T6: Strategy zones and the forgiving win, measured on our own data

Worker T6, 2026-09-26 00:37 to 00:51 Beirut. Read-only on copies in data/. Every number below comes from scripts/T6-zones.py or scripts/T6-symmetry.py and is saved as CSV in data/T6-*.csv. Rerun: cd projects/caverio/companion-2026-09-26 && python3 scripts/T6-zones.py && python3 scripts/T6-symmetry.py (about 25 s CPU each, no network). Evidence grade for every number here: A (our stores, script saved).

0. The answer in six lines

1. A forgiving win on its own mostly measures volatility, not direction. After an alert, 24% of tokens touched +20% within an hour, but 28% touched -20%. At a random moment on the same tokens it was 8% and 7%. Triggers mark moments when things move, in both directions (T6-symmetry.csv).
2. So the definition we can defend in public is a first-passage win: the price rose X% above an honest entry, after costs, held for 5 minutes, before it ever fell Y%. Shown next to the rate for a random moment on the same tokens. Without that baseline the number flatters.
3. On that measure (+20% before -30%, within 24 h), crossings 46.6% vs 38.3% random (+8 pts), alerts 34.3% vs 30.5% (+4 pts), discovery 29.3% vs 32.4% (no advantage, n=41). The advantage is concentrated in the first 1 to 4 hours: crossings +22 pts at 1 h, alerts +10 pts at 1 h.
4. The strict win loses money on every trigger set. The follower's shape (stop -45%, +40% arms a 20% trail, 72 h time exit), net of costs, has a median result of -3.6% (crossings), -11.1% (alerts), -17.0% (discovery). A random moment on the same tokens does about the same: -5.1%, -5.5%, -7.2%.
5. The headline is robust. No cleaning, delay, haircut or trigger-clock change moves it by more than a third. The biggest moves are -24% (discovery, 15 min delay, n=45) and -26% (crossings, 60 min delay). Those are buy-window decay, not fragility. The 5-minute hold rule absorbs bad prints on its own. Strict means do not: without cleaning they reach 10^19 %.
6. Zones can honestly be descriptive clocks and rulers per source, not predictions. "Minutes since trigger versus when past winners from this source peaked" and "this dip versus the dips past winners sat through" are backed by numbers. "Buy here" and "sell here" are not. A fading alert is not a sell signal: it is followed by +20% more often than an entry alert (48.7% vs 40.7%).

1. Definitions (exact, with the implementing line in scripts/T6-zones.py)

Price path. Deep-paper crossings use deep-paper/book.sqlite observations. Discovery and alerts use ticks/ (both are dexscreener prints via market-refresh). The candle stores cannot serve: candles-onchain and candles-bsc keep a rolling ~12 h window (719 candles per pool), and ohlcv/ keeps 48 candles per token from one outcomes pull.

Term	Definition	Line
Stale repeat	A print whose price equals the previous raw print exactly. Dropped first.	60
Misquote episode (new, see §1.1)	A print more than 3x away from the last accepted print, where liquidity moved the same way by at least 0.8 of the price move in log terms. Rejected until a print comes back within 3x.	63, 72
Phantom	A lone print under 0.75x of both neighbours (the audit test). Base settings also drop the mirror lone-high, over 1.333x both neighbours, because a lone high inflates any best-price measure.	81, 82
Trigger time	Discovery: seenAt (when our intake saw the call); calledAt is tested as a sensitivity. Crossings: observationAt of the deep_crossing event. Alerts: at of the alert row.	107, 191, 203, 209
Honest entry	The first cleaned print at or after trigger + reaction delay (base 60 s; also tested 0 s, 300 s, 900 s, 1800 s, 3600 s). If there is no print within 15 min of that time, there is no entry and the trigger is excluded and counted.	108, 109
Cost haircut	Per side, on entry and exit: 1% if pool liquidity at entry is under $50K, 0.5% at or above. This stands in for DEX fee plus slippage on a small ticket. Sensitivity: flat 1.5% per side (the deep-paper replay fee). Net gain = exit × (1 - h) / (entry × (1 + h)) - 1.	90, 92
Touch best	The highest cleaned print within H hours after entry, net of the haircut. Reported only as the loose comparison.	135
Held best ("for long enough")	The highest level L such that every cleaned print from some time t until the first print at or after t + 5 min was at or above L, with the window closing inside H. Net of haircut.	122, 128, 137
Forgiving win (X, H)	Held best within H hours ≥ +X% net.	137, fw()
First-passage win (X, Y, H)	The held +X% level started holding before any cleaned print at or below -Y% (gross) from entry, within H.	156, 162
Strict win	Enter as above. Exit at the first print at or below -45% (stop). Once a print reaches +40%, exit at the first print at or below 80% of the peak (trail). Otherwise exit at the last print before 72 h. Fills happen at the crossing print, not at the level. Costs are charged on both sides. Win = net > 0. If the path ends before 72 h, the trade is marked at the last print and flagged unresolved.	171 to 177
Time to best	Minutes from entry to the highest cleaned print within 24 h (and 72 h).	145
Drawdown before best	The lowest cleaned print between entry and that best, relative to entry (gross).	147
Giveback after best	The lowest print after the best, within the same window, relative to the best.	149
Matched control	Same tokens as each trigger set. A pseudo-trigger every 4 h from the token's first print + 1 h, same engine and cleaning. Lift = trigger share minus control share.	249
Band	Market cap at the entry print: <250K, 250K-1M, 1-5M, 5-20M, ≥20M.	94
Censoring	If the path ends before H, non-wins count as non-wins (a lower bound). The censored non-win count is in every grid row (T6-grid.csv, column censored_nonwins).	139

