit found three defects in how alerts became a cap.
- Repetition. Crowded positioning produced up to five separate alerts (funding, long/short, open interest, basis, liquidations), each counted.
- Overlap with other caps. Macro, valuation and ETH-downgrade conditions were counted by the rules cap and again by the cap that owns them.
- Severity inversion. Danger and non-DeFi critical alerts counted zero; warnings counted one.
A live ETH case: the rules cap was binding at 0.50 on two conditions counted three times. Counted once, it is 0.86.
In score_v4 each rule family contributes once, by its heaviest alert, with weights that rise with severity. Families owned by another cap are excluded. Rules adverse only to short positions are displayed and not counted.
A relative crowding rule. The long/short rule used fixed thresholds. Measured on 180 four-hour bars, the BTC ratio was above the warning threshold in 77% of bars and the ETH ratio in 91%: the rule was describing the normal state. It now ranks the ratio within its own last 30 days and fires only at a high percentile. The squeeze classifier used a similar fixed cut-off for its "longs crowded" label, which gave the positioning reading a permanent bearish tilt; it now uses the same percentile test.
On 465 recorded decisions (13 October – 16 November) the mean rules cap rises from 0.72 to 0.83 for BTC and from 0.51 to 0.85 for ETH. The rules engine became less restrictive and more specific at once.
6. Correctness of the sizing arithmetic
The cap arithmetic was ported to C and compared with the production JavaScript on 200,000 random inputs, including NaN and infinities. All rows are identical (largest difference 5.8 × 10⁻¹⁶).
Both ports were then given to an external automated-reasoning review with a list of stated properties.
On score_v3.2 it produced counterexamples that were real defects:
- A health factor of exactly 0 was read as "no position" and returned the most permissive level with leverage allowed.
- The DeFi cap did not bound the result: a health factor of 1.3 (cap 0.10) could return size 0.28.
- A NaN in one input could pass through to a non-zero size.
- The capitulation boost could lower the cycle cap in one phase.
All four are fixed in score_v4.
On score_v4 it tested 17 properties. Three were refuted, none a defect in the engine:
- A non-finite health factor on a wallet with debt closes size. This is intended: an indebted wallet cannot have an infinite health factor, so the value is treated as invalid.
- The geometric mean of four equal caps can sit a few units in the last place below that cap (rounding in
logandexp). - A one-unit-in-the-last-place inversion exists for a rules load the engine cannot produce.
No counterexample was found in which an invalid number yields exposure, or in which the DeFi cap fails to bind. The property statements were tightened to say exactly what was shown.
7. Effect on recorded decisions
Replaying the final rule on the engine decisions recorded between 13 October and 16 November 2026 (234 BTC, 231 ETH):
score_v3.2 | score_v4 | |
|---|---|---|
| Mean size, BTC | 0.28 | 0.20 |
| Mean size, ETH | 0.31 | 0.19 |
| 5% worst size × 7-day adverse move, BTC | −3.8% of NAV | −2.1% |
| 5% worst size × 7-day adverse move, ETH | −4.4% | −3.4% to −3.8% |
BTC was below its 50-day average on 167 of 234 decisions and ETH on all 231, so this window shows the gate at its most restrictive. The reduction comes from the gate; the rules cap moved the other way (§5).
8. Limitations
- One calm regime in the live record. About 18 independent weekly windows. The trend-gate case rests on the reconstruction, not on 2026.
- The reconstruction is partial. Rules and DeFi caps are fixed at 1.0, the volatility regime uses three of five inputs, BTC only.
- One asset, one path. Eight years of BTC is a single history with two bear markets. There is no second path on which to confirm the 50-day window or the 0.5 multiplier.
- The replay in §7 reuses stored
score_v3.2values for quality, volatility regime and the non-rules caps. The funding rules and the relative crowding rule cannot be replayed on past rows. - No forecast is claimed. The composite, structural and tactical scores remain in the response as market description. They do not drive size.
9. What is frozen now
score_v4 is frozen for six months from its release: no changes to weights, thresholds, cap formulas or hash inputs. Decisions continue to be recorded with inputs, policy_hash and forward returns. The gate matrix (separate limits for spot, long perps and short perps) stays in shadow until its own test completes.
Citation block
RiskState (2026). score_v4: What Six Months of Live Decisions Changed.
RiskState Research, 16 November 2026.
https://riskstate.ai/docs/research/score-v4-2026-11-16
Live record (2,069 snapshots, Feb 17 – Oct 13 2026, exchange-entry labels):
composite vs forward return: block-permutation p > 0.56 at every horizon,
family-wise p = 0.98; pre-specified contrast +5.6 pp, P(<=0) = 20.5%
echo: rho(layer, trailing 7d) up to +0.74
Reconstruction (BTC, 3,002 d, 2018-03-05 – 2026-05-23, production module):
CAGR / maxDD: trend gate 0.61 · no gate 0.39 · constant 0.37 · v3.2 0.34
maxDD -31.8% vs -33.3% constant; smaller in 5 of 9 calendar years
realized-vol targeting: Sharpe 0.094 vs 0.098 constant (not adopted)
Arithmetic: C port identical to production on 200,000 inputs;
17 properties reviewed externally, 3 refuted, none a defect.
Recorded decisions (Oct 13 – Nov 16 2026, 234 BTC / 231 ETH):
mean size 0.28 -> 0.20 (BTC), 0.31 -> 0.19 (ETH)
BibTeX
@techreport{riskstate2026scorev4,
title = {score\_v4: What Six Months of Live Decisions Changed},
author = {{RiskState}},
institution = {RiskState},
year = {2026},
month = {November},
day = {16},
url = {https://riskstate.ai/docs/research/score-v4-2026-11-16},
note = {Closes the score\_v3.2 freeze. No detectable return signal in
the composite; trend gate adopted on an 8-year reconstruction
with the production module; realized-volatility-first sizing
tested and not adopted; rules counted once per family;
sizing arithmetic checked against a C port and an external
automated-reasoning review.}
}