Independent Research · Built by retail, for retail Scored every trading day
The Proving Ground · shadow harnesses · pre-registered gates

The live models are champions. Everything here is trying to take their seat.

Most shops show you their winners. This page shows you our attempts to beat ourselves — on every desk we run. The autopilot portfolios, the call options ledger, the put lab, the swing desk and the scoring core itself each carry registered challengers running in shadow: identical dated marks, frozen rules, control lanes built to lose, and promotion gates written down before the data existed. No challenger touches capital. The one that proves better becomes the champion — and the old champion's record stays on the books, never rewritten.

We are never satisfied. The champion you can audit on the track record page is only the current holder of the seat — we are always auditing it, always testing against it, always hunting for the version of ourselves that does the job better than the best traders and institutions do it.

We have a long way to go, and we say so in public. But the drive to be the best will never let us stop improving, auditing, and putting our own methods on trial. That is the strength of this desk: not a claim that we are right — a system that finds out.

01How the arena worksSame marks · frozen rules · controls
The champion
Runs live. Defends the seat.

The Pulse and Strike model books, selected monthly and risk-managed by the live pipeline. On every chart below the champion is the zero line — a challenger above it is winning, below it is losing.

The challengers
Paper books. Frozen rules.

Each lane's rule was written down and frozen on the day it was registered. Books are formed once and never edited; a lane that needs to change gets a new name and starts over. Every lane is marked on the same dated prices as the champion, costs included.

The controls
Built to lose. Raced anyway.

A random portfolio and an anti-momentum portfolio run beside the real candidates. If a control starts beating the champion, the test is broken, not the market solved — and the page will say so.

The line that matters
In-sample vs. out-of-sample.

Shaded history is calibration — computed from data that existed when the lanes were chosen, worth nothing as proof. The competition only counts what happens after registration. That line is drawn by the harness, not by us.

02Autopilot PortfoliosFoundation · Pulse · Strike
Read this before the numbers

Nothing in this section is evidence yet.

These lanes were frozen on 14 August 2026. The books they are scored against were formed before that date — so every figure below is calibration: a rule being marked against the very data it was built from. It is printed because the ledger should be complete, and it is worth nothing as proof. A rule graded on its own training data is expected to look good.

The tell is in the numbers themselves: the best-looking lane is the one defined most recently. That is the signature of contamination, not skill, and it is exactly why we do not let calibration count.

The first book that counts is the 1 September 2026 selection — month one of twelve. Earliest possible verdict: September 2027. Until a lane has out-of-sample months on the board, the competition column below reads “—”, and that dash is the honest answer.

Champion · the allocation map

Tactical Radar Foundation

Decides how invested the whole desk is and holds the steady ETF core. No standalone shadow book challenges Foundation today — but its biggest lever, the live invested-%, is inherited by the two-mandate lanes below, so every autopilot test is also a test of Foundation's signal.

Champion · the steady core

Tactical Radar Pulse

The 15-name diversified model book. Defending against: the three selection challengers (CH1_m12_top15, CH2_m3_top15, CH3_fundrev_top15), the construction lane P1_pulse_theme3_emp, and both weekly-resize lanes — every one graded against Pulse on identical dated marks. Lanes are named here exactly as they are named in the logs those numbers come from.

Champion · the aggressive sleeve

Tactical Radar Strike

The concentrated 5-name book we advertise as aggressive. Defending against: the construction lane S1_strike5_emp and, pre-registered on 2026-08-16 and not yet writing, S2_strike5_disp — the same book as S1_strike5_emp, differing only in how the weights are set, so the pair is a clean read on one idea rather than two. Its registered gate is deliberately harder than the others on this page: it must beat S1_strike5_emp, not merely the incumbent, because a sizing rule that only matches equal weight has bought nothing. It also defends against both weekly-resize lanes. Its record, wins and misses, lives on the track record.

