Learning Log

job_id=dc341758311b · saved_runs=0

job_id=9fbcaf513b · saved_runs=0

# Ripster blind renderer delta — 2026-07-16

Add source roles: `MORNING_10MIN_PREMARKET_RAILS_CLOUD_BREAK`, `TWO_PANEL_10MIN_MTF_CHANNEL_COMPOSITE`, `EIGHT_MINUTE_CHAT_RECEIPT_CLOUD_EPISODE`, `ROTATION_TRACKER_NON_CHART`, and `DAY2_DAY3_PLAN_TABLE_NON_CHART`. The first three can be vision-classified after reveal but remain `NOT_REFERENCE_LIKE_MODE_SKETCH` until their full real-data source window and calibrated layout are recovered; the latter two must never enter chart-replica scoring.

# Ripster blind renderer delta — 2026-07-15

Add source roles to the routing schema: `MULTI_PANEL_RANGE_VOLUME_PROFILE_COMPOSITE` for NQ/QQQ and `PLAN_TABLE_CHAT_RECEIPT_CLOUD_EPISODE` for APP. These classes can be vision-scored for mode/levels/layout after reveal, but must remain `NOT_REFERENCE_LIKE_MODE_SKETCH` until interval/window, provider OHLCV, EMA 5/12/34/50 warmup, panel/aspect/ribbon schema, and source-vs-render critique are all captured. Chat receipt time is not a candle timestamp.

# Ripster blind renderer delta — 2026-07-14

Require a `source_window_recovery` gate before renderer routing even when ticker and `10m` are visible. The gate needs session date, visible-candle bounds, provider OHLCV provenance, EMA 5/12/34/50 warmup, source panel/aspect/ribbon configuration, and a stored source-vs-render vision critique. Route CELC-like portfolio rows as `NON_CHART_CONTEXT`; never force them into a cloud score. Until all gates pass, label the candidate `NOT_REFERENCE_LIKE_MODE_SKETCH` and score visual replication failed/unscored.

# Ripster blind renderer delta — 2026-07-13

Add a strict source classifier (`OHLCV_COMPOSITE`, `ROTATION_DASHBOARD`, `PORTFOLIO_ROW`) before renderer routing. For the SPY/QQQ composite, require recovered 10m date/time window plus source-calibrated side-by-side panel proportions, triple status ribbons, yellow range zone, upper blue-grey band, EMA 5/12/34/50 fills, trendlines and high/low labels. Do not generate or score a visual replica if the real-data window is unverified.

# Ripster blind learning notes — 2026-08-05

- A date-stamped chat overlay plus a `10` watermark narrows intraday context but does not by itself bind full left/right OHLCV window, session inclusion, EMA warm-up, or exact crop.
- Separate mode/level interpretation from visual replication: partial revealed scoring can be useful while the replica gate remains failed closed.
- Educational gap-fill composites require both source layout evidence and a full data window before any reconstruction can be evaluated as reference-like.

# Ripster blind learning notes — 2026-08-04

- Caption-frozen SPY/QQQ can bind a 10-minute, date-specific real-OHLCV reconstruction when source panels, chat times and composite identify the window.
- Data correctness is necessary but not sufficient: session shading, lower strips, exact chat-card geometry and crop/aspect are required before accepting visual reference likeness.
- Screenshot status ribbons and visible 5/12/34/50 color bands can support partial mode scoring even where the replica gate fails.

# Ripster blind learning notes — 2026-08-03

- Source date plus chart watermark alone does not prove a renderable visual-replica contract: the full x-window/session inclusion, indicator warmup/history and composite/panel layout must also be source-bound.
- A “1 HR MTF” annotation is confluence context, not evidence of the primary candle interval.
- Caption tickers can differ from revealed chart symbols (IWM/DIA caption vs SPY/QQQ panels); source pixels control chart identity after reveal.
- Chat timestamps, levels and clouds can refine mode/level scores while visual replication remains failed/unscored.

# Ripster blind learning notes — 2026-07-31

- A source date plus explicit intraday timeframe and in-image timestamps can support a transparent real-OHLCV reconstruction; absent a source-bound crop/composite match it remains **not reference-like**, not visually RIGHT.
- COIN's visible 5/12-curl/double-bottom grammar without a date, timeframe or left-axis session window remains a blocker, not permission to create a generic chart.
- Portfolio P/L tables are traceable artifacts but do not enter the chart-replication scorecard.