明鑑  SHEFA vALPHAPENGUUSDT PERPETUAL · THE BRIGHT MIRROR

--connecting

起動 Ignition · next 60 min

--≥1.5% either way
vs baseline this hour--
percentile of all minutes--
P(≥2.0% within 120m)--
P(≥3.0% within 240m)--
PENGU base rate 40.5%. The σ-decile lookup scores AUC 0.792; PEPE's 126-feature gradient-boosted model scores 0.767 on its own tape. Read the percentile, never the raw percentage — the raw number moves with volatility, the percentile does not. Bottom decile hours are when to be away from the screen.

行程 Expected runway · next 60 min

--ticks · median excursion
q25 → q75 band--
4-hour median--
8-hour median--
current 1σ hour--
expansion vs σ--
Spearman 0.605 / 0.515 / 0.466 at 60m / 4h / 8h. A median is not a target — half of all outcomes fall below it. The q25→q75 band is the middle half; treat anything outside it as unusual rather than impossible.

機況 Engine state

status--
size for constant risk--
my runway (8h q50)--
my good case (q75)--
range pos (24h)--
fade tail lean--
MY SURVIVE-TO--
survive to (4h, either side)--
survive to (8h, either side)--
hold8h
expansion-surprise 4m--
burst alert (4m)--
expansion-surprise 1h--
expansion-surprise 4h--
expansion-surprise 8h--
path quality (payoff:heat)not predictable
compression rank (60m)--
minutes since last burst--
loudness vs prior σ--
Surprise divides σ out and asks whether the move beats what volatility already explains — the only genuinely anticipatory channel here. PENGU AUC: 4m 0.546, burst-4m 0.581, 1h–8h 0.58–0.69. Path quality permanently reads "not predictable" — on PEPE it scored AUC 0.5503 against a 0.5654 volatility baseline, worse than the trivial benchmark. The row is kept as a gravestone so the question is not asked twice.

餘勢 What is left on the current leg

LEG--
TRAVELLED--
pullback from leg extreme--
leg age--
residual q25 → q75--
median residual--
PENGU differs from PEPE here. On PEPE remaining extension is memoryless — flat in every travelled×age cell. On PENGU the raw spread is 135.8%, but that is mostly volatility: σ-normalised it is 41.8%. So it is modestly conditional — neither flat nor usefully predictable. Read the band as a distribution, never as a forecast.

階梯 Target ladder · odds only

movehorizonlive oddsbase ratelift
Seven rungs, live σ-conditional odds against PENGU's own unconditional base rate. Rungs are a percentage of price, never pip counts — PENGU's tick is 1e-6, ten times PEPE's, so a pip ladder carried across would be wrong by an order of magnitude. Probabilities are for a move of that size in either direction. No side is implied.

逆風 Adversity budget · trade through

--ticks · p90 heat, 4h
median heat before payoff--
survive-to price (long)--
survive-to price (short)--
payoff / heat at median--
my payoff : heat--
8-hour heat p90--
survive-to 8h (long / short)--
size if holding 8h--
Heat is the excursion against you before the move resolves; p90 means nine times in ten it will not get worse than this. If you cannot sit through survive-to, you are too big. That is the most directly actionable line on the page. Payoff:heat is descriptive only — it is not an edge, because path quality is not predictable.

時辰 Time ahead · next 24 hours

MYTactivity
Median 60-minute range by hour of day, relative to the day's median. PENGU is busiest at 23:00, 22:00 and 00:00 MYT, deadest at 19:00, 18:00 and 12:00. Hour-of-day structure is the most stable thing in this market.

遠見 Reach · how far ahead it is real

aheadrhotop decile still hits
How well σ now predicts the 60-minute excursion beginning later. Decays 0.97 → 0.89 then rises again at 24 hours — the daily clock coming back around, the same signature PEPE shows. Corrected during build: the first version shifted σ instead of the outcome and printed rho 0.9987. That was lookahead, not skill.

價跡 Price · ignition history · last 240 min

Last four hours of price. Red marks are minutes where loudness reached 1.5 or above — a genuine burst for this market. Use it to see whether the current reading sits inside a run or is an isolated spike.

戒 What vAlpha will not do

It will not tell you which way. On PEPE direction was closed twenty-eight ways, then reopened on a trade-based clock — genuinely predictable at 58% on 5% of bars, t≈10, 19 of 19 folds — and still worthless: right 54.6% for +59bp, wrong 45.4% for −70bp, netting +0.6bp. Present, and exactly priced away.

No PEPE model runs here. The 31 PEPE heads (5,800 trees, 11.1MB) were not ported: their rungs are pip-denominated, PENGU's tick is ten times larger and its σ-hour differs, so both labels and targets move. Every head is PENGU's own σ-conditional table — and it matches or beats the PEPE model on its own tape (0.79/0.78/0.77 against 0.767/0.743/0.720).

No profit, loss or position size in money appears on this page. The cost gate lives offline.
連線  Connection · feed routing
venue--
bars loaded--
last bar--
refresh60 s
Primary is the USD-M perpetual endpoint. If it is unreachable the page falls back to the public spot mirror and says so — spot and perp differ by basis, so treat the figures as close, not exact, when the mirror is in use.
飛行記錄  Flight recorder · 0 entries
JSON carries full precision and the engine constants — that is the one to send me. TXT is a fixed-width table readable as-is. CSV is the raw bars.
實時讀數  State readout · live feature values
校準  Calibration · measured on PENGU
餘勢何以不變  Why "what is left" barely moves · tested
Remaining extension was tested against how far the leg had already travelled and how old it was, in a 4×4 grid. On PEPE every cell returned the same answer — a conditional lookup scored −0.6% pinball skill, worse than quoting the flat distribution, so the panel was removed and the unconditional band shipped instead. On PENGU the raw spread looked large (135.8%) until σ was divided out, leaving 41.8%. Modest, and not enough to forecast on. The band you see is a distribution, deliberately.
方法與界限  Method and limits · read once
Archive: 896,145 one-minute bars, 2024-12-17 to 2026-08-31, zero gaps.
Heads: σ-decile conditional tables, not machine learning. Ignition AUC 0.74–0.80, surprise 0.55–0.69, runway spearman 0.47–0.61, heat spearman 0.35.
Sampling: every table uses non-overlapping windows at its own horizon.
Three bugs caught in build, all of which would have shipped silently: an AUC function that ranked in the wrong order and printed 0.35 — below chance; a reach curve that shifted σ instead of the outcome and printed rho 0.9987 — lookahead; and a memoryless test that was not σ-controlled.
Not done: no direction head, no order-flow features, no event-clock rebuild. Direction on PEPE proved real and worth +0.6bp, so it was not worth repeating here.