1.1 A trap the audit's phantom test misses

The lone-print test cannot catch multi-print misquotes. Example from ticks: solana:JE3HT7SbCgXDQWV6xp3oiiAisDzq4HyZ8wyEVBDCs45Z, a token around $0.001 with $150K liquidity, prints $2.56, then $10.7 for five consecutive ticks, then $9.09 and $14.10 in pairs. Liquidity reads $370M to $938M on those ticks. This happened eight separate times between 09-11 and 09-25. The quote is for a different pair or unit. Because the prints come in runs, no single print is "lone".

Before this filter, the strict mean on alerts was +485%, and with no cleaning at all it was 2.7 × 10^19 %. The organic signature is different: trading moves pool USD liquidity by about the square root of the price move, so a price and liquidity jump in proportion is not organic. The filter rejects 229 tick prints and 33 book prints. Bolo should add this test next to the phantom check (audit fix 4). The raw sidecar (bead 44ht) is what would confirm it.

Cleaning stage counts, base settings, all stored paths (T6-summary.json):

Store	Raw prints	After stale repeats out	After misquote out	After phantom (low+high) out
ticks	422,443	300,833 (28.8% were repeats)	300,604 (-229)	300,251 (-353)
deep-paper book	165,889	112,363 (32.3% were repeats)	112,330 (-33)	112,272 (-58)

2. Trigger sets and coverage

Set	Triggers	With honest entry (60 s, ≤15 min)	Distinct tokens	Window	Note
(a) Discovery calls	652 rows → 479 with chain+address → 308 unique (post, token)	41	35	09-10 to 09-25	Ticks start when a token becomes a situation, not when it is called. Median gap from seenAt to the first tick is 70 h. Discovery cannot be measured properly today.
(b) Deep-paper crossings	165 deep_crossing events (59 entered, 36 pre_start, 23 repeat_token, 19 quiet_hour, 10 unconfirmed_print, 1 liquidity_floor skips)	148	77	09-22 04:24Z to 09-25 21:20Z	3.7 days only
(c) Alerts, entry-type (forming / confirming / active)	5,469	3,337	591	09-10 to 09-25	2,809 are repeats on a token already alerted
(c') Alerts, fading	2,088	1,232	190	same	Kept separate to test "sell zone" meaning
Controls	discovery 1,279 / crossings 777 / alerts 17,767 pseudo-triggers	516 / 499 / 5,771	35 / 74 / 564	same tokens

The per-trigger register (token, chain, trigger time, entry time, lag, entry price, liquidity, band, outcomes) is data/T6-triggers.csv.

3. Results

3.1 Forgiving win grid, held measure (share of triggers, %; rows H hours, columns X% net)

Full grid for both measures, all sets and controls, with n and censored counts: data/T6-grid.csv. None of these cells has n under 20.

Discovery, n=41 (35 tokens). Control n=516.

