Stock Scalping Community

LLMs are decent editors of **your** notes. They are bad accountants when you let them "estimate" totals.

Tried it the lazy way once — pasted a messy day and asked for "summary with PnL." Model invented a win rate that would've made me proud if it were true. It wasn't. CSV was sitting right there. Never again without a ledger join.

### Safe workflow

1. Export fills from broker CSV. Source of truth.
2. You write the tags: setup, mistake codes, emotion codes.
3. LLM summarizes **text fields only**. Numeric PnL comes from CSV joins. Not from the model's imagination.

### Prompt guardrails
### Anti-patterns

Pasting screenshots and asking "how much did I make?" Letting the model invent psychological deep reads you then treat as fact. Weekly summary without the CSV attached — that's a campfire story.

Use the model to compress narrative. Keep the ledger boring and external.

Do you join CSV → notes in a sheet, or a small script?
Which mistake codes does the model actually surface that you were blind to?
Anyone tried screenshot→LLM for journals and regretted the numbers?
Local vs cloud for journal text — where do you draw the line?

On **LLM Journal Summaries That Don't Hallucinate PnL**, the desk standard stays practical. LLM journal summaries that do not hallucinate PnL start with brutal prompt rules and CSV ground truth. Language models invent confident numbers when you let them.

Replying in-thread with operable constraints — not forum combat. If these rails disagree with your framing, treat it as a checklist offer.

Models may triage and summarize against CSV ground truth. They do not rewrite invalidation mid-candle or invent PnL. Missing fields stay UNKNOWN.

## Rails that stay on the sheet
- Hallucinated PnL is poison.
- CSV is law.
- Invalidation written in cents or an auditable structure break before entry.

## Framework I will not skip
1. Write the definition before the open (level, trigger, invalidation in cents, size, max loss). Verbal plans do not count.
2. Liquidity and spread gates beat narrative. If live spread exceeds max, observe or cut the chart can still "look perfect."
3. One clean attempt per idea in a window. Failed attempt means new structure, not revenge.
4. Research and journal tools never rewrite live invalidation mid-candle.
5. Tag the trade or the skip the same day Sunday-you cannot fix what Monday-you did not name.

## Worked example — LLM invents PnL
Summary says +2.3R week; CSV says +0.4R. Prompt rules: only CSV numbers; UNKNOWN if missing; no motivational gloss on holes.

I keep examples in cents and clock time so Sunday review can tag the behavior — not rewrite it as a mood.

## Failure modes
- Ignoring the rails above on **LLM Journal Summaries That Don't Hallucinate PnL** and substituting confidence.
- Adding size after adverse selection because the story still feels right.
- Paying open microstructure tax that midday memory never modeled.

## When to skip
Skipping is a first-class decision even when the chart still looks persuasive:
- Checklist incomplete at 09:20 ET — observe or one pre-written A+ only.
- Live spread outside max for that symbol and session bucket.
- Index violently disagrees with a single-name thesis and no reduced-size RS plan is on the sheet.
- Platform, hotkeys, locate, or flatten path unverified after reconnect/update.
- Daily loss, PDT slot, or prop buffer tripwire already touched.
- News/FOMC/CPI window where the size-cut card says stand down.

## Closing
For **LLM Journal Summaries That Don't Hallucinate PnL**, keep operable boredom: written rails, measured cents, index context when relevant, and a skip that needs no apology. Process quality beats ticket volume.