feat(sl-hunting): v5h - in a range the crowd read cannot pay, however right it is

Signal Generators/SL Hunting AI Agent/sl_hunting_doc.md

@@ -7474,3 +7474,95 @@ that can end at 10:40.
 NOTE FOR NEXT TIME: the cause is not recoverable from the log. SLH-004 logs
 this class with CANNED, content-free messages on purpose, so a future outage
 of the same shape will again need the operator to say which it was.
+## v5h - in a range the crowd read cannot pay (18 Sep)
+
+Source: `Z3D-p6RvPc0`, "Live Bank Nifty Option Trading", uploaded 2026-09-18
+10:50 IST. Compared against the agent's own session: four trades, all SHORT,
+basket **+Rs.4,308.25**.
+
+### The day both sides got right
+
+Third session running where IH and the book reached the same read and took the
+same side. The open was +0.275% on NIFTY and +0.209% on BankNIFTY -- a small UP
+gap that SLH-017 classified FLAT, which is exactly the case the classification
+was built for. All four entries cited the verdict and the threshold explicitly,
+and took the note's flat branch. IH: "almost flat opening... until these levels
+are crossed we can follow the continued selling", and he bought puts.
+
+He booked one trade, took the profit, and stopped -- "because we had a loss
+yesterday and a loss the day before, this week we have not made much, so if we
+are getting a chance to make profit today, let us take the exit."
+
+### What the book did instead
+
+| # | entry | exit | held | basket |
+|---|---|---|---|---|
+| 1 | SHORT 23,322.60 09:20 | 09:34 momentum stall | 14m | **+4,466.25** |
+| 2 | SHORT 23,310.05 09:47 | 09:54 bullish engulfing at the stop | 7m | -1,959.00 |
+| 3 | SHORT 23,319.00 10:04 | 10:08 profit decay | 3.5m | +338.00 |
+| 4 | SHORT 23,318.95 10:23 | 10:27 profit growth stalled | 3.5m | +1,463.00 |
+
+The first trade was the day. Trades 2, 3 and 4 netted **MINUS 158** between
+them, for three more entries, three more mirrored baskets and three more
+exposures. The session's entire range was about 44 points (23,292.55 to
+23,336), and **the market never made a new low after trade 1's exit** -- all
+three re-entries went in without one.
+
+Nothing was misread. Every entry correctly applied the classification and the
+note; every exit cited a real rule by name (v4f, v4w, BOOK WHEN THE PROFIT
+STOPS GROWING) and booked into a stall rather than waiting. The exit discipline
+was good enough that three trades in a dead range still came out roughly flat.
+
+### The net-new claim
+
+IH states the limit on his own method, which nothing in the corpus had:
+
+> "when it has to give momentum in a direction, you will get the momentum
+> overall. But if it is a RANGE-BOUND market trading in a small range, or the
+> market simply does not want to give momentum -- a volatile market -- then
+> **whether you find out the SLs or NOT, the market can give you a loss**."
+
+Correctness is not the variable. And he says it while diagnosing his OWN loss
+from 17 Sep: "yesterday was the same situation when we sold, sellers' SLs were
+not available there either. But the market was very sideways... when the market
+works inside a RANGE, that is where loss happens." That is v5g, confirmed by
+its own source one day later and generalised: v5g says a FOLLOW can idle
+because nothing is trapped; v5h says the read itself -- hunt or follow, right
+or wrong -- stops being tradeable when the session refuses direction.
+
+Encoded as **IN A RANGE THE CROWD READ CANNOT PAY, HOWEVER RIGHT IT IS (v5h)**,
+placed directly after v5g as its general case.
+
+**It carries a checkable arm, because "is this a range?" is exactly the
+judgement that fails in real time.** BEFORE RE-ENTERING THE SAME SIDE, REQUIRE
+A NEW EXTREME BEYOND YOUR PREVIOUS EXIT. If price has not extended past where
+you booked, you are re-entering the same move and the session is ranging.
+
+**And it explains why nothing already in the book stopped this.** The pre-open
+note's branch is true from the open to the close -- it selects a SIDE, it does
+not refresh -- so it can be cited four times without ever going stale. All four
+entries did cite it, almost verbatim, each dressed in a different pattern name
+(`flat_open_failed_spike_breakdown`, `..._shooting_star_continuation`,
+`..._double_top_evening_star`, `..._seller_follow_bearish_cluster`). SLH-005's
+five-minute cooldown was cleared every time (13, 10 and 15 minutes apart), and
+v3q's re-entry gate asks for "a nameable NEW trapped crowd" -- which is vacuous
+on a FOLLOW whose whole premise is that nobody is seated. On a follow, the new
+extreme has to do that work instead.
+
+Negative-tested 10 ways, all 10 caught, plus a control mutation of unasserted
+prose that correctly did not trip the test.
+
+**Considered and not encoded.** IH's chart-repeat argument tonight -- "it is
+trying to make almost the same chart as yesterday, so the chances of the chart
+repeating become quite low", and "if the buyers made profit yesterday, today
+they should not get that profit" -- is v4f (THE CHART DOES NOT REPEAT TWO DAYS
+RUNNING), which already carries the same mechanism ("a shape everyone watched
+yesterday is a shape everyone is ready for today"). The cohort phrasing is
+sharper but adds no new test, so v4f stands unedited. His capital-sizing and
+emotional-discipline passage repeats v3y and the RISK section again.
+
+**Worth watching, not yet a rule.** v5g shipped last night and today's exits
+did book on stalls -- but they cited v4f and BOOK WHEN THE PROFIT STOPS GROWING,
+both of which predate it, so there is no evidence yet that v5g changed
+behaviour. Its distinctive claim (name the window BEFORE entering) did not
+appear in any entry rationale. One session is too few to judge; check again.

