Read-only strategy opening preview¶

scripts/strategy_preview.py is a standalone F1 diagnostic, not a trading command. It compares T1–T10 and TB's 50:50 T1/T8 opening blend, using an existing model run's forecast cutoff and cached Polymarket bid/ask quotes. Choose any F1 event with an existing diagnostic; there is no default source run or hardcoded event.

It never fetches market data, saves forecasts/signals, changes active profiles, or places orders. PostgreSQL connections use repeatable-read transactions with default_transaction_read_only=on, verified before reading. Model inputs are loaded through that same transaction. No migration or database write is involved.

Run and freeze inputs¶

Run from the repository root in the installed project environment. Replace EVENT_KEY and RUN_ID below with an existing event/source diagnostic pair. Store generated outputs under ignored reports/, never commit them.

LOCAL (Mac) — create the 4k preview (reads the configured database):

mkdir -p reports/strategy-preview
PYTHONPATH=. .venv/bin/python -m scripts.strategy_preview \
  --event EVENT_KEY --source-run RUN_ID --sims 4000 \
  > reports/strategy-preview/4k.json

LOCAL (Mac) — compare 16k simulations with the same frozen inputs:

PYTHONPATH=. .venv/bin/python -m scripts.strategy_preview \
  --event EVENT_KEY --source-run RUN_ID --sims 16000 \
  --snapshot reports/strategy-preview/4k.json \
  > reports/strategy-preview/16k.json

LOCAL (Mac) — render saved JSON, without a database connection:

PYTHONPATH=. .venv/bin/python -m scripts.strategy_preview \
  --render reports/strategy-preview/4k.json \
  --output-dir reports/strategy-preview/4k

The renderer writes Markdown and CSV: a candidate comparison, per-profile picked.csv / considered-not-picked.csv, and combined decision tables. Every supported market appears in each HK/QK decision table, including excluded, illiquid, stale, incoherent, or unpriced markets. A rejected decision has zero allocated cost/shares; any standalone watchlist size is a separate hypothetical target, without the portfolio cap, not an approved allocation.

HK/QK default to a fresh $1,000 balance, $50 per market, $250 total. Override them with --bankroll, --max-market, and --max-total; keep those values identical in frozen simulation comparisons. Original sizing retains the profile's supported sizing fields. These are empty-position opening allocations, not rolling season bankrolls.

Frozen JSON pins quotes, recorded 24-hour liquidity (including explicitly empty venues), observation time, source run, event, cutoff, code commit, sizing and model input fingerprint. A mismatch fails rather than mixing old/current inputs. The fingerprint covers the repository model's cutoff view; it is not a full database backup. A 16k run may reuse a matching stored diagnostic model; other variants are computed in memory with each profile's seed (default 42). A flushed heartbeat runs every five minutes during computation. Ad-hoc snapshots predating this version's schema must be recreated.

What this does and does not show¶

Timing gates alone are bypassed: “enter now” is hypothetical, even when a profile would normally wait for another stage. Other model settings and market filters apply.
Buys use cached asks, or 1 - bid for NO, plus strategy slippage and venue fees. This differs from historical Polymarket sweeps that execute at the midpoint.
Quotes must be valid two-sided quotes no more than ten minutes old. Touch depth and live fee schedules have not been verified.
Zero recorded tape is not evidence of zero live trading activity.
Linked Kalshi markets, unknown prediction kinds/market parameters, maker settings, and thin-market exceptions are explicitly unsupported and fail closed. No linked Kalshi markets is reported as empty coverage, not a successful Kalshi strategy test.
This is not an execution recommendation. There is no trade execution or active configuration change.

Publishing this script requires focused synthetic tests and lint. A merge is not proposed until the usual local racinglines check, non-live pytest, and strict MkDocs gates have also been run. Merging code deploys code only: no migration or data step.