Programme A · registered 2026-08-14

The Weekly-Resize Challengers

"What if, every Monday, the book shifted weight toward the month's winners and re-struck its stops?" Our own backtests say the resize is roughly a coin flip and the trailed stops cost money — so both ideas now have to prove themselves in public, against the live books, before either touches a dollar.

gate progress: 0 / 24 out-of-sample resize Mondays · verdict no earlier than Feb 2027
Weekly-resize lanes vs. the live Pulse book
calibration · in-sample+0.00-0.34+0.38CHAMPION (0)winner-tilt -0.17trailed +0.36anti +0.1705-2206-0306-2907-2208-1309-03
Cumulative net difference vs. the champion, percentage points, chained month by month on identical marks. The champion is the zero line.
Weekly-resize lanes vs. the live Strike book
calibration · in-sample+0.00-0.83+0.32CHAMPION (0)winner-tilt -0.67trailed -0.75anti +0.2805-2206-0306-2907-2208-1309-03
Cumulative net difference vs. the champion, percentage points, chained month by month on identical marks. The champion is the zero line.
LanePulse · calibrationPulse · competition Strike · calibrationStrike · competition
W1_resize30_mtd winner-tilt resize COMPETING-0.17 pp+0.00 pp-0.67 pp+0.00 pp
W2_resize30_trail resize + trailed stops COMPETING+0.36 pp+0.00 pp-0.75 pp+0.00 pp
C1_anti30 anti-momentum control CONTROL — NEVER PROMOTED+0.17 pp+0.00 pp+0.28 pp+0.00 pp
Programme B · registered 2026-08-14

The Selection Challengers

"Forget our methodology entirely — what would the auditable price history build from scratch?" Answer: a 15-stock, equal-weight book picked purely by 12-month momentum. In a disciplined walk-forward test it beat all 200 random portfolios we raced it against. Now it has to beat the champion where it counts: out of sample, in public. A third lane joined on 2026-08-15 — our first selection rule that reads no price at all, ranking instead on analyst revisions and profitability, including a free-cash-flow test built to catch the stock that only looks cheap. It carries no calibration history whatsoever, because the historical ratings data we could have tuned it on turned out not to be point-in-time. Starting a lane with nothing to show is the more honest way to start one.

gate progress: 0 / 12 out-of-sample months · verdict no earlier than Sep 2027
Selection challengers vs. the live Pulse book
calibration · in-sample+0.00+14.08CHAMPION (0)12-month +10.873-month +5.75fundrev +14.08random +3.8005-2205-2508-0308-1308-24
Each lane is a full 15-stock portfolio picked by one frozen rule on the month's selection day, marked on the same dated prices as the champion. Champion is the zero line.
LaneCalibration (in-sample) Competition (out-of-sample)
CH1_m12_top15 12-month momentum CALIBRATION+10.87 pp
CH2_m3_top15 3-month momentum CALIBRATION+5.75 pp
CH3_fundrev_top15 fundamentals + revisions CALIBRATION+14.08 pp
C1_random random control CONTROL — NEVER PROMOTED+3.80 pp
Programme C · registered 2026-08-15

The Two-Mandate Test

Our two portfolios are meant to be different products, not two flavours of the same one — Pulse the steady core, Strike the aggressive sleeve we actually advertise as aggressive. These lanes test that idea honestly: same stock-picking rule, different shape. One holds five names; one holds fifteen but refuses to spread itself one-name-thin across every theme. Both hand the same job to the same safety switch — our macro model decides how much of the book is invested at all, and moves the rest to cash. On our own history that switch is worth far more to a concentrated book than to a diversified one, which is exactly why a high-risk sleeve can be run responsibly. Until these lanes clear their gate, that remains a hypothesis with numbers attached.

