# Refined Plan — Full Market Scan + Execution-History Audit
Sat Oct 24, 2026 ~2:35 AM PT · Prepared for Param · Supersedes all earlier informal weights

## 1. Execution history — the true record
- Live orders ever placed by Muse: **ZERO.** Two auto-trade attempts (Oct 20: $10 one-off; $30 BTC ≥85%-at-3:00 rule) both died to task latency — no order ever reached Kalshi. Account last verified: $39.37 cash, $0 positions (Oct 20).
- Paper bot (Phase 0, Crypto Leader WILL NOT WIN, started Oct 24 ~2:08 AM): **0 trades, 2 decisions — both SKIP, both correct** (XRP No 99¢ and SOL No 100¢, both above the 95¢ hard cap). Price discipline is working; nothing has been lost to a fill.
- All other "losses" in the logs are graded CALL losses (no money attached): GBP from-5 feed 42/54 RIGHT (77.8%) at stop; GBP 6:00 feed 49/67 at last audit; Treasury feed mixed with whole-curve tiny events. Post-result audit log: 40 WRONG entries — **the majority TINY** (FX finals within $0.00011 of strike; Treasury finals within 0.2bp), i.e. coin-flip finishes, not model misses.

## 2. Mistakes and false signals found in the logs (each with its fix)
1. **Tiny-lead trap (the #1 real error).** Repeated losses where the printed % was 78–83% but the lead at the mark was under $0.00005 (T bucket): GBP 1:00 PM UP 83% → WRONG real (lead +$0.00002); 1:15 PM UP 81% → WRONG real (lead +$0.00005); 12:15 PM UP 79% → WRONG tiny (lead +$0.00005). History says a sub-$0.00005 lead holds only ~55.8% — the pattern label was overriding the lead size. **Fix: lead size is now a HARD GATE applied before any pattern % (see §4).**
2. **Book-wall veto cost wins, saved none.** In the GBP audit tail alone, 4 mandate aborts on book walls were graded FALSE vetoes (11:30 AM, 12:00 PM, 1:15 PM, 1:30 PM — including a 20.7× wall on a call that held DOWN by −$0.00048 real). No audited window shows the wall veto saving a loss. **Fix: wall weight → 0 as a veto. Info only, until it is graded on the full book log and earns its way back.**
3. **TradingView gauge is noise.** GBP audit tallies at the end: calls held 5/6 when TV AGREED and 6/8 when TV was AGAINST — no discrimination in either direction. **Fix: TV gauge weight → 0 everywhere. It stays logged (data collection is free) but prices nothing.**
4. **Counting one event as three losses (Treasuries).** Tenors move together 96.7%; whole-curve flips were logged as 3 independent WRONGs (e.g. 1:15 PM Oct 23: 2Y tiny + 5Y/10Y real, one event) and curve splits were misread as tenor signals. **Fix: the curve is ONE position. One vote, one size, one grade.**
5. **Friday-UP GBP weakness ignored.** Deep study: Friday GBP UP holds only 47.5%. Multiple Friday UP calls were printed at full % anyway (several lost tiny/real). **Fix: Friday UP in FX is discounted to its measured bucket or passed — never printed at the generic %.**
6. **Settlement-basis slippage (Crypto Leader).** Plain Coinbase open-to-open matched Kalshi's official winner 80.5–83.7%; the CF-Benchmarks-aligned TWAP-proxy method matched 89.9%. **Fix: TWAP-proxy method only; plain-open reads are retired.**
7. **Latency killed the only live attempts.** Long single sleeps overshot checkpoints by up to ~24 minutes (Oct 20). **Fix (already the standing rule, now a gate): start near the mark, re-sync to a live countdown, hard abort at +150s, stale = suppressed, never delivered.**
8. **Data-gap false reads.** Coinbase serves no candle for the mark minute itself and none for HYPE no-trade minutes — two feed failures (Oct 24 1:06 AM, 1:36 AM) before fixes. **Fix: mark price = close of the candle ending at the mark; forward-fill from last traded close; NO READ line instead of a guessed ranking.**
9. **Known open bug — DOWN-side price formula.** P = 100 − no_best returns the YES price, inverting DOWN soft/strong labels (found Oct 20, still unfixed). **Fix: DOWN soft/strong labels are not used for any decision until the fix is approved and tested.**
10. **Dead signals that must stay dead (holdout-killed):** standalone momentum, time-of-day/afternoon discount, pre-window prediction for Crypto Leader (momentum 21–24%, reversal 22.8%, repeat-winner 20.2% — all at the 20% random line), and "follow the market on disagreement" for BTC/ETH/Copper (disagreements graded 50.5–54.7% = coin flips). **Fix: weight 0, permanently, unless a fresh chronological holdout revives them — none has.**