H \ X	10	20	30	50	100
1	29.3	12.2	7.3	4.9	0.0
4	39.0	26.8	14.6	9.8	2.4
12	43.9	31.7	24.4	14.6	2.4
24	46.3	34.1	29.3	22.0	7.3
72	51.2	36.6	31.7	24.4	9.8
control 1 h	6.6	3.1	1.4	0.6	0.0
control 24 h	48.3	34.5	23.8	15.1	6.8

Crossings, n=148 (77 tokens). Control n=499.

H \ X	10	20	30	50	100
1	45.3	29.1	18.2	6.8	0.7
4	57.4	47.3	34.5	23.0	8.1
12	62.2	54.7	45.9	32.4	15.5
24	64.9	57.4	48.6	37.2	20.9
72	67.6	60.1	51.4	40.5	24.3
control 1 h	12.6	7.2	4.4	1.8	0.8
control 24 h	56.1	43.1	33.3	25.1	13.6

Alerts (entry-type), n=3,337 (591 tokens). Control n=5,771.

H \ X	10	20	30	50	100
1	23.7	13.5	8.9	4.6	1.3
4	38.0	24.9	18.0	9.7	3.4
12	47.9	34.6	26.3	15.0	5.3
24	54.0	40.7	31.5	19.5	7.4
72	63.0	50.9	40.2	27.0	12.9
control 1 h	6.1	3.2	1.8	0.8	0.2
control 24 h	47.7	33.9	25.4	16.1	7.0

Alerts, fading, n=1,232 (190 tokens): 1 h: 26.5 / 16.3 / 11.8 / 6.8 / 1.6; 24 h: 62.9 / 48.7 / 35.7 / 27.1 / 11.7.

The 72 h row is a lower bound. For alerts at X=20, 890 of the non-wins had paths that ended before 72 h; for crossings, 55.

Touch vs held. Touch runs 7 to 12 points above held at 24 h: crossings 67.6 vs 57.4, alerts 49.6 vs 40.7, discovery 43.9 vs 34.1. That gap is the set of prices that were printed but did not stand for 5 minutes.

3.2 Direction versus volatility (T6-symmetry.csv, gross touch)

Set	+20% within 1 h	-20% within 1 h	+20% within 24 h	-20% within 24 h	Both within 24 h
Discovery (n=41)	26.8	31.7	43.9	68.3	34.1
control	5.8	6.6	43.0	45.2	18.4
Crossings (n=148)	48.0	43.2	68.2	76.4	50.7
control	16.2	16.4	51.7	63.7	31.5
Alerts (n=3,337)	23.9	27.9	51.3	67.3	34.0
control	7.8	6.7	42.2	50.0	19.6
Alerts fading (n=1,232)	27.5	36.2	59.3	75.3	46.5

Triggers roughly triple the chance of a 20% move in the next hour, in both directions. Over 24 h, a -20% move is more common than a +20% move after every trigger type. A forgiving-win number shown alone would be read as "the signal points up". It does not, or only weakly.

3.3 First-passage win with lift (T6-firstpassage.csv)

Share whose +X% net level held 5 minutes before the price ever printed -Y%. Rows are (X, Y); cells are trigger / control / lift in points.

	H	Discovery (n=41)	Crossings (n=148)	Alerts (n=3,337)	Alerts fading (n=1,232)
+20 before -30	1 h	12.2 / 3.1 / +9.1	29.1 / 7.0 / +22.0	12.9 / 3.1 / +9.7	15.7 / 3.1 / +12.6
+20 before -30	4 h	26.8 / 12.4 / +14.4	43.2 / 25.7 / +17.6	22.3 / 14.3 / +8.1	28.1 / 14.3 / +13.8
+20 before -30	24 h	29.3 / 32.4 / -3.1	46.6 / 38.3 / +8.3	34.3 / 30.5 / +3.8	37.1 / 30.5 / +6.5
+50 before -30	24 h	14.6 / 13.2 / +1.5	25.7 / 20.6 / +5.0	15.1 / 13.9 / +1.2	20.4 / 13.9 / +6.5
+20 before -20	24 h	26.8 / 28.5 / -1.7	41.2 / 32.7 / +8.6	28.7 / 27.1 / +1.6	29.3 / 27.1 / +2.2
+20 before -45	24 h	34.1 / 33.5 / +0.6	51.4 / 41.1 / +10.3	38.2 / 32.8 / +5.3	45.5 / 32.8 / +12.6

3.4 Strict and forgiving side by side (T6-strict-vs-forgiving.csv, T6-firstpassage-dims.csv)

Columns: n; forgiving held +20%/24 h; first-passage +20 before -30 in 24 h (FP); strict win rate; strict median net; share that never traded above entry within 72 h. Rows with n under 20 are marked SMALL and should not be quoted.

