docs(sl-hunting): 21 Sep — the refused gate, the v5a correction, and the 22 Sep pre-open note

[DoRmAmMu1997]

Three commits, and no gate was built: today’s comparison, the backtest that refused the gate I proposed, and the v5a correction that backtest forced.

The comparison

Source: YLhsZu6D658, uploaded 2026-09-21 10:16 IST. Agent session: two trades, both LONG, basket −₹3,200.50.

Fourth session running with the same read and the same side — both LONG on a flat open, both citing the measured FLAT classification. IH booked one long at his normal target and stopped.

#	entry	exit	basket
1	LONG 23,387.25 @ 09:26	09:32, stall at the 23,400 round number	+₹631.50
2	LONG 23,385.95 @ 10:10	10:14, AI_STOP	−₹3,832.00

Why nothing was encoded

Each candidate resolves to a rule already in the corpus:

• "Two indices break out while the third cannot cross — their breakout converts into a failure" → v4i, which already records IH saying almost exactly this and already spells out the equal-lot mirror consequence. Sharper phrasing, no new test, so v4i stands unedited.
• "At the open the biggest risk is the PREMIUM" → PREMIUM NON-CONFIRMATION's IT CAN GO NEGATIVE, NOT MERELY WEAK, mechanism and measured case already present.
• The discipline passage → fourth repeat this week; v3y and RISK carry it.

The gate was priced and refused

I proposed a stop-distance floor (or mirror-premium cap) and it was approved to build. Priced first against all 149 closed journal trades, 2026-07-02 → 2026-09-21, base +₹57,988.50. The data refuses it on both shapes.

floor	refused	P&L of refused	delta
8 pts	3	+₹14,208.75	−₹14,208.75
10 pts	16	+₹28,458.75	−₹28,458.75
12 pts	34	+₹23,393.25	−₹23,393.25
14 pts	58	+₹65,413.00	−₹65,413.00
16 pts	85	+₹36,906.50	−₹36,906.50

A lot cap is negative at every level too — cap 5 costs −₹8,417, cap 2 costs −₹25,534.

The relationship runs the other way in this book:

	n	total	mean	win %	stopped out %
stop < 12 pts	34	+₹23,393.25	+₹688.04	53%	32%
stop ≥ 12 pts	115	+₹34,595.25	+₹300.83	43%	45%

Tighter stops earn more per trade, win more often, and are stopped out less often. Today's −₹3,832 sits in the 8–10 point bucket, which is +₹14,250 across thirteen trades.

The three "instances" don't survive their own dates

date	wider stop	tighter stop
2026-09-08	24.40 pts → −₹786.25	17.90 pts → +₹716.00
2026-09-09	15.15 pts → −₹1,202.00	11.85 pts → −₹499.50
2026-09-21	15.25 pts → +₹631.50	9.45 pts → −₹3,832.00

Across the three dates the tighter stop was the worse trade exactly once — today. My first commit's framing ("three measured instances of the same mechanism") was wrong; it's corrected in place, and the section title no longer claims a third violation.

v5a is now corrected, not just flagged

You agreed to soften it, so the rule carries the correction itself. The bullet keeps the mechanics —

"MECHANICALLY the probability the stop is hit rises as it tightens, and a constant rupee loss suffered more often is not a smaller loss. BUT THIS BOOK DOES NOT SHOW IT…"

— then states the 32%/45% stop-out split, the +688/+301 means, and the priced floor result. It closes with what to do with tightness, because a correction without an instruction just leaves a gap: treat it as a reason to CHECK the entry, never on its own as a reason to refuse it or to size down.

The two bullets beside it are arithmetic — the equal-lot mirror sitting outside the budget, and per-unit slippage — and are unchanged. Only the statistical claim moved.

The v5a marker test now pins both halves, the mechanics and the refutation. One pre-existing assertion was relaxed from "A constant rupee loss…" to "constant rupee loss…" because that sentence is now a mid-sentence clause; its content is unchanged.

Negative-tested 6 ways, all 6 caught, plus a passing control. The mutation that matters most — reverting the bullet to its original unqualified wording — is caught, so the stop-floor gate this PR refused cannot be quietly re-derived from the rule later.

Gates: 606 master · 1579 pytest · ruff 0.16.7 · mypy (80 files) · compileall.

---

Addendum: the pre-open note for 22 SEP 2026

Source: Prediction For 22 SEP 2026 (1ariOZ4dAVQ, uploaded 2026-09-21 21:25 IST).

The branch shape moved, and it moved asymmetrically. Both gap branches reversed from last night; the flat branch did not.

open	21 Sep	22 Sep
gap up	SELL	BUY
flat	BUY	BUY (unchanged)
gap down	BUY	SELL

Carrying yesterday's read forward inverts both gap cases while still looking right on the flat case — a half-correct carry-over is harder to spot than a wholly wrong one. The note spells out both shapes side by side, and the test pins both strings.

The two branches have different premises, and only one is the familiar seated-crowd argument.

• flat / gap up → BUY. The rise was positive but gradual, never sharp, and SENSEX never crossed its round number. A slow move breeds no greed, so buyers came but "did not go holding the trade". An unseated buyer crowd is nobody to squeeze — so follow.
• gap down → SELL. Not because sellers are seated. Because a fall is only covered if an operator commits money; retail cannot produce that bounce, and "we can get TRAPPED there" planning for one.

That second premise is net-new for this channel. Every previous note reasoned about who is trapped; this one reasons about who can pay. The test asserts the negation — "NOT because sellers are seated" — precisely because substituting the house default would look harmless while silently replacing the claim.

All three indices carry identical branches with no per-index qualification, unlike 21 Sep where NIFTY alone was limited to a small gap down. That absence is asserted too, because an absence is what a summariser invents into.

Levels came through clean for the first time in several nights — nothing to reconstruct:

	resistance	support
NIFTY	23500 / 23570	23350 / 23270
BANKNIFTY	57000 / 57200	56400 / 56100
SENSEX	75200 / 75500	74700 / 74350

SENSEX is an exact one-rung shift of yesterday's ladder, and 75000 — the round number the gradual move never crossed — sits between the new support and the new resistance, which corroborates the spoken context. NIFTY and BANKNIFTY lift without being clean shifts (NIFTY carries 23270 and 23500 over, BANKNIFTY only 56100), and the test comment says so rather than generalising the SENSEX pattern to all three.

Negative-tested 10 ways, all 10 caught, plus a passing control — including flipping the flat branch alone, carrying yesterday's gaps forward, swapping in the seated-sellers premise, inventing a per-index qualifier, drifting a support by 50 points, and re-dating the note to the previous session.

Gates re-run on the merged tree: 606 master · 28 market-data-health · 1579 pytest · ruff 0.16.7 · mypy (80 files) · compileall · bandit.

🤖 Generated with Claude Code