> ## Documentation Index
> Fetch the complete documentation index at: https://shumi.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Decision record: rules

> Version 2 of the rules behind the public decision record: the coin pick, the daily question, the scoring of each plan and the pass bar.

## Summary

Each day the record picks the coin Shumi's hourly sentiment summaries mentioned most over the past 24 hours, from the top 500 by market cap. At 13:15 UTC Shumi is asked for a long and a short plan (entry zone, stop, target, score out of 10) and which side the data favours, if either. The entry goes live by 19:15 UTC, and the favoured plan is scored from then.

A plan fills when price reaches its entry zone. It stays open until a daily close beyond its stop, the trend turning against it, or a reversal signal once it is 3 % in profit.

The record is read at 30 finished plans from at least 8 weeks, and again at 60 from at least 12. It passes only if, after costs, the average plan and the average week made money, Shumi's chosen side beat taking the long plan every day, at least 40 % of plans ended in profit, and no scheduled day was missed. Every plan and result is in [the decision record](https://shumi.mintlify.site/evidence/decision-record.md).

## Version

Version 2, registered 2026-10-29. Day one is 2026-10-31, and every count starts there.

Amended 2026-10-30, before day one and before any version 2 entry: section 5 "Prices" adds Hyperliquid spot for a coin with under 150 days of Binance history.

[Version 1](https://shumi.mintlify.site/evidence/pre-registration-v1.md), registered 2026-10-14, missed scheduled days, so it can no longer pass. Its entries, for days before 2026-10-31, stay published and unedited and are not scored. Its picks count for the 7-day rule. A version 1 plan still open keeps its coin out of selection until it closes, if its entry went live by 00:00 UTC the next day.

A rule change is a new version with its own page and date. Each entry is scored under the version it was published under. No rule changes after the first outcome has been looked at, and the bar is never changed against outcomes already seen.

## 1. Selection

At 12:15 UTC a query picks the coin:

1. Take the eligible coins (section 2).
2. Count each coin's hourly sentiment summaries, matched by upper-case symbol, in the 24 hours to 12:15, keeping the 50 highest counts; any other coin counts zero. Daily, weekly and monthly summaries do not count.
3. Remove coins picked on the previous 7 scheduled days and coins with a plan still open.
4. Pick the highest count. Ties, an all-zero day included, go to the better market-cap rank.
5. Two eligible coins sharing a symbol: the better-ranked one gets the count, the other zero.
6. No eligible coin: the day publishes a `no eligible coin` entry, an attempt with no take.

Each entry lists the five highest counts and whether a tie-break or an open-plan skip decided the pick.

## 2. Universe

A coin is eligible at 12:15 UTC when it:

* ranks 1 to 500 by market cap, as a crypto asset (stocks and other real-world-asset rows are out);
* has no category ending in "stablecoin", other than "Stablecoin Issuer";
* has no category matching `wrapped`, `bridged` or `liquid staking`, in any case;
* has a current site trend computed from complete data.

## 3. The question

At 13:15 UTC a scheduled job sends one message in a new shumi.ai chat, once, with no follow-up and no regenerating. The message is "XRP (XRP): which plan does the data favour today: long, short or neither?", with the day's coin in place of XRP, then layout instructions. The answer must give, in order:

* `Data favours: LONG`, `Data favours: SHORT` or `Data favours: NEITHER` as its first line;
* one line per side, `$XRP | long | entry zone | stop | target | score/10`, from Shumi's price levels, n/a in an empty slot;
* this exit line, exactly: `Exit: the plan ends on a daily close beyond its stop, when the trend turns against it, or on a reversal signal once it is 3% in profit. The target is not an exit.`
* the biggest risk, the venue Shumi would act at, and the source time of every input. An input past its age limit is marked stale and left out.

The stance is the first line matched against `^Data favours: (LONG|SHORT|NEITHER)$`, otherwise `unparsed`. NEITHER is a full answer.

Plan lines are parsed with `^\$(\S+) \| (long|short) \| ([0-9.]+)-([0-9.]+) \| ([0-9.]+|n/a) \| ([0-9.]+|n/a) \| (\d+|n/a)/10$` into entry low, entry high, stop, target and data score. The line on the stance's side is scored, with its numbers as published. With that line missing, unparsed or without a stop, the entry is labelled `no plan`. A missing or reworded exit line is labelled `exit line missing` and still scored.

The model and Shumi's instructions may change during the record; each entry names the model. The X/10 figure is called the data score. A request that fails with no answer is retried once after 10 minutes. Both attempts are published, and the retry is the take of record.

## 4. Times and input ages

Every day, weekends included, on UTC. Each entry states its publication time `T`, and scoring counts from `T`. An entry live after 19:15 is published late with its real `T`. One not live by 00:00 the next day is a missed day, and so is a missed prompt, whether or not anyone was at a keyboard.