Set / slice	n	Forgiving 20/24	FP 20/-30/24	Strict win	Strict median	Never above entry
Discovery, all	41	34.1	29.3	36.6	-17.0	7.3
control	516	34.5	32.4	41.1	-7.2	10.3
PowsGemCalls	25	44.0	36.0	44.0	-6.6	12.0
sugarydick	14 SMALL	21.4	21.4	21.4	-33.8	0
passed screen	13 SMALL	46.2		23.1	-45.7	15.4
not passed	28	28.6		42.9	-11.5	3.6
Crossings, all	148	57.4	46.6	49.3	-3.6	7.4
control	499	43.1	38.3	45.7	-5.1	10.8
entered	59	54.2	42.4	42.4	-22.0	8.5
skip: pre_start	36	58.3	50.0	47.2	-5.2	2.8
skip: repeat_token	23	60.9	56.5	52.2	+2.8	13.0
skip: quiet_hour	19 SMALL	68.4	47.4	68.4	+12.1	10.5
skip: unconfirmed_print	10 SMALL	40.0	30.0	50.0	-2.5	0
solana	98	54.1	43.9	46.9	-8.1	9.2
robinhood	44	63.6	50.0	52.3	+6.7	4.5
bsc	6 SMALL	66.7	66.7	66.7	+9.3	0
Alerts, all	3,337	40.7	34.3	42.8	-11.1	6.1
control	5,771	33.9	30.5	42.5	-5.5	11.6
forming	2,505	39.2	35.0	43.4	-8.5	5.9
confirming	455	40.9	30.1	39.6	-27.4	7.3
active	377	50.9	34.7	42.7	-45.7	6.1
trigger: first covered buy	1,854	39.1		43.1	-8.5	6.7
trigger: distinct-trader confirmation	768	45.7		40.9	-32.7	6.5
trigger: thesis-first	466	38.6		47.9	-3.8	3.9
robinhood	1,979	41.3	34.9	44.0	-8.6	6.2
solana	1,085	41.7	35.0	41.5	-14.8	5.6
bsc	273	33.0	27.5	39.9	-15.2	7.3
Alerts fading, all	1,232	48.7	37.1	48.5	-3.0	6.6

By band (alerts, n; forgiving 20/24; FP; strict win; strict median):

Band	Alerts	Control-alerts	Crossings
<250K	56; 14.3; 12.5; 28.6; -42.5	432; 28.5; 26.4; 32.6; -15.0	23; 52.2; 39.1; 30.4; -22.0
250K-1M	585; 42.2; 32.3; 41.2; -21.7	1,340; 40.2; 34.6; 44.6; -4.5	77; 61.0; 46.8; 54.5; +4.9
1-5M	1,659; 47.3; 38.7; 42.6; -23.8	2,446; 38.3; 34.3; 43.4; -7.4	38; 63.2; 57.9; 47.4; -6.0
5-20M	642; 39.1; 37.1; 43.1; -8.0	940; 30.7; 30.1; 41.8; -5.3	2 SMALL
≥20M	392; 17.6; 17.6; 47.7; -0.8	595; 10.9; 10.8; 42.4; -2.8	8 SMALL

Reading the band table: ≥20M almost never gives a +20% forgiving win (18%), but has the least bad strict result. Under $250K has the worst strict result, and 27% never trade above entry at all (rugs and dead launches). The $1-5M band is where an alert adds the most over a random moment on the forgiving measure (+9 pts). Its strict median is still -24%, because the moves are large in both directions.

Strict outcome. Strict win rates sit at 37 to 49% everywhere, with medians negative on every trigger set. The strict result after a trigger is no better than at a random moment on the same token (alerts median -11.1% vs control -5.5%; crossings -3.6% vs -5.1%). The forgiving advantage exists only in the first hours, and a fixed 72 h rule gives it back.

3.5 Time to best, drawdown before best, giveback (T6-timing.csv, T6-holdzone.csv)

Forgiving winners (+20% held within 24 h). Minutes after entry; percent vs entry.