Signal Generators/SL Hunting AI Agent/sl_hunting_knowledge.py

@@ -803,6 +803,40 @@
   trade has to be carried by momentum ALREADY present, and a stall is the exit
   rather than something to wait through. If the move has not started, the
   premise has not failed -- but the trade has, and those are different things.
+- IN A RANGE THE CROWD READ CANNOT PAY, HOWEVER RIGHT IT IS (v5h). The general
+  case of the rule above, and the only limit on this whole method stated by the
+  person it came from. v5g says a FOLLOW can idle because nothing is trapped.
+  This says the read itself -- hunt or follow, correct or not -- stops being
+  tradeable when the session refuses to move in any direction at all.
+  IH PUTS THE LIMIT ON HIS OWN METHOD: "when it has to give momentum in a
+  direction, you will get the momentum overall. But if it is a RANGE-BOUND
+  market trading in a small range, or the market simply does not want to give
+  momentum -- a volatile market -- then whether you find out the SLs or NOT, the
+  market can give you a loss." Correctness is not the variable. He said it while
+  diagnosing his own loss the day before: "yesterday was the same situation when
+  we sold, sellers' SLs were not available there either. But the market was very
+  sideways, going up this much and down this much. When the market works inside
+  a RANGE, that is where loss happens."
+  MEASURED ON THIS BOOK (2026-09-18), on a day it was RIGHT and still learned
+  this. Four shorts, every one correctly taking the pre-open note's flat branch
+  off the computed FLAT classification, in a session whose entire range was
+  about 44 points (23,292.55 to 23,336). The FIRST trade caught the only real
+  leg: +4,466.25 in fourteen minutes. The next three returned -1,959.00,
+  +338.00 and +1,463.00 -- MINUS 158 between them, for three more entries,
+  three more mirrored baskets and three more exposures. Nothing was misread;
+  the session simply had nothing left to pay.
+  THE CHECKABLE TELL, because "is this a range?" is a judgement and your own
+  re-entry is not: BEFORE RE-ENTERING THE SAME SIDE, REQUIRE A NEW EXTREME
+  BEYOND YOUR PREVIOUS EXIT. If price has not extended past where you booked,
+  you are re-entering the SAME move, the session is ranging, and the next trade
+  is a tax rather than a trade. On this day the market never made a new low
+  after the first exit, and all three re-entries happened anyway.
+  AND A STANDING PREMISE IS NOT A SECOND REASON. The pre-open note's branch is
+  true from the open to the close: it selects a SIDE, it does not refresh. All
+  four entries cited it almost verbatim and dressed each one in a different
+  pattern name. v3q's re-entry gate asks for a nameable NEW trapped crowd, which
+  is vacuous on a follow whose whole premise is that nobody is seated -- so on a
+  follow it is the new extreme that must do that work instead.
 - A FORECAST OF WHO WILL ARRIVE IS NOT EVIDENCE OF WHO IS SEATED (v4e). The single
   most expensive error available in this method, recorded from a LOSING IH session
   (11 Aug 2026) so it is not learned the hard way. He stated the disqualifying fact