Each input shows its own source time, apart from "asked at" (when the question was sent). Limits at "asked at":

| Input | Source time | Maximum age |
| - | - | - |
| Trend | the 00:00 close it was computed from | as of the selection day, from complete data |
| Social sentiment | stored read: the hour it covers; live read of the last 6 hours of X: the newest post | 6 hours |
| On-chain KOL traders | fetch time | 60 minutes (10 if the coin missed its 12:20 pre-fetch registration) |
| On-chain holders, wallet tags, largest-pool liquidity | fetch time | 60 minutes (6 hours if the coin missed its 12:20 pre-fetch registration) |
| Price levels | the open of the newest daily bar | the previous day's 00:00 close, so they read the previous day's date |

On-chain data covers Ethereum, BNB Chain, Base, Robinhood Chain and Solana; a coin with no contract there has no on-chain field. On-chain data is never called sentiment. An input past its limit labels the entry `stale input` with its actual age, and a missing sentiment read labels it `no current sentiment read`. The entry is still scored and also reported as its own group.

The venue Shumi names is published verbatim and does not change the score. Pool liquidity is never shown as depth near spot, checked on every entry.

## 5. Scoring a plan

**Prices.** Public 4h spot bars from one venue per coin, kept for the whole record and never mixed. The venue is Binance `<SYMBOL>USDT` when its 4h history starts at least 150 days before the coin's first entry in the record. Otherwise it is whichever of Binance `<SYMBOL>USDT` and Hyperliquid `<SYMBOL>/USDC` starts earlier. Prices are in the pair's quote coin, USDT or USDC. The bars make one daily bar per UTC day: first open, highest high, lowest low, and the close of the 4h bar ending at 00:00 UTC. A pair whose last close is more than 1.2 times, or less than 1/1.2 of, the coin's CoinGecko USD price is another asset on the same ticker and does not count. Day D's close is used only from 00:00 on D+1. On the publication day, only 4h bars opening at or after `T` count. Each response is stored raw with its request URL, and bars are only requested 4h wide. Each open day's bar and trend are recorded with the plan. A coin with neither pair, or a needed day with no price, makes the outcome `incomplete`, published and not scored.

**Trend.** The site's trend: SuperTrend 5 × 1.5 on the coin in USD, BTC and ETH. All three up is UP, all three down is DOWN, anything else HODL. It is read for the previous completed day, and the record never shows a trend the site contradicts.

**Entry.** 4h bars after `T` are checked in order. A LONG fills on a bar whose low reaches the entry high, at the entry high or at a lower open. A SHORT fills on a bar whose high reaches the entry low, at the entry low or at a higher open. Before a fill, the plan lapses on the first of:

* `invalidated before entry`: a daily close beyond the stop;
* `ran without entry`: price reaches the target;
* `trend turned`: the trend turns against the side (DOWN for a LONG, UP for a SHORT).

With no fill and no lapse after 30 daily closes the plan is `not triggered`, which closes the line.

An unfilled plan shows the move from `T` to the lapse and how near price came to the zone, in ATR.

**Exit.** A filled plan ends at the first of these, at that day's close:

1. **Stop:** a daily close beyond the stop printed at `T`, as Shumi's price levels print it: the nearest swing\_5 low below price within 120 daily candles minus 0.5 ATR (shorts: the swing high above price plus 0.5 ATR). It sits at least 1.5 ATR from the entry midpoint, with no maximum. A swing closed through and reclaimed still counts. The stop never moves, and a gap through it can lose more than 1R.
2. **Trend change:** the trend turns to the opposite side, at any result. HODL does not end the plan.
3. **Reversal signal:** an opposite mean-reversion point under CoinRotator's exit rule (including its last-14-band-re-entries behaviour), with the plan at least 3 % in profit at that close.

No time limit applies. A plan open at a reading is marked at that close and labelled `open`. Reaching the target is published as `reached target` and the plan stays open. Continuation re-entries are not scored. Exits are labelled `exit (stop)`, `exit (trend change)` or `exit (reversal signal)`.

**Result.** Size is the plan's size from Shumi's price levels for a fill in the zone: 1 % of the account ÷ stop distance, long or short. Above 100 % of the account the entry is labelled `needs_leverage`. Risk = |zone midpoint − stop| ÷ zone midpoint. The move g runs from fill to exit, signed for the side. Net R = (g − c) ÷ risk, with c from section 7, and account result = net R × 1 %.

NEITHER, `unparsed`, `no plan`, lapses, `not triggered`, `incomplete`, `no short venue`, errors and `no eligible coin` count as attempts and are not scored. An outcome is a new line appended below its entry, naming the entry id.