Set	n winners	Time to best p25 / p50 / p75 / p90 (min)	Drawdown before best p50 / p75 / p90	Median best gain	Median giveback after best (24 h)	+20% level first held, p50 / p80 (min)
Discovery	14 SMALL	132 / 310 / 933 / 1229	-13.0 / -27.0 / -41.4	+70%	-49%	86 / 310
Crossings	85	67 / 282 / 776 / 1213	-12.3 / -26.8 / -46.1	+91%	-58%	52 / 186
Alerts	1,359	129 / 407 / 921 / 1233	-12.5 / -25.8 / -40.5	+63%	-47%	144 / 570
Alerts fading	600	101 / 414 / 901 / 1177	-19.0 / -31.1 / -43.7	+77%	-63%	97 / 486
Control-alerts	1,954	263 / 570 / 988 / 1283	-7.6 / -18.6 / -31.2	+60%	-40%	273 / 679

Non-winners peak early: 52% of alert non-winners had already printed their 24 h best by 60 min, and 79% by 8 h. For crossing non-winners the median time to best is 20 min.

Hold curve (share of eventual winners still below their 24 h best at elapsed time):

Elapsed	15 min	30 min	1 h	2 h	4 h	8 h	12 h
Crossings winners (85)	94%	85%	78%	68%	53%	38%	26%
Alerts winners (1,359)	96%	92%	85%	77%	62%	46%	36%
Alert non-winners already at their best	34%	43%	52%	62%	71%	79%	85%

3.6 Sensitivity (T6-sensitivity.csv)

Headline A = forgiving held +20%/24 h. Headline B = first-passage +20 before -30 in 24 h. Cells show the relative change vs base.

Variant	Discovery A / B	Crossings A / B	Alerts A / B
base value	34.1 / 29.3	57.4 / 46.6	40.7 / 34.3
delay 0 s	-5% / -15%	+6% / +7%	+4% / +7%
delay 300 s	+5% / -2%	-6% / -11%	-1% / -1%
no phantom classification	0 / 0	+1% / +1%	0 / 0
phantom low only (audit test)	0 / 0	0 / 0	0 / 0
stale repeats kept	0 / 0	+1% / +1%	+1% / +1%
misquote episodes kept	0 / 0	0 / 0	0 / 0
no cleaning at all	0 / 0	+2% / +3%	+1% / +1%
haircut flat 1.5% per side	0 / 0	-7% / -10%	-7% / -7%
discovery trigger = calledAt	-12% / -23%	n/a	n/a

Verdict: not fragile. Nothing moves either headline by a third. The discovery calledAt case (-23%, n=40) comes closest and is a small cell. The strict mean, by contrast, is fragile: without cleaning it is 10^19 %, and with the misquote filter off it is +485%. Quote strict medians only. The held measure is robust because the 5-minute hold requires a price to stand on consecutive prints, so a single bad print or a short misquote run cannot create a win. That is a second reason, besides Thomas's "long enough", to put the hold in the public definition.

3.7 Repeat-token effect

Set	First sighting: n / FP / strict win / strict median	Repeat on same token: n / FP / strict win / strict median
Discovery	22 / 36.4 / 40.9 / -19.5	19 SMALL / 21.1 / 31.6 / -17.0
Crossings (nth crossing in this set)	76 / 48.7 / 51.3 / +2.6	72 / 44.4 / 47.2 / -6.7
Crossings (engine's own repeatToken flag)	106 / 57.5 (A) / 52.8 / +4.0	42 / 57.1 (A) / 40.5 / -19.1
Alerts	528 / 29.2 / 38.4 / -21.2	2,809 / 35.3 / 43.6 / -9.4

On crossings, repeats lose under the strict rule: median -19.1% on the engine flag vs +4.0% for first episodes. They do not lose on the forgiving measure. That matches the v1 loss: repeats spike and fail fast. On alerts it is the other way round. The first alert on a token is the worse one (strict median -21.2%), because first alerts include fresh launches that die. Repeat handling should be per source, not global.

4. The recommended public definitions and page wording

Forgiving win (first-passage, with baseline). Parameters: X = 20% net of costs, held 5 minutes, before -30%, within 24 h. Also show the 4 h figure, because that is where the signal lives.

Page sentence: "After this signal, the price rose 20% above the first price you could have bought at, after costs, and stayed there for five minutes before it ever fell 30%, in N of M cases within 24 hours; at a random moment on the same tokens it was K of L."

