The ₹26.96L replay, corrected

7 October 2026 · the same 127 opportunities, run under the rules we actually trade · research, not realised performance

This page is a backtest. It is not the account. The live record — 159 closed trades, ₹4.16L realised, a 32.1% win rate — is a separate page and a different kind of evidence. Nothing here is money that was made.

What was wrong

The published replay allowed several ₹5L lots in one stock, and it applied no market filter. Neither matches how the book is run: one ₹5L lot per stock, and no entry while the breadth and trend gates are shut.

Both were documented rather than hidden, which is why nobody caught them. book-v3-policy.md said “twelve independent lots, including separate lots of the same stock”, and EXECUTION_PROFILE.md said “several independently tagged lots in the same symbol are permitted. This is not permission to pyramid an existing lot” — a sentence that contradicts itself, since several ₹5L lots in one symbol is pyramiding whatever the tags say.

In practice that meant ₹15L in LAURUSLABS and ₹10L in WELCORP, and a twelve-slot book holding twenty stocks of which four were doubled or tripled.

The correction, step by step

variant	trades	stocks	realised	open mark	total
as published	25	20	₹7,06,989	₹19,88,658	₹26,95,647
one lot per stock	26	26	₹4,49,163	₹11,98,090	₹16,47,253
+ market gate	18	18	₹7,01,272	₹6,10,267	₹13,11,539

Removing the duplicates does not simply delete lots — the freed slot and cash go to the next eligible name, so trade count rises from 25 to 26 and twenty-six distinct stocks get funded. The book ran to twelve slots full and ₹4,462 of cash: capital was the binding constraint, not the supply of ideas.

Why the fall is this large

Two effects, and the second is the one that is easy to miss.

The dropped lots were the three biggest winners. WELCORP's second lot alone was +₹5.63L, more than the entire realised P&L of the corrected book. That asymmetry is structural rather than bad luck: you only get a second lot in a stock that is already working. The list offers WELCORP again on 22 May because WELCORP was strong on 21 May.

And every replacement lost money — six lots, −₹37k to −₹65k each. The capital freed in mid-2026 reached the names next on the list, which were PSU-bank and IT sector ETFs, BHARATFORG, BHEL and VEDL: the sectors that then fell.

The market gate, measured on real setups

The gate was open on 19 of the 56 entry dates — 34%. It refuses 25 of the surviving opportunities and cuts the book from 26 trades to 18. What it does to the trade profile is the finding:

closed trades only	n	win %	avg win	avg loss	payoff	expectancy
as published	19	63.2%	₹67,483	−₹14,687	4.59	₹37,210
one lot per stock	23	52.2%	₹67,483	−₹32,785	2.06	₹19,529
+ market gate	16	62.5%	₹87,182	−₹28,424	3.07	₹43,830

The gate improves every metric at once, which is rare. Win rate 52.2% to 62.5%, average win ₹67k to ₹87k, payoff 2.06 to 3.07, and expectancy per closed trade more than doubles — ₹19,529 to ₹43,830 — on seven fewer trades. Note it lifts the average win, not only the loss rate: a breakout entered while breadth is rising has more room behind it.

The published 72% win rate and 10.59 payoff were an artefact of the pyramiding. Not because duplicates win more often, but because they diluted the average loss across more, smaller entries — −₹14,687 against −₹32,785 once each stock gets one lot. A 10.59 payoff was never real.

This reverses an earlier conclusion of ours

On 30 September we published that the entry gate showed no measurable benefit — a basket spread of −0.67pp at p = 0.871, stop-out rates of 94.5% against 92.9% — and kept it on asymmetric grounds rather than evidence. That page named its own weakness: it selected on trailing return, a momentum proxy, not the real seven-gate scanner.

This is the test that page said was missing, and it says the opposite. On the 127 actual setups the gate roughly doubles expectancy. The earlier result is not withdrawn — it measured a different population and remains true of that one — but the gate should no longer be described as unproven.

What this is not

Sixteen closed trades. One six-month window containing one correction and one recovery. The slot count, the ADR band and the gate thresholds are all parameters that fit beautifully to a sample this size; two of them have already died here when measured properly.
₹6.10L of the ₹13.1L is still an open mark on two names. The realised part is ₹7.01L.
The gate reconstruction is partial. It uses the two legs price can settle — breadth rising against its prior five-session mean, and the Nifty above its 20MA. It does not reproduce the distribution-day count or the follow-through-day machinery. The real gate is at least as strict, so ₹13.1L is an upper bound.
Charges are still absent. The replay models 0.5% slippage and no brokerage, STT or stamp duty. On the real book those run 0.1044% of turnover — 10.1% of gross realised profit, three quarters of it STT.
The scanner is still not in the loop. These are 127 supplied opportunities, not what the live seven-gate scan would have proposed.

The number to quote

For the 2026 April–September window, on ₹60L, twelve slots, one ₹5L lot per stock and no entry while the market filter is shut: ₹13.1L marked, of which ₹7.01L is realised. Not ₹26.96L.