## 3. What the record says actually works (kept, and weighted up)
- **Kalshi's own % at the checkpoint** — best single probability in every book (e.g. BTC lead-side ≥90¢ at 6:00 held 95.6–96.0% across BTC/ETH/Copper).
- **Lead size buckets** — the dominant sorter inside every asset (BTC $50–100 × % grid; FX lead buckets; All-8 big-lead cutoffs at 5:00, CLEAR hold 95.6–100%).
- **Full-record coin×place tables (Crypto Leader)** — 5,844 official windows: at 5:00 left, 1st holds 66.9% (HYPE 74.9% → ETH 59.7%), 5th wins only 1.7%; loser stacks with gap tiers reach 98.1–99.5%.
- **CAD gatekeeper for GBP** — confirm branch 85.6% at 4:00 vs 75.9% against (audit tally on the confirm side 7/8 at last log). EUR+AUD both against GBP = NO TRADE (69.0% / real 63.5%).
- **ETH filter** — |move| ≥$7 + lead-side 90–95¢ at 6:00: 98.44% in the practical tier (holdout-verified).
- **Price caps with fee netting** — paper bot's 2/2 correct SKIPs came from this alone: never pay above record% − 3¢, hard cap 95¢, taker fee 0.07·P·(1−P) netted before any EV claim, live depth checked before a fill.

## 4. Refined strategy weights (v2 — tighter than last iteration)
Signal weights for pricing a call (was → now):
- Kalshi own % at mark: primary → **primary (unchanged, it earned it)**
- Lead/move size bucket: co-primary → **HARD GATE first** (below the asset's tiny threshold = NO TRADE, whatever the % says — kills mistake #1)
- Confirmation (FX: EUR/AUD/CAD/JPY pattern; Treasuries: curve-as-one; Crypto Leader: coin×place table): full weight → **full weight, and EUR+AUD-both-against / curve-split = NO TRADE**
- Book wall / orderbook imbalance: veto → **0 (info only)** — kills mistake #2
- TradingView gauge: confirmatory → **0** — kills mistake #3
- Momentum (standalone): already 0 → **0 (locked)**
- Pre-window / time-of-day: already 0 → **0 (locked)**
- Session flags: none → **NEW: Friday-UP FX discount to measured 47.5% bucket** — kills mistake #5

Entry rule (all markets, no exceptions):
1. Live fetch at the mark, re-synced; >150s late = no call, no trade.
2. Lead ≥ the asset's real threshold (FX ≥$0.00010 at 6:00 / ≥$0.00005 absolute floor; BTC grid cell n adequate; Crypto Leader: pick priced from the 5,844-window table only).
3. Confirmation not against (FX full pattern; Crypto Leader coin×place; Treasuries one curve vote).
4. Price: ask ≤ measured record % − 3¢, hard cap 95¢, fee netted, live depth ≥ size. Otherwise SKIP — a skip is a saved loss, per the paper ledger.
5. Size: Phase 0 paper only ($10 nominal, max 20/day, −$30/day stop) until 50–100 graded paper trades; live only on Param's exact asset/side/size/limit command.
6. Every result graded in the same run (tiny vs real); weights change only on ≥25 new graded windows or 3+ REAL losses on the same pattern — never off one window.

## 5. Market scan — Sat Oct 24 ~2:35 AM PT
- **Crypto (BTC/ETH/SOL/XRP/HYPE 15m + Crypto Leader): LIVE 24/7** — the only markets open now. Feeds running: Crypto Leader places/BUY/predictions in their chats; Phase 0 paper bot; **new in this chat: WINNER at 5:00 left + RESULT every window (started tonight, first pick 2:39 AM run for the 2:40 mark).**
- **FX (EUR/GBP/AUD/CAD/JPY): CLOSED** — Friday close was 2:00 PM PT; no markets until the Sunday/Monday reopen. All FX feeds stopped by Param Oct 23 night; restart only on his word in each chat.
- **Gold / Copper: CLOSED for the weekend** (CME); copper also has its daily 2:15–3:15 PM PT halt on trading days.
- **Treasuries: CLOSED** weekend; feeds stopped Oct 23 night.
- Net: tonight the entire book is Crypto Leader + crypto 15m. That is where the refined weights are live-tested first (paper), which matches the Phase 0 plan.

## 6. What changed vs the last iteration — demonstrably
- 3 signals zeroed on graded evidence (wall veto, TV gauge, plus locked dead signals), 1 hard gate added (lead size before %), 1 session discount added (Friday UP), 1 method retired (plain-open crypto reads → TWAP proxy, +6 to +9 pts winner agreement), 1 risk unit corrected (Treasury curve = 1 position).
- Every change traces to a named log entry or holdout above — nothing is tuned on a single window.