Strict win.

Page sentence: "Bought one minute after the signal and sold by a fixed rule (out at -45%; after a 40% rise, out on a 20% fall from the high; otherwise out at 72 hours), the trade made money after costs in N of M cases, median result R%."

Rules for showing either number: - Always show n, the date window, and the random-moment rate next to it. - Hide any cell with n under 20. Never show a mean, only medians and shares. - The costs model and the 60 s reaction delay are stated in a footnote under the number. - Neither number may be shown for discovery calls until the capture gap in §5 is fixed; n=41 of 308 is not a sample of the calls.

Why these parameters: +20% is the smallest move that survives realistic costs and the 5-minute hold with a margin. +50% is too rare to be a "forgiving" measure: 15% for alerts, 26% for crossings. A -30% adverse limit is where a normal person gives up; -45% is the follower's stop, and at -45% the measure approaches the volatility-only version. 24 h is the default; 4 h is the honest window, where the lift is largest (crossings +18 pts, alerts +8 pts).

5. Zones: what they can honestly mean on this data

Principle: a zone is a descriptive state computed from the source's own history, with its n, never an instruction. The page shows where the token is now on a ruler built from past triggers of the same source (and band, once cells reach n ≥ 20).

Buy zone: "early window"

• Meaning: "Minutes since the trigger, inside the window where this source's triggers have beaten a random moment."
• Backing: Crossings' first-passage rate falls from 50.0% at a 0 s delay to 46.6% at 60 s, 41.5% at 5 min, 40.1% at 15 min, 42.4% at 30 min and 34.5% at 60 min. The control rate is 38.3%. The crossing window closes at about 15 to 30 minutes. Alerts barely decay: 36.8% at 0 s, 34.1% at 60 min, control 30.5%. An alert marks a regime of activity, not a moment, and its window is about the first 1 to 4 hours, where lift is +8 to +10 pts. Discovery shows +9 to +14 pts at 1 to 4 h, but n=41 is not enough to draw a zone.
• Cannot claim: that buying inside the window wins; that the direction is up (§3.2 says -20% is as likely as +20% in the first hour); any price-level buy zone. Entry price relative to the trigger print was not tested.

Hold zone: "clock and ruler"

• Clock: elapsed time against this source's time-to-best curve. Example wording: "At 1 h, 85% of past alert winners had not yet reached their high; at 8 h, 46% had not." Crossing winners peak sooner: median 4.7 h vs 6.8 h for alerts.
• Ruler: the current drawdown from entry against this source's drawdown-before-best distribution. Example wording: "Half of past winners from this source dipped 12% or more before their high; one in four dipped 26% or more; one in ten dipped 40 to 46%."
• How the hold-stage view should use them: draw both as bands behind the token's own path since the trigger. The clock is the x-axis band (p25 to p75 time-to-best: 2 h to 15 h for alerts). The ruler is the y-axis band below entry (p50 and p90 drawdown-before-best). The token's live position is a dot.
• Required caveat, shown on the page: being inside the ruler does not mean it will recover. Non-winners dip too, and half of alert non-winners had already made their high by the first hour. The ruler is context, not a signal.
• Cannot claim: a probability of recovery for this token; any per-token forecast.

Sell zone: "late window" and "past-best giveback"

• Late window: past the time by which most of this source's winners had peaked. For alerts, 64% of winners had peaked by 12 h; for crossings, 74% by 12 h.
• Giveback ruler: after winners made their 24 h best, the median later low was -47% (alerts) and -58% (crossings) from that best. Wording: "Past winners from this source gave back half from their high within the day."
• Gain ruler: the median best gain of past winners (+63% alerts, +91% crossings). Being above it is shown as "above what the median past winner reached", not as "sell".
• A fading alert does not define a sell zone. Fading alerts are followed by a +20% held move more often than entry alerts: 48.7% vs 40.7% forgiving, 37.1% vs 34.3% first-passage. The fading state is not a directional exit signal on this data and must not be drawn as one.
• Cannot claim: that selling in the zone beats holding; any exit timing. No exit rule tested here (the strict rule) makes money after a trigger.

Where zones are not allowed today

Discovery (n=41). Every cell with n under 20 (bsc crossings, ≥20M crossings, most per-caller cells). Any band-by-source split below n=20.

