The Year SkyAnalyst Learned to Trust Claude: 99 Trades, +16.57R YTD

Through May 7, 2026, SkyAnalyst has banked plus 16.57R YTD across ninety-nine trades from system inception on Jan 12. Fifty-eight winners, forty-one losers, a 58.6 percent win rate. The simulated $100,000 account at 2 percent risk per trade now sits at $133,139 on the static line and $136,266 on the compounded line — a 33-percent year-to-date return on the published baseline, with the spread between static and compounded a $3,127 read on how steadily winners and losers alternated. The headline number is real. The story underneath it is the one worth reading. This year did not open with ninety-nine trades. It opened with three. The system spent January in soft launch — the journal published from Jan 12, the trades were live, the broker fills cleared, but the desk took a single trade in three weeks because not every model was running and not every setup was clearing threshold. February brought half the models live and twenty-one trades to the published record. March stood the rest of the desk up and produced forty-two trades — the first month with the full agent stack, every model attributing, every setup library active. Toward the end of March, the desk consolidated around Claude Opus 4.6 as the master flow's primary model. April was the first calendar month run that way. May has continued the cadence into a fifth. Plus 16.57R YTD reads as a result. The work underneath it was a four-month decision about which model carries weight, which setups earn capital, and which instruments produce the system's edge. This article is the unedited record.

Act 1: Jan 12-31, the soft launch and the first three trades

The system went live on Jan 12, 2026. Three trades cleared the published record across the rest of the month. Every one was a winner. The Trend Agent, the Macro Agent, the Cross-Asset Agent, and the Risk Agent were all running, but the cross-model flow was not yet provisioned and the threshold conservatism that ships in any new automation was tighter than where it would settle. Three winners on three entries reads cleanly on the line; the operating reality was that the system was rejecting more entries than it took.

Net R closed January at plus 3.02R from a 100 percent win rate. The simulated $100,000 account at 2 percent risk ended Jan 31 at $106,045. We do not draw any expectancy reading from a 3-trade sample. The published January 2026 monthly recap walks the three case-study writeups in order. The honest framing of January is that this was the month the system proved it could clear a complete trade lifecycle — agent evaluation through broker fill — not the month it proved its edge.

Act 2: Feb 1-28, half the models live and the first real sample

February brought additional models online and twenty-one entries to the published record. Net R closed plus 4.41R at a 61.9 percent win rate. The first month with a sample large enough to read.

The shape of February was that variance compressed. Thirteen winners against eight losers. No window inside the month produced a +4R bumper. No window produced a -3R drawdown either. The system held expectancy through a 21-trade sample without leaning on any single high-grade winner. The simulated account closed February at $114,505 — up roughly $8,460 from January's close, with the static-vs-compounded spread already opening because winners and losers were alternating cleanly rather than clustering.

The architecture proved out in February. The agent panel was running the same logic that would carry the rest of the year. Confluence math, regime detection, macro gates, cross-asset correlation. The setups that paid in March and April were already paying in February. See the February 2026 monthly recap for the per-window breakdown across Feb 2-7, Feb 9-15, Feb 16-22, and Feb 23-28.

Act 3: Mar 1-31, full force and the Claude consolidation

March was the first calendar month with every model live in production and every setup library armed. Forty-two trades cleared the published record — twice February's count. Net R closed minus 0.13R at a 52.4 percent win rate. The largest sample of any 2026 month so far, and the first month the system survived variance both directions inside one window.

The mid-March drawdown was the year's deepest stretch. The published record carried a minus 9R run-down across an Mar 16-22 window before recovering on the closing bumper of plus 4.19R in the Mar 23-29 window. Forty-two trades, expected variance for a positive-expectancy playbook, and a net result close enough to flat that no tuning signal cleared threshold. The system did not change posture across the drawdown. It did not change posture across the recovery. The published March 2026 monthly recap walks the regime-by-regime read.

The other story of March was the model decision. Through the first three weeks the desk ran a parallel cross-model evaluation: every published trade carried both Claude and GPT analysis, and the master automations rotated between the two on the criteria each agent panel cleared. As the sample accumulated, the dispersion settled. By the closing week of March, Claude Opus 4.6 had produced the more consistent discipline across regime shifts — particularly in macro-event windows where the agent panel's gating depended on getting the macro read right the first time. Toward the end of March, the desk consolidated the master flow around Claude. GPT continues to run on the parallel benchmark cadence and the case-study pipeline; the master automations now route through one model.

