THE ESTIMATOR · KXBTC15M · estimate, don't echo

connecting to brain…

OFFICIAL v1.0 TRACK 7¢ · FROZEN · forward test from 2026-09-04 17:03 UTC

Graded

—

Modeled-fill

—

Next checkpoints

163 sanity · 500 audit · 1,000 persistence · 2,000 durability

Wins: — / — · Success rate: — — descriptive outcome frequency, not proof of edge (hit rate alone ≠ +EV)

Evidence-collection phase — the model is frozen; this tracks progress, not score. Panel 6 is the historical / all-time record (v0 10¢ + v1.0 blended), not this forward track.

V1.0 FORM · recent performance · monitoring only

MONITORING ONLY — recent form of the frozen v1.0 track. Not an audit, not a tuning input. No window here may trigger or justify a model / rule / threshold / calibration / hypothesis / version change. First serious performance audit remains the 500-signal milestone. Reads the served immutable graded log client-side; the daemon, schema, and grading are untouched. loading…

Last 20 — form (oldest → newest)

loading…

← 20 agomost recent →

—

Rolling windows (newest-anchored — these shift as new signals grade)

Window	N	W–L	Success	EV/sig (modeled)
loading…

Historical 20-signal blocks (start-anchored — strictly chronological, non-overlapping, permanent identity)

Block	Signals	Dates (UTC)	W–L	Succ	EV/sig	form
loading…

Two anchors: rolling windows anchor at the newest signal (they re-slice as signals grade — the "recent form" read); historical blocks anchor at the first signal, so B1 is always signals 1–20 and never changes (the "drift over time" read). That is why Last 20 ≠ the final block unless the total is an exact multiple of 20.
Definitions (all from immutable stored fields): Success = wins ÷ N (stored win, graded against Kalshi's official result — never recomputed). EV/sig (modeled) = mean stored pnlCentsSlip = the modeled / pessimistic-fill (+2¢) assumption, ¢ per signal — an execution model, not verified fills. Only graded v1.0 rows (loggedAt ≥ cutover) count; pending + archived v0 (10¢) excluded.

1 · Window

contract—

target strike—

resolvesspot ≥ strike at close

closes in—

2 · Fair estimate (independent · it dances, not a score)

—

σ (1-min, realized)—

spot vs strike—

flow nudge—

2b · Spot vs strike (last ~20 min — the distance you're betting on)

– – strike — spot now —

3 · ⭐ Edge — fair vs implied

our fair (YES)—

market YES (bid/ask)—

gap (edge)—

edge magnitudeamber line = threshold

—

4 · Expected value (the deciding number)

—

gross edge—

− est. Kalshi fee—

= net EV / contract—

5 · Order flow / context (nudge, not driver)

book imbalance (±0.15%)—

bidask

tape (aggressor buys)—

buysell

spot (Coinbase)—

read—

6 · Track record · BTC — HISTORICAL / ALL-TIME (v0 10¢ + v1.0 blended) · graded at real close · study EV, not hit-rate

signals logged0

graded0

realized hit-rate (vanity)—

EV / signal (quote fill)—

★ Modeled/pessimistic-fill EV (+2¢)—

paper P&L (net)—

★ same-window: est vs favorite (clean test)—

vs "always favorite" aggregate—

vs "always underdog" aggregate—

window	side	fair	dist σ	entry	result	¢P&L
no signals yet — WAIT is the default

7 · Discipline

PAPER MODE threshold 7¢ stake 10 daily cap none one brain · cloud

🧠 learner (beginner's mind · calibration)—

Never auto-trades. Computed once, server-side, by the always-on cloud brain — this page is a pure window into it. BTC and ETH are scanned by the same engine; the toggle above swaps which one you're viewing. You place any trade yourself, on your own keys.

8 · Threshold test · BTC — observe-only: does a lower bar earn its keep?

bar	candidate signals	hit	EV / sig	EV (pessimistic fill)
gathering — no graded candidates yet

Candidate signals the engine WOULD take at each bar (first-touch, no confirm), graded at close for this coin. The live bar is now 7¢ — the shadow test (n≈196) showed it nets more after pessimistic fills than 10¢. This panel keeps watching all three so the choice stays honest: if 7¢ ever starts bleeding relative to 10¢, the data will say so (panel 6 is the real record).

9 · Calibration · BTC — when it says X%, does it happen? (all windows, not just bets)

accuracy log—

predicts UP (≥50% fair)—

predicts DOWN (<50% fair)—

it says (confidence)	n	actually happens	read
gathering — needs graded windows

Every window's raw prediction is graded here (not just the bets), so your track record stays untouched. Well-calibrated = "actually happens" lands inside or above each band; below the band means over-confident — the risky kind.

10 · Pattern scan · BTC — observe-only: where does the edge concentrate? (buckets are small — hypotheses, not signals)

condition	n	hit	EV / sig
gathering

Slices the graded record by condition to surface possible patterns — it bets on none of them. Small buckets, so treat any gap as a hypothesis to watch, not a signal to act on. A real pattern must hold across a big, multi-regime sample before it earns a place in the engine.

11 · Hypothesis board · two minds · —

pattern	found (in-sample)	split	forward (out-of-sample)	verdict
gathering — the creative mind proposes once a bucket clears its gate

Creative mind proposes a pattern the moment a bucket beats (or trails) the coin's own average by ≥5¢. Logical mind refuses to rule until 15 fresh signals arrive AFTER the proposal — true out-of-sample — because a pattern that only shows up in the data it was found in is usually noise. CONFIRMED = held on forward data (a candidate to graduate into the engine, only with your go-ahead). WATCHING = still gathering forward proof. REJECTED = didn't hold. Most will read WATCHING for a while — that patience is the discipline. Bets on nothing.

Engine: fair_up = Φ( ln(spot/strike) / (σ·√t) ), σ = realized 1-min vol, t = minutes to close. Flow is a ±2¢ capped nudge, never the anchor. The market price is TESTED, never an input. σ and flow are EMA-smoothed and a verdict must hold ~3 cycles before it shows; no signal is issued in the final 90s or when spot is within 0.7σ of the strike (a coin flip). Abstain by default — most cycles are WAIT, and that is the point. Paper-only; you place trades on your own keys. · This page reads the cloud brain's live state; it computes nothing itself.