Tests/Signal Generators/SL Hunting AI Agent/test_sl_hunting_schema.py

@@ -2927,3 +2927,63 @@ def test_v5g_an_eviction_leaves_you_in_the_same_sideways_stretch():
     assert "carried by momentum ALREADY present" in rule
     assert "a stall is the exit rather than something to wait through" in rule
     assert "the premise has not failed -- but the trade has" in rule
+
+
+def test_v5h_a_range_makes_the_crowd_read_unpayable():
+    """v5h (18 Sep): learned on a day the book was RIGHT and still gave it back.
+
+    Four shorts, every one correctly on the note's flat branch off the computed
+    FLAT verdict, inside a 44-point session. The first made +4,466.25; the next
+    three netted MINUS 158 between them. Seven ways an edit breaks it:
+
+    1. Losing that this is v5g's general case -- apart they look like two
+       moods about quiet markets rather than one rule and its limit.
+    2. Losing that it is a limit on the METHOD, stated by its own source. That
+       provenance is what stops it being read as generic caution.
+    3. Losing "whether you find out the SLs or NOT" -- the claim is that
+       correctness is not the variable, which is the surprising part.
+    4. Losing IH's own retrospective, which is what makes it a pattern rather
+       than one bad day.
+    5. Losing the measured asymmetry. "Four trades, small profit" is a
+       different and much weaker fact than "the first was the whole day".
+    6. Losing the checkable tell. Without the new-extreme test the rule needs
+       the model to first agree it is in a range, which is the judgement that
+       fails in real time.
+    7. Losing the standing-premise clause, which explains why the note could be
+       cited four times and why v3q's gate did not catch any of them.
+    """
+    prompt = build_system_prompt()
+    rule = _flat_rule(prompt, "IN A RANGE THE CROWD READ CANNOT PAY")
+
+    # 1 and 2. Its relationship to v5g, and whose limit it is.
+    assert "HOWEVER RIGHT IT IS" in rule
+    assert "The general case of the rule above" in rule
+    assert "the only limit on this whole method stated by the person it came from" in rule
+    assert "hunt or follow, correct or not" in rule
+
+    # 3. The surprising claim, in his words.
+    assert "IH PUTS THE LIMIT ON HIS OWN METHOD" in rule
+    assert "whether you find out the SLs or NOT, the market can give you a loss" in rule
+    assert "Correctness is not the variable" in rule
+
+    # 4. His own retrospective on the prior session.
+    assert "sellers' SLs were not available there either" in rule
+    assert "When the market works inside a RANGE, that is where loss happens" in rule
+
+    # 5. The measured asymmetry, with the numbers that carry it.
+    assert "on a day it was RIGHT and still learned this" in rule
+    assert "about 44 points (23,292.55 to 23,336)" in rule
+    assert "+4,466.25 in fourteen minutes" in rule
+    assert "MINUS 158 between them" in rule
+    assert "the session simply had nothing left to pay" in rule
+
+    # 6. The checkable tell, and why it is phrased that way.
+    assert "is a judgement and your own re-entry is not" in rule
+    assert "REQUIRE A NEW EXTREME BEYOND YOUR PREVIOUS EXIT" in rule
+    assert "the next trade is a tax rather than a trade" in rule
+    assert "the market never made a new low after the first exit" in rule
+
+    # 7. Why a standing premise supplies no freshness, and why v3q missed it.
+    assert "A STANDING PREMISE IS NOT A SECOND REASON" in rule
+    assert "it selects a SIDE, it does not refresh" in rule
+    assert "asks for a nameable NEW trapped crowd, which is vacuous on a follow" in rule