That decision is the consequential operational choice of 2026 so far.

Act 4: Apr 1-May 7, the rhythm

April was the first complete calendar month run on the post-consolidation flow. Twenty-four trades, plus 0.92R net, 50 percent win rate. The median month — variance inside the central tendency, posture held, no window swinging far in either direction. The simulated account closed April at $115,685. The published April 2026 monthly recap argues that the absence of a story is the story.

The first week of May has carried the same cadence into a fifth month: nine entries, the win-of-window the May 5 EURUSD short on VWAP rejection at plus 2.22R credited, and the simulated account climbing toward $133,139 as of May 7. We do not call a four-month sample a track record yet. We call it a system holding expectancy across a meaningfully large window.

Pattern of the week

The pattern that defined 2026 so far is threshold conservatism. The system rejects more entries than it takes. Every winning trade — every loser too — passed an agent panel that started with the Macro Agent's regime read, walked through the Trend Agent's confluence stack, cleared the Cross-Asset Agent's correlation tolerance, and survived the Risk Agent's veto pass. Setups that look identical to the human eye disagree at the bar level on whether the macro gate is open, whether the confluence threshold has cleared, or whether a correlated cross-asset position changes the risk math.

Why threshold conservatism produced 58.6 percent across ninety-nine trades

The win-rate floor in any positive-expectancy playbook is the fraction of taken entries that work. The threshold sets that floor by deciding which entries get taken in the first place. Loosen the threshold and the system takes more trades, the win rate drops, and the per-trade expectancy compresses. Tighten it and the inverse happens. The 58.6 percent the system delivered across ninety-nine entries is the rate at which a moderately tight threshold cleared during a year that did not produce an unbroken trend regime.

Every monthly recap of 2026 has carried evidence of the threshold doing its work. February's plus 4.41R came on twenty-one entries — a single-digit fraction of the candidate setups the agent panel evaluated. March's forty-two entries against the cross-asset complex's broader opportunity set means the system rejected the majority of bar-level continuation prints because the macro gate was not aligned. April's twenty-four entries against the same opportunity set says the agent panel's pass rate was inside its working range. The threshold is the architecture's most consequential parameter, and the 2026 record is the sample that says where it should sit.

We tune the threshold against the rolling 100-trade window, not against the calendar month. Through May 7, the rolling window has just turned over for the first time. Whatever signal that distribution produces will be the first credible tuning input the architecture has had since launch.

Decision highlights

The end-of-March consolidation around Claude Opus 4.6 is the year's defining operational decision. Through the first three weeks of March the desk ran cross-model: every published case study carried both a Claude and a GPT analysis, and the master automations rotated between the two on the criteria each agent panel cleared. As the sample accumulated, the macro-gate dispersion settled — Claude's pre-trade reads held discipline through regime shifts more consistently than the parallel GPT runs. The desk consolidated the master flow around Claude in the closing week of March; April was the first complete calendar month operated that way. The benchmark cadence continues running in parallel.

The Mar 16-22 drawdown was the year's deepest test of the architecture's posture-hold rule. The published record carried a minus 9R run-down across the window before the closing bumper of plus 4.19R in the Mar 23-29 week. Through the drawdown the agent panel did not change threshold, did not flip the macro gate's bias, and did not consult the recent record to size the next decision. Forty-two March entries, expected variance for a positive-expectancy playbook, a net result inside the rolling-100 distribution, and zero tuning changes. The discipline of holding posture through a 6-day stretch where every entry was producing minus 1R outcomes is the part of the system that's hardest to validate with a small sample.

The decision to publish every trade in window — winners, losers, drawdowns, and all — is the year's editorial decision the system itself depends on. Every monthly recap has carried the full trade index. Every losing window has been published with the loss-teardown intact. The Apr 23-24 cluster of three correlated stops, the Mar 16-22 drawdown, the XAUUSD slow bleed across two months: each one sits on the public record as plainly as the bumpers. A subscriber evaluating the system reads the same data the desk does. That commitment is what lets the +16.57R YTD figure carry weight; without it, the number would be marketing.

Loss worth learning from

The cleanest loss the year produced was not a single trade but the Mar 16-22 drawdown stretch. Across the window the published record carried a minus 9R run-down — six trading days of negative-leaning outcomes that pulled the cumulative line back toward break-even after February's clean accumulation.