Two-mandate construction lanes vs. the live Pulse book
calibration · in-sample+0.00+15.20CHAMPION (0)strike-5 +12.59pulse-15 +6.49random +3.8005-2205-2508-0308-1308-24
Identical stock-picking rule in every lane — only the shape of the book changes, and both candidates scale their exposure to the same EMP cash signal the live books follow. Champion is the zero line.
LaneCalibration (in-sample) Competition (out-of-sample)
S1_strike5_emp concentrated 5, EMP-scaled CALIBRATION+12.59 pp
P1_pulse_theme3_emp 15 names, max 3 per theme CALIBRATION+6.49 pp
C1_random random control CONTROL — NEVER PROMOTED+3.80 pp
03Call OptionsThe desk lab · registered filters
Desk lab · candidates registered 2026-07-21

Can a volatility filter beat taking every call?

The champion is the live, unfiltered Call Options ledger — every Focus setup taken, both books, marked in public. Three pre-registered challengers each drop the day's call when the name's realized volatility crosses a line (80%, 85%, 90%). The harness scores them on the exact P&L numbers the live ledger prints, so the lab and the page members see can never disagree.

Adoption rule, written before the evidence: a filter is a proven improvement only when — out of sample — the ledger has enough matured trades, the filter has actually excluded some (a filter that never acts has proven nothing), and the trades it excluded are net-negative on both books: dropping them must have helped the call ledger and the spread ledger. Until then the lab says "watching."
First out-of-sample status file publishes with the next lab run. A performance chart appears with it — this desk never draws a curve before it has matured out-of-sample trades to draw.
04Put OptionsFour candidates · three controls
Put lab · lanes frozen 2026-08-12

Is there a durable way to be long downside?

We searched 22,116 put-strategy combinations and learned something humbling first: pure chance produced a better "best result" than the real data did. So instead of shipping the winner of a data-mine, the four most defensible hypotheses were frozen as lanes — beside three control lanes chosen to have no view at all — and every one now has to earn its number on trades it had never seen. Nothing here is a product; it is the audition to become one.

Stale — not today’s numbers. The put harness last wrote its status on August 28, 2026, 7 days ago. The figures below are real and are shown exactly as that run left them, but they are not current. We would rather say so than let an old number wear today’s date.
gate progress: 18 / 60 matured out-of-sample trades on the leading candidate · and ≥12 independent exit days
LaneMatured (OOS)Independent exit daysWin rateNet P&L ($, on $100/day)Cumulative P&L (drawn by the harness)Open
L1 wk52high10 7d 5otm COMPETING
7-day · 5% OTM · 3 picks
18411%-1,295.77trade 1: cumulative +187trade 2: cumulative +87trade 3: cumulative -13trade 4: cumulative -113trade 5: cumulative -213trade 6: cumulative -96trade 7: cumulative -196trade 8: cumulative -296trade 9: cumulative -396trade 10: cumulative -496trade 11: cumulative -596trade 12: cumulative -696trade 13: cumulative -796trade 14: cumulative -896trade 15: cumulative -996trade 16: cumulative -1096trade 17: cumulative -1196trade 18: cumulative -129615
L2 lowvol5 30d atm AWAITING FIRST MATURED TRADE
30-day · ATM · 3 picks
L3 highvol 30d spread10 AWAITING FIRST MATURED TRADE
30-day · ATM · spread 10% · 1 pick
L4 highvol3 30d spread10 AWAITING FIRST MATURED TRADE
30-day · ATM · spread 10% · 3 picks
C1 random CONTROL — NEVER PROMOTED
30-day · ATM · 1 pick
C2 highvol atm CONTROL — NEVER PROMOTED
30-day · ATM · 1 pick
C3 highvol3 atm CONTROL — NEVER PROMOTED
30-day · ATM · 3 picks

Out-of-sample rows only. The 494 in-sample rows that chose these lanes are logged and drawn on the research monitor — and counted for nothing. Status generated 2026-08-28 06:06 by the put harness.

05Swing TradingLive rule vs. four variants
Swing lanes · pre-registered 2026-08-15

The swing rule, challenged on every front.

