Does “4 in 5” actually mean 4 in 5?

What “4 in 5” looks like on a real stock

Each evening Shunt draws the range for the next day, using only what had happened by then. The shaded lane is that range. Each dot is the move that followed. If “4 in 5” is honest, about one dot in five lands outside the lane, and that is what you see.

The same test on every stock

For every stock, the range was built only from events before each day, then checked against that day. Price history reaches back ten years; each row below says which window it uses. Each bar is the share of moves that stayed inside the range. The dark tick marks the 80% target, so a bar that reaches it means the range was honest.

Honest is not the same as useful. A range that is very wide catches nearly every move and tells you little. So each row also shows how wide the range is for a typical stock today, how far the result sits from 80% in percentage points, and the 95% interval around it. Close to 80% and narrow is what useful looks like.

Weekends fall short: 71% overall, 78% in the most recent half. A correction learned on older weekends overshot on newer ones (90%), so it is not applied; Shunt shows the weekend number as it is. Ranges are scaled by each stock's volatility now, refreshed after every US close. Data as of 2026-09-28.

Live, written before the outcome

After every US close, Shunt locks in tomorrow's 4 in 5 range for 30 stocks (the earnings range if one reports). After that session closes, each one is graded. Entries are never edited once written. This only counts as evidence once enough sessions have been graded; until then it shows the method, not a result.

Graded

510

Inside the range

80.0%

Waiting for their session

60

+0.0 pts from the 80% target, with a 95% interval of 76.3% to 83.2% over 510 graded ranges and 17 sessions. Median range ±3.3%. Thirty stocks on the same day move together, so the real uncertainty is wider than that interval. A rate above 80% in a calm stretch can simply mean the ranges are a little wide; this becomes evidence once many more sessions are graded, kept separate from the history above.

Updated Thu, 29 Oct 2026 00:52:50 GMT. Script: scripts/journal.mts, run by a scheduled GitHub Action.

Does it read you right?

Shunt turns your sentence into a trade before it measures anything. To test that, 24 sentences were written the way traders type, each with the trade a careful person would read from it. Every field (stock, side, size, days, loss limit) is checked.

First run, rules alone

14 of 24

First run, with Claude

22 of 24

After the fixes, rules alone

24 of 24

After the fixes, with Claude

24 of 24

The fixes were made after seeing the misses, so the second row is not an unbiased score. The first run is. Claude only reads words here; it never produces a number you see. Claude version claude-opus-5-5.

A second test, sealed before the change

Those 24 sentences were fixed after their misses were seen, so they cannot score the language layer again. Before Shunt learned to answer questions, 42 new cases were written and committed with the answer each should get: trades in slang, unclear or contradictory inputs that must be asked about, attempts to push past the limits, follow ups, and ten questions about a trade. The same cases ran on the live site before and after the change, paced under the model's allowance. The cases were not edited between runs.

Missed after the change: "rTSLA 3k, 2 days, risk 2%" (asked conflict instead). That miss was new: the check for two different loss limits read “risk 2%” as both $2 and 2%. It was fixed after this run and is not counted here. Cases and runner: scripts/lui-sealed.

Orders through Bitget Agent Hub

Shunt predicts what getting in and out will cost: the fee plus how far your size walks through the order book. To test that, each trade below was checked, sized to its loss limit, sent to Bitget Demo through Agent Hub's CLI (bgc --paper-trading), then closed straight away. The fill was compared with the price on the screen just before the order.

Orders placed

118

Round trips

57

Stocks

8

Median cost predicted

15.7 bps

Median cost charged

16.8 bps

Within 5 bps of the call

54 of 57

Charged no more than predicted

24 of 57

Every attempt, not just the good ones. 113 trades were tried and 57 completed a round trip. The rest: 31 on stocks Bitget Demo does not list, 4 no measured history for the stock, 9 Bitget Demo order size cap, 11 Shunt priced no cost for it (book or data unavailable), 1 order lookup failed. The trades were drawn from a fixed random seed, and most came out as “doesn't fit” at the size asked (47 of 57), so most were placed at a smaller, fitted size. The charged cost was higher than the predicted cost in 33 round trips.

One bps is one hundredth of one percent of the trade. 18 bps on a $5,000 trade is $9. Demo fills are simulated matching on Bitget's paper account, not live liquidity, so this tests the fee and book arithmetic, not the real market.

Bitget Demo (paper trading) via Agent Hub bgc. Demo does not list 31 other stock perps Shunt checks, so those could not be ordered. NVDA, META, AMZN, AAPL and TSLA orders were placed on the same Demo account that another of the builder's projects also trades on; every order here carries a shunt client id so the two can be told apart. The public site never holds an exchange key; these were placed from the builder's machine.