The Same Six Months, on the New Rules

The new rules take 19 positions over the same six months — and they are not the 132 the book took; only two of those clear all four price gates, so this is a different set of names, not a filtered version of ours (see the reconciliation below) — 16 of the 19 that cleared the gate fell in June to September, and only 3 in April and May, because breadth was below its own mean for most of that stretch. Every one of them is a name that cleared all four price gates, carried RS ≥ 90 and EPS growth ≥ 25%, ranged 3–7% a day, and had a pivot low close enough to put the stop inside 5% — and each was entered only on a session when universe breadth was above its prior five-session mean.

Median hold 14 sessions. The left tail is bounded by construction: the gate refuses a stop more than 5% below the close it was judged on, and the exit is taken at the next open after a close beneath it. The worst trade in the replay lost -6.5% and every loss is inside 7% — against a book whose six-month record contains losses of 11% and 45%.

Why 10 rows show a stop wider than the cap. The gate is judged on the signal close; the fill is the next session's open, which can gap above it. The stop stays where the chart put it, so the distance measured from the fill is wider than the distance the gate approved. That is not a breach — it is the gap risk the disaster line at stop × 0.95 exists for, and it is why an entry more than a few percent above the signal close is worth refusing.

The replay above asks what today's rules would have bought. This asks something different and, for a book that already exists, more useful: keep every name we actually bought — including the ones the entry gates would now refuse — and change only the two things that are not about selection. Size each position to a ₹40,000 risk budget under a ₹5 lakh cap, and exit on a close beneath the stop at the next open, with the disaster line at stop × 0.95 and breakeven at +8%.

This page says the ₹21.9L harness could not be reproduced and asks you to treat ₹13–17L as the finding. That was wrong. Replayed against the real bars of all 132 trades, the figure comes back at ₹23.8L — a 3.4% gap that is slippage and horizon assumptions, not a different result.

What was never true is that it came from sizing, or from a different entry rule, or from a shortlist. The same 132 fills at ₹40,000 of risk and a ₹5,00,000 cap give:

Every rupee of that gap is the exit rule: stop on a close below structure rather than a touch, sold at the next open, breakeven at +8%, a 20-day trail at 0.97, and the disaster line at stop × 0.95. It is the single largest measured effect in this system, and for a day it was being reported as an unexplained discrepancy between two harnesses.

Both figures are now pinned in tests/regressions/test_the_book.py against the real fills and 10,705 real bars, so a change to sizing moves one number, a change to the exits moves the other, and neither can drift unnoticed again. The missing-stop caveat below still stands: 39 of the 132 have no recorded entry stop and both figures assume 5% for them.

Two things above need qualifying. Neither is an arithmetic error — the ₹12.44L → ₹23.8L pair is properly paired, same fills, same sizing, same 60-session horizon, and it still reproduces.

First, it assumes capital we do not have. Both rows size every one of the 132 trades independently at ₹40,000 of risk capped at ₹5L, with no cash pool and no slot limit. Measured afterwards, that book needs a peak of ₹242.5L across 49 simultaneous positions on 8 May. The account is ₹60L. Sequenced through one ₹60L pool with twelve slots — so that a full account simply misses the next signal — the same exit rule returns ₹12.60L, or 21.0%, taking 47 of 127 signals and missing 80. The ₹23.8L is the right number for the question "what are these exits worth per trade"; it is not an amount this account could have made.

Second, "the single largest measured effect in this system" is no longer true. Holding sizing fixed at the book's own quantities, the exit rule is worth +3.4 points over what the book actually did (6.1% → 9.4% on ₹60L). Holding the exit rule fixed and changing only the sizing — ₹40,000 of risk, ₹5L a name, twelve slots, against the 35 thin positions the book really ran at its April peak — is worth +11.6 points (9.4% → 21.0%). Position sizing is roughly three and a half times the exit rule. The exit rule remains the largest thing we have measured about exits, which is what the sentence should have said.

Both corrections come from cortex/india_strategy/retrospectives/2026-09-25_ms_vs_ours_v3.py, which runs every arm through a single ₹60L pool. See MarketSmith's Exits Against Ours for the full study and for two superseded versions of it that got this wrong in the other direction.

An independent audit of this work found an assumption this page never stated, and it is a material one. 41 of the 134 closed trades carry no initial_stop_price at all — 39 of the 132 in the research sample. Checked directly against production: count(*) filter (where initial_stop_price is null) on trading_trade returns 41. Twenty-eight of those rows say explicitly that they were backfilled from broker fills, and the audit inspected five database backups from May to September without finding a single original entry stop.

Every number on this page treats those trades as having a 5% stop, because that is the convention the reconstruction adopted. That was never disclosed here, and the result is sensitive to it. The audit’s cash-constrained replay puts the range at:

The audit’s primary figure of ₹16.87L sits inside the ₹13–17L range this page already published, which is the one thing here that holds up unchanged. What does not hold up is the silence: a result that halves under a different treatment of a third of its inputs needs that treatment on the page, not in a script. The missing stops should stay recorded as unknown rather than filled in.

The audit also separates the effect of the new exits from the effect of sizing and selection, which this page does not: at equal sizing, limits and admission rules it puts the exit change at +₹46,430, far below the headline improvement. Most of the gain here is sizing and selection, not the exit rule itself.

So the ladder is: −₹1.9L from measuring six months instead of ten, and −₹4.3L from the breadth gate, which removes roughly seven of every ten setups that clear the shortlist. The ₹21.9L is a different exercise altogether — it keeps every name we actually bought, including the ones the entry rules would refuse, and only changes the sizing and the exit.

Eighteen against 132 invites the reading "the new rules rejected 114 of our trades". That is not what happened. Measured at entry, our own 132 fills fail the gates like this:

The number to take from it is the comparison, not the level: far fewer positions, each sized to the same risk, with the left tail bounded by a cap the old book did not enforce.