The swing desk's live rule was written down in full before it ran — and a register of named variants now replays every archived scan beside it, each lane isolating exactly one question: is the entry bar too high? Do the quality gates help timing? Is money flow the dominant pillar? Do earnings drift, insider buying, the 200-day line, or relative strength add real edge? Every lane was registered before its evidence existed — adding one after peeking at results is the sin this register exists to prevent.

gate progress: 0 / 60 matured out-of-sample picks on the leading lane · 15-session maturity per pick
LanePicks so farMaturedIndependent exit daysGate checks passed
C0 live CONTROL — THE LIVE RULE
The control — the live rule itself · What did the live rule do?
3000 / 6
L1 v11 ready70 AWAITING FIRST MATURED PICK
The retired v1.1 rule, kept running · Was retiring it right, on data that did not exist then?
6000 / 6
L2 no foundation AWAITING FIRST MATURED PICK
Drop the CRC quality gates · Does Train 2’s opinion help timing?
3000 / 6
L3 fuel forward AWAITING FIRST MATURED PICK
Re-weight pillars toward money flow · Is institutional flow the dominant signal?
3000 / 6
L4 firing65 AWAITING FIRST MATURED PICK
Entry bar 65 · The sweep runner-up — tracked instead of trusted
3000 / 6
L5 rotation AWAITING FIRST MATURED PICK
Relaxed bar inside leading themes · Does a group tailwind justify lighter single-stock evidence?
0000 / 6
L6 theme veto AWAITING FIRST MATURED PICK
Live rule minus weak-theme names · Does removing dead-group trades help?
2000 / 6
L7 pead AWAITING FIRST MATURED PICK
Post-earnings drift overlay · Does fresh earnings momentum add edge?
2000 / 6
L8 insider AWAITING FIRST MATURED PICK
Longs require insider cluster buying · Does follow-the-insiders improve entries?
0000 / 6
L9 defense AWAITING FIRST MATURED PICK
Longs require a 200-day defense · Does the long-term trend line protect entries?
3000 / 6
L10 rsline AWAITING FIRST MATURED PICK
Longs require a rising relative-strength line · Is leadership vs. the market the tell?
1000 / 6
L11 confluence AWAITING FIRST MATURED PICK
200-day defense AND rising RS line · Do the two guards compound?
1000 / 6
L12 focus AWAITING FIRST MATURED PICK
The daily Focus pick, inside this harness · How does the flagship signal grade under swing rules?
0000 / 6
L13 shorts on AWAITING FIRST MATURED PICK
Registered lane — see the lane register
6000 / 6
L14 volume shock AWAITING FIRST MATURED PICK
Registered lane — see the lane register
1000 / 6
L15 bestofday AWAITING FIRST MATURED PICK
Registered lane — see the lane register
1000 / 6

Lane picks are replayed deterministically from the archived daily scans — the harness re-fetches nothing and cannot touch the live record. Status generated 2026-09-01T21:58:15 by the swing harness (vv1.1).

Why there is no chart here yet. Performance curves on this page are drawn only from matured out-of-sample results, and every swing lane is still at zero matured picks — the honest chart would be an empty axis. The curves appear on their own the day the first picks mature. We never plot data that does not exist.

06The Scoring CoreEven the referee gets audited
Score lab · incumbent registered 2026-05-22

Five challengers to the CRC score itself.

Every number on this site flows from one place: the CRC score. So it gets the same treatment as everything downstream of it — an incumbent defending against registered challenger composites, graded daily on how well today's scores predicted the next month's returns.