## 7. Addendum — independent recount + Phase 0 tightening (chat ff43ba66, Oct 24 ~2:36 AM)
Recount of the raw logs, done independently of §1–§6 above:
- Post-result audit log (post_result_audit_log.md, Oct 23, treas6 + gbp6 result jobs): **83 graded entries — 49 RIGHT / 34 WRONG. Of the 34 WRONG: 24 TINY (71%), only 10 REAL.** The dominant "loss" in Muse's history is the coin-flip finish at the strike, which is why the lead-size hard gate (§4) is the single highest-value change in this plan.
- Book-wall veto, FULL live record across all GBP sessions through Oct 23 1:30 PM (stronger than the §2 tail count): **11 vetoes — 10 cost would-be wins, 1 saved a loss.** Veto weight 0 is not a close call.
- Auto-trade latency, exact count: **3 qualifying BTC signals missed on Oct 20 alone** (3:00, 3:45, 4:15 PM windows, incl. a DOWN 98% call) — zero orders reached Kalshi all day. Execution path stays API/demo only; browser execution weight stays 0.
- Phase 0 paper bot, first 2 windows (2:15–2:30 AM): both marks SKIPped correctly on price (XRP No 99¢ at 5L, SOL No 100¢ at 3L — caps did their job, $0 lost). But the pattern is the lesson: by 5:00–3:00 left in thin overnight hours the loser is usually already priced 99–100¢, so the bot as built will rarely fill. **Phase 0 tightening (proposed, awaiting Param's go in chat ff43ba66): add the 6:00-left mark as the FIRST entry attempt — loser records at 6L are BTC 99.7 / ETH 99.4 / SOL 98.3 / XRP 97.8 (5,841 windows) with the same caps and depth check — then 5L and 3L remain as re-checks. Still ONE paper trade per window: first qualifying mark wins.** No cap is loosened; the cascade only moves the first attempt earlier, before the price runs away.
- False-signal ledger for Phase 0 itself so far: **0 false trades, 0 losses, 0 missed fills that should have traded** — both SKIPs were the system working as designed.
- Live scan snapshot from this chat (Sat Oct 24, 2:30–2:45 AM window, ~9 min left, Kalshi public API): BTC OPEN (strike $80,474.74, Yes bid/ask 32/33¢, vol 662k) and ETH OPEN (strike $2,359.68, Yes 19/20¢) — both leaning DOWN at scan time; Crypto Leader live Yes asks HYPE 45¢ / BTC 36¢ / XRP 33¢ / SOL 8¢ / ETH 6¢; Gold, Copper, EUR/USD, GBP/USD, AUD/USD, USD/CAD, USD/JPY and US 2Y/5Y/10Y all CLOSED / no market (weekend). Confirms §5: tonight's entire tradeable book is crypto.

## 7. Addendum (chat ff43ba66 recount, Oct 24 ~2:34 AM)
Independent recount of the post-result audit log: 83 graded — 49 RIGHT / 34 WRONG, of which 24 TINY and only 10 REAL. Wall veto full record: 11 vetoes — 10 cost wins, 1 saved → weight 0 confirmed. Live scan from that chat: BTC + ETH + Crypto Leader OPEN only; Gold/Copper/all FX/all Treasuries CLOSED (weekend). Pending Param's go there: Phase 0 cascade entry at 6:00 left.

## 8. Addendum (chat 5c614b36, Oct 24 ~2:40 AM) — Crypto Leader SPREAD GATE, wired
The coin×place table alone mispriced tonight's 2:00 AM window: HYPE printed 60% as 1st at 10:00 left and lost — it led at −0.066% with all five negative and the field packed inside a 0.14% spread. Measured on all 5,844 official windows (holdout-stable): leader win rate by 1st-to-5th spread at the mark — @10L: <0.05% → 24.2% / 0.05–0.10 → 38.2% / 0.10–0.20 → 46.6% / 0.20–0.35 → 54.3% / ≥0.35 → 66.0%; @5L: 25.5 / 45.8 / 60.6 / 70.1 / 78.4%. Gap 1st–2nd @5L is the same physics (<0.02% → 46.1% vs ≥0.10% → 86.5%).
- WIRED into the prediction feed (crypto5_predict.py): every 10:00/5:00 block now prints the spread + the measured leader rate for that spread bucket; <0.10% flagged "THIN SPREAD: no-edge window", ≥0.35% flagged "WIDE SPREAD: the strong zone". Hist-verified on the 2:00 AM window (spread 0.137% → 47% @10L; 0.177% → 61% @5L — the gate would have discounted the losing 60% print).
- Acting weights: winner pick full weight only at spread ≥0.20% (58% @10L / 74% @5L vs 49/67% ungated), half at 0.10–0.20%, no winner-side trade <0.10%. Gating to ≥0.10% lifts 5:00-left leader accuracy 66.9% → 69.3% on 91% of windows.
- Loser side reinforced, not changed: 5th-place win rate falls as spread widens (@5L 14.5% → 0.5%), consistent with the existing gap tiers; HYPE-in-5th stays quarantined at its own 95.9% record.