## 6. Shadow lines and baselines

Scored on the same fill and published, outside the bar:

| Line | Rule |
| - | - |
| Reversal signal at any result | an opposite mean-reversion point ends the plan |
| Stop beyond the reversal bar | the stop sits beyond the latest same-side reversal bar |
| 2-bar reversal | once 3 % in profit, a second consecutive daily close against the plan |
| Faster average | once the trend on the plan's side is 14 days old, a close beyond the 20-day EMA |
| Target | exit at the printed target |
| Slower trend | CoinRotator's slower trend replaces the site trend |
| EMA tie-break | a 20/50 EMA settles days the two trends disagree |

**Always-long plan, the bar's baseline:** each published day, whatever the stance, Shumi's LONG line is scored under section 5 on the same coin, dates, costs and size. For context: holding from the first price after `T` to the scored plan's exit or lapse, and BTC over that window.

## 7. Costs and funding

Round-trip cost c: 0.4 % for ranks 1 to 100 at selection, 1 % for ranks 101 to 500, baselines included.

For a filled plan, the funding realised over the hold on the coin's Binance USD-M perp (`<SYMBOL>USDT`) is published as a signed figure, with the result including it. A coin with no Binance perp, or no funding data, reads `funding: not available`, never 0. Shorts are reported with and without funding. Also reported: the mean result with every stop exit filled 1 % worse.

A SHORT whose named venue has no perp or borrow for the coin at `T` is `no short venue`.

## 8. The bar

**First reading:** the first day on which 30 scored plans (filled and exited) have resolved, entered in at least 8 different UTC weeks. Short of either, it moves forward a week at a time. **Confirmed reading:** 60 scored plans from at least 12 weeks.

Each reading passes only if all four hold, after costs, in account result:

1. The mean of the scored plans is above 0, and so is the mean of weekly means (each entry week counted once).
2. Over every published day, Shumi's total beats the always-long total. Shumi's day is its scored plan, or 0 for NEITHER, `unparsed`, `no plan` or no fill. The long day is the LONG line's result, or 0 without a fill or a line. Open plans are marked at that day's close.
3. At least 40 % of scored plans end above 0 net R.
4. No scheduled attempt in the period is missing.

A NEITHER on a day the long plan lost counts for Shumi, and on a day it won, against. Each reading is published, dated, as "passed" or "not passed". Until then the record is "registered, unresolved".

Reported alongside, outside the bar:

* median net R, and the mean without the best plan;
* the busiest entry week's and most-repeated coin's share;
* splits by side and BTC above or below its 200-day average, month, rank band, and data score (0 to 4, 5 to 7, 8 to 10, 8 and up);
* `one-sided` when over 85 % of scored plans share a side;
* fill rate, lapses and NEITHER share, shadow lines and baselines.

## 9. Publication

Each entry gives:

* id `R-YYYY-MM-DD`, or `R-YYYY-MM-DD-b` for a retry;
* asked at, published at `T`, and the delay;
* coin, symbol and rank, with the selection counts;
* input ages;
* context at `T`: site trend, its age and days since its last flip, the shadow trends, whether the stance is with or against the trend, BTC's trend and its position against its 200-day average and prior cycle high, funding rate;
* labels, stance, both plan lines verbatim and parsed, and the scored side;
* size as printed and account risk %, exit line and scored rule, risk and venue;
* the full answer, verbatim, and the live Shumi link;
* the model and its provider;
* daily lines while open: the bar, the site trend, and `reached target` when it happens;
* the outcome: fill price and date or lapse reason; exit kind, price and date; gross move, net % move, net R, account result with and without funding; days held;
* with the outcome: best and worst close, highest high and lowest low in R and %; pullback at fill in ATR; every shadow line; the always-long result; holding and BTC over the same window.

The record says "take", "reading", "plan" and "result", gives counts with their sample size, and never uses "accurate", "win rate", "proven" or "profitable". No expected hit rate per regime is published. A cycle-high price or alert ladder in the answer is published as given, outside scoring. The page is append-only on docs.shumi.ai. Each entry also goes to Shumi's X and Telegram by 19:15 UTC and to the coin's community the same UTC day, each post linking the entry and its live Shumi link (`https://shumi.ai/s/<id>`).

## Changes from version 1

* Day one is 2026-10-31; version 1 entries are not scored. Version 1 picks, and open plans from version 1 entries that went live by 00:00 UTC the next day, still count in selection.
* Entries drop the CoinGecko id and instruction version tags.
* Each entry states its own `T`; version 1 read `T` from the page history.
* Prices come from 4h spot bars on Binance `<SYMBOL>USDT`, or on Hyperliquid `<SYMBOL>/USDC` for a coin with under 150 days of Binance history; version 1 named CoinGecko 4h OHLC.
* Funding comes from the coin's Binance USD-M perp, as the scorer has read it; version 1 named the plan's venue, or else Hyperliquid.
* Made explicit, as scored since version 1's day one: risk uses the zone midpoint, the move uses the fill price, and fills are read on 4h bars.
* Wording only: research notes are gone, and version 1's open items (baselines, costs, the faster-average shadow's 14 days and 20-day EMA) are fixed at their registered values.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.