Anatomy of a 14-Day Silence

For fourteen consecutive days — May 19 through June 1 — the BV-7X oracle published NEUTRAL while every one of its sub-signals screamed bearish. Bitcoin fell 9.6% over that window. The model wasn't wrong. It wasn't even uncertain. It had a high-conviction, unanimous directional view the entire time — and a single line of risk-management logic refused to let it speak.

This is the post-mortem: what broke, what it cost, the fix, and — because I get asked this — exactly what the fix did and did not do to the numbers. No spin. The headline is not "we got more accurate." The headline is "we stopped going dark."

*Resolved 7-day windows only (6 of the 9 recovered SELLs have completed their horizon as of June 2). See the caveats section — these are not yet in the scored backtest.

1. The symptom: a flat line where a signal should be

The oracle computes one directional call per day at 21:35 UTC, pins it to IPFS, and attests it on Base. Every call carries four sub-signals — trend, momentum, flow, value — that vote, and a decision layer that turns the vote into BUY / SELL / HOLD. On May 19 the published calls went flat and stayed flat for two weeks.

Here is the tape. The line is BTC spot; each red dot is a day the model published NEUTRAL. The directional view underneath every one of those dots was SELL.

A 9.6% decline, called by no one — because the model that called it was muted. BTC: $76,959 → $69,585. Intra-window high $77,693.

2. The diagnosis: a model that worked, gated by logic that didn't

The first thing to rule out was a dead data feed or a crashed signal layer. It was neither. Every single day, the decision layer produced the same trace:

The model reached SELL. Then the SELL proximity gate — a risk filter that only lets sell signals fire when price is "close enough" to a reference moving average with a sane RSI — overrode it to HOLD. Every day. The gate exists for a good reason: historically, selling deep into an already-collapsed, far-from-mean market is a coin flip. But it had two coverage holes that, together, swallowed an entire regime.

Hole 1: the dead band

The gate had two branches. A structural-bear branch (for price more than 15% below the 200-day MA) and a normal branch (for price within 10% of it). Price that sat 10–15% below the MA fell through both — too far for normal, not far enough for structural — and defaulted to HOLD. A textbook off-by-one in threshold space.

Hole 2: the RSI floor

The normal branch also required RSI between 40 and 55. But a market in a confirmed downtrend is, by definition, oversold — RSI sat at a median of 32.6 across the streak. So even after closing the dead band, the normal branch still muted every bearish call whose RSI lived below 40. This was the bigger hole, and it's the one the first patch attempt missed.

The two holes mapped cleanly onto the data. Below: every day of the streak, its distance below the 200-day MA, its RSI, and which hole trapped it.

Bar width = distance below the 200-day MA (Jun 1 = −10.7%, the deepest). †Days with RSI below 30 are deliberately still held — that's the capitulation zone, where you don't want to initiate new shorts into a possible bottom.

3. What the silence cost

A signal product that goes dark for two weeks has a cost even if you can't book it as P&L: it's the track record you didn't build, the attestations you didn't write, the calls subscribers didn't get. And in this case the calls would have been right.

Of the nine days the fixed gate now fires SELL, six have completed their 7-day evaluation window as of June 2. All six were correct — price was lower seven days later, every time.

Date	Dist MA200	RSI	7d ROC	Old	New	Realized 7d
May 19	−5.5%	40	−4.7%	HOLD	SELL	−1.36%
May 20	−4.6%	34	−2.5%	HOLD	SELL	−3.53%
May 21	−4.2%	35	−4.5%	HOLD	SELL	−5.20%
May 22	−6.2%	40	−4.0%	HOLD	SELL	−3.28%
May 23	−5.2%	33	−2.1%	HOLD	SELL	−3.57%
May 24	−5.2%	36	−2.3%	HOLD	SELL	−3.80%
May 27–Jun 1	≤−6.8%	30–32	<−3%	HOLD	SELL	pending
Resolved	6 of 6 correct · avg realized					−3.46%

4. The fix (v5.6.9)

Two changes, both in the SELL proximity gate. Both are exposed as tunable thresholds and mirrored exactly into the backtest engine, so the live model and the simulator can never silently disagree.

The second change is the load-bearing one. It says: when a death cross confirms a structural downtrend and price is actively declining, an oversold reading (RSI 30–40) is a reason to sell, not a reason to abstain. Below 30, the model still steps aside — that's the capitulation zone, where shorting the bottom is how you give back the whole trade.

5. Validation: the part where I refuse to oversell it

Here is the question a CEO should actually ask: "You loosened a risk filter. Did you just trade discipline for noise?" The answer, from a full-history backtest over 5,600 days (2011–2026), is no. The change is accuracy-neutral.

Metric	Before (floor 40)	After (floor 30)	Δ
Overall accuracy	63.04%	63.01%	−0.03pp
BUY accuracy	64.66%	64.66%	0.00pp
SELL accuracy	57.49%	57.36%	−0.13pp
SELL near MA200	—	62.4%	≥58% floor
SELL far below MA200	—	59.4%	≥58% floor
Actionable signals	2,013	2,014	+1

Read those last two rows carefully, because they're the whole story. Over thirteen years, loosening the floor changed the historical signal count by exactly one. The regime this fix targets — death cross, price hugging the MA, RSI stuck in the 30s — occurred on just 7 of 5,600 days (0.13%), and five of those seven are the streak we just lived through. The fix is, by construction, almost invisible to history and surgical to the present.

6. The governance footnote

Two smaller things shipped with v5.6.9, because a model is only as trustworthy as its paper trail:

• Version drift, closed. The previous fix attempt changed live logic but never bumped the version constant — so predictions were being stamped 5.6.8 while running newer code. That's exactly the kind of silent drift that erodes an attestation's meaning. The stamp is now 5.6.9, enforced by a test that ties the version constant to the changelog's LIVE row.
• 13 new regression tests. The gate is now pinned at every boundary — the dead band, the RSI floor, the death-cross requirement, the capitulation guard. The bug that produced a 14-day blackout cannot silently return.

Every call BV-7X has ever made is on-chain, timestamped, and verifiable — including the fourteen that said NEUTRAL. We don't get to quietly delete an embarrassing fortnight; it's attested on Base forever. The only honest response to a public track record is to fix the machine in public and show the math. That's this post.

A model that abstains when it's certain isn't being cautious. It's being broken quietly. The fix wasn't to make it smarter — it was to stop muzzling it when it already knew the answer.