What the system saw that was right

A regime where the agent panel's confluence threshold kept clearing on continuation setups that had paid through February and into the first half of March. The Trend Agent's bar-level reads were consistent. The Macro Agent's gate was open on most of the relevant window. The Cross-Asset Agent flagged correlations as inside tolerance. The setup library was the same one that had produced February's plus 4.41R and would produce the closing bumper of plus 4.19R one week later.

What the system got wrong

The Mar 16-22 tape inverted intraday in a way the bar-level evaluation rhythm could not detect ahead of time. Continuations resolved as fades inside the same session. The Cross-Asset Agent did not veto the correlated cluster of stops because the bar-level structural premises were independent at trigger time. Six trading days of regime mismatch followed. None of the individual trades were misjudged at entry; the regime they entered into changed shape at intraday timeframes the agent panel does not gate on.

What we would do the same

The entry rule is the entry rule. The system does not consult the recent trade record to size or decline a new entry. Every entry inside the Mar 16-22 window cleared the same threshold that produced winners on either side of it. Streak-aware filters would lower realized expectancy across calendar windows by rejecting entries that, on long-run sample, average to positive expectancy. The plus 4.19R closing bumper of Mar 23-29 paid for the drawdown by structure rather than design. Plus 16.57R YTD and 58.6 percent across ninety-nine entries says the threshold is sitting at a working level the rolling-100 sample is now starting to confirm.

From the desk

The honest reading of 2026 so far is that an architecture launched in soft mode in January is now operating with four months of unbroken cadence and a result that, on the published baseline, would clear the median public-record AI-trading benchmark we are aware of. Plus 16.57R net across ninety-nine entries at 58.6 percent does not say the system has solved any of the durable problems in algorithmic trading. It says the system has produced a positive-expectancy result on a sample large enough that the rolling-100 distribution can begin to read.

The simulated $100,000 account at 2 percent risk per trade now sits at $133,139 on the static line and $136,266 on the compounded line. The 2.3 percent spread between the two lines is the read on how steadily winners clustered after winning streaks across the four-month window — compounded outpaces static slightly, which is the signature of winners that came after winning runs more often than after losing runs. That spread is the disciplined-sizing math working as designed across a sample where variance landed close to its central tendency. A subscriber sizing 2 percent of an actual $100,000 account through the same trades would have realized substantively the same return.

What carries forward is the same playbook. The agent panel runs the same threshold. The Macro Agent gates the same regime variables. The Cross-Asset Agent enforces the same correlation tolerance. The Risk Agent retains its veto. The master flow runs through Claude Opus 4.6 with the parallel benchmark cadence continuing on GPT-5.4. We will publish the next monthly recap when May closes, the next weekly recaps as each window completes, and the next case study when a trade clears the threshold for editorial coverage. Every trade will land on the published record.

If you read one thing from this article, read this: the system is not optimized. It is calibrated. The threshold sits where it sits because four months of live record have not yet produced a signal saying it should sit elsewhere. The 16.57R YTD is what calibration produces when the architecture is built right and the threshold is honest. We expect the rest of 2026 to test that thesis, and we expect to publish whatever the test produces — winning months, losing months, drawdowns and all.

Monthly recaps for the year so far: January · February · March · April.

What we're tuning

2026 has not produced a tuning signal that cleared threshold yet. The architecture has shipped without a single retune of the agent-panel threshold across one hundred trades. Plus 16.57R net at 58.6 percent across the published record sits inside the rolling-100 distribution the playbook implies. The per-instrument dispersion (US500 plus 5.05R against XAUUSD minus 0.30R, NAS100 plus 1.00R) is the variance the threshold-conservative architecture produces across a multi-instrument book. Tuning on a sub-100 sample would over-fit the noise.

What the desk tracks forward into the rest of the year: whether the rolling-100 turn produces a signal worth acting on (the first credible tuning input since launch); whether XAUUSD's drag persists or reverses through the next regime rotation; whether the equity-index complex keeps producing the bulk of the system's net R or whether the FX side closes the gap; and whether the next cross-model window produces a different macro-gate dispersion than March's. The benchmark cadence continues running in parallel. None of these are immediate tuning items. They are the variables that, if they shift, would surface as a signal the rolling sample has now grown enough to read.