starchild-ai-agent/@2195/quotesim-tape-validation

QuoteSim real-tape validation (agent-driven, no CLI for the user)

QuoteSim (open source: 2195/quotesim in the Starchild community-projects repo, live demo at https://community.iamstarchild.com/2195-quotesim) simulates a two-sided MM quoting strategy on any oracle-backed asset. Its fast OHLC model overcounts fills (no queue position), so it self-reports a low trust score and withholds