6. What data we lack for better zones, and what each would add

Gap	Today	What it would add	Cost (estimate, grade C)
Price capture at call time for discovery	Median 70 h between a call being seen and its first tick; 41 of 308 calls measurable	Makes discovery zones possible at all. One price read per call at seenAt, then the same poll as situations for 72 h.	~650 calls over 15 days ≈ 45 calls/day. At a 60 s cadence for 72 h that is ~190K dexscreener reads/day, which is a lot. At 5 min cadence, ~39K/day.
Raw payload sidecar (bead 44ht)	Phantoms and misquotes are classified by heuristic	Confirms whether a jump is the provider or our handling. Turns the misquote filter from inference into fact.	Disk: about the size of the normalised stream, ~1 to 2 GB/month (estimate).
Provider clock on ticks	Ticks carry no source and no providerAt	Honest entry needs the price's own time, not our receive time. A 60 s-late quote on a fast token is the paper-book quote-lag artifact.	Writer change only.
Executable quotes at entry	Entry priced off the indexer print	Replaces the haircut model with a real fill for the ticket size (router getAmountsOut on BSC; Jupiter quote on Solana). Required before any strict number is shown as "what you would have got".	One quote per trigger, ~400/day for alerts.
Holdings on Solana and BSC	46% of holdings files are "chain has no tape" stubs	A sell-zone ruler based on flows, not prices: covered wallets exiting, top-holder share falling. That is the only candidate for a non-price sell state.	BSC tape exists (09-18); Solana tape is a fossil. The Solana tape restart is the real cost.
Solana trade tape / on-chain candles with retention	candles-onchain and candles-bsc roll at ~12 h; trades-solana is one hour on 09-17	True traded prices (highs, lows, volume per minute) instead of indexer prints. Removes the phantom and misquote problem at the root, and allows a volume-confirmed hold.	Disk (the BSC tape is 1 GB in 7 days); RPC credits for Solana.
Longer crossing history	3.7 days, 148 crossings	Band-by-chain cells with n ≥ 20; stable time-to-best curves.	Time only.
Event ids on alert evidence	Alerts carry alert ids only	Would let zones be computed per underlying event (which wallet, which caller) rather than per alert state.	Writer change.

7. What I did not do

• Did not test price-level buy zones (entry distance from the trigger print or from the trailing high). Only time-based windows were tested.
• Did not use executable quotes. The haircut is a model (1% / 0.5% per side, and a flat 1.5% sensitivity), not a measured fill.
• Did not use candles or the trade tape. Retention is too short (§1). The deep-paper path and ticks are both dexscreener, so there is no independent cross-check of any single price.
• The controls are one construction (every 4 h from first print + 1 h). They are not matched on time of day or on the token's own recent activity, and they can fall near a real trigger. A stricter control, such as random times drawn at matched volatility, would probably shrink the 1 h lift further.
• Alerts are not independent. 3,337 triggers sit on 591 tokens, with repeats in the same regime. The "first entered per token" rows in T6-strict-vs-forgiving.csv give the one-per-token view: forgiving 20/24 is 38.7% for alerts (vs 40.7% all) and 61.0% for crossings (vs 57.4% all). No clustered confidence intervals were computed. With 35 to 77 tokens behind discovery and crossings, treat differences under ~10 pts as noise (my inference, grade C).
• Did not model the follower's actual clip ladder (50% at +40, 25% at +200, breakeven after clip, 40% relative trail). The strict rule is a simplified single-exit version of its shape.
• Did not split by caller beyond the 2 callers with any n, or by hour of day.
• Did not touch any live store or writer. All reads are from copies in data/ (T6-book.sqlite via sqlite .backup, T6-ticks/, T6-discovery.jsonl, T6-alerts.jsonl, T6-deep-paper-status.json, ~480 MB).
• Did not read events.sqlite or price-paths.jsonl. price-paths is derived from the same ticks and candles.

Files: scripts/T6-zones.py, scripts/T6-symmetry.py; data/T6-triggers.csv, T6-grid.csv, T6-lift.csv, T6-firstpassage.csv, T6-firstpassage-dims.csv, T6-strict-vs-forgiving.csv, T6-timing.csv, T6-holdzone.csv, T6-sensitivity.csv, T6-symmetry.csv, T6-summary.json; logs in logs/T6-run*.log.

Status: DONE