Stale — not today’s numbers. The score lab harness last wrote its status on August 07, 2026, 28 days ago. The figures below are real and are shown exactly as that run left them, but they are not current. We would rather say so than let an old number wear today’s date.
CompositeRegisteredScoring runs graded
INCUMBENT CRC (M+V+Q) THE CHAMPION2026-05-2238
C1 Defense sleeve (R2) CHALLENGER2026-07-1838
C2 Horizon shift (R3) CHALLENGER2026-07-1838
C3 Defense + horizon CHALLENGER2026-07-1838
C4 Rank rebuild sketch (R6) CHALLENGER2026-07-1838
CRC 3.0 three-pillar (T/E/S) CHALLENGER2026-07-1710
What happened in August is the whole philosophy in one story. On 2026-08-11 the lab reported all four candidates as clearing the promotion bar at once. Four ideas all winning together is not four good ideas — it is a signal to audit the test. The audit found two holes in the harness itself, either of which would have promoted something the evidence did not support. The holes were fixed, the fix was written down, and the decision was reset to 2026-08-27 — registered before the deciding evidence existed. When the referee looked wrong, we investigated the referee. That date has now passed; what the lab decided on it is on the liveness strip below.
07The promotion gatesWritten before the data existed
A challenger is promoted only when, on out-of-sample data alone, it has: enough independent observations (24 resize Mondays or 12 matured months for the portfolios; 60 matured trades and 12 distinct exit days for the put and swing lanes — trades that settle on the same day share one market move and are not independent); a lead over the champion that is positive in both halves of its record; a lead that is statistically real (clustered t > 2, not a lucky streak); and its matched control failing the same test — because an edge a random book can match is not an edge. Even then, promotion is a written, dated decision with a name on it — never automatic, and never quiet. Most months the honest scoreboard reads "nothing is ready yet." That is the expected result, and we publish it anyway. First formal gate dates already on the calendar: Feb 15, 2027 (resize), Sep 1, 2027 (selection). Already reached: Aug 27, 2026 (scoring core) — what the lab decided is reported on the liveness strip below, read from the research runner rather than restated here.
07BIs the machine actually running?Harness liveness · read from the runner

A promotion gate is only as honest as the harness feeding it. A programme that quietly stops writing would leave its last numbers on this page looking current forever — so the runner reports, every day, when each programme last wrote, and this page prints that report rather than its own opinion of it.

Harness liveness unavailable. The roll-up that reports whether each research programme is still writing is not being read this run — roll-up itself is 6 days old (written August 28, 2026). Each desk above still carries its own freshness check, which is the weaker signal but cannot go stale without its harness going stale too.

08Why the boring parts matterControls · gates · the losing lanes

Elsewhere you can watch AI models run stock portfolios in public, and some of the numbers are spectacular. We read them closely — it is part of the job. Three things are usually true of them, and all three are worth naming, because avoiding them is most of what this page is.

The leader keeps changing. Run enough unconstrained portfolios side by side and one of them will look brilliant this quarter — that is arithmetic, not skill. It is the same trap our own options research walked into and caught: when we searched twenty-two thousand strategy combinations, pure chance produced a better "best result" than our real data did. That is why every programme here ships with a control lane built to lose. If the control wins, we say so and throw the finding out.

One position is usually the whole story. A book with a third of its capital in a single winner is not a strategy that worked; it is one holding with commentary attached. Our books cap any name at 8% of capital, across both portfolios combined, and the cap is enforced by the risk layer rather than by good intentions.

Confident writing is not evidence. We fact-check the reasoning we read — including our own — against the filings. It is routine to find a specific, fluent, load-bearing number that simply does not reconcile with the 10-Q. Every figure on this page is computed by a harness from dated data, not written by a model that sounds sure.

None of that means we are ahead. It means we would rather be slow and checkable than fast and lucky — and that when a challenger here finally beats the champion, you will be able to tell the difference.

What this page is, and is not. Challenger lanes are hypothetical shadow portfolios — research, not tracked performance of any account, and not investment advice. They are marked on the same dated model prices as the live books, net of assumed costs, exactly as our harnesses compute them; this page publishes the harness numbers and computes nothing of its own. Shaded "calibration" history is in-sample and is never counted toward promotion. The champion's own record — including every miss — lives on the track record page and is never rewritten. Methodology in full on The System. Page rebuilt September 04, 2026 by the same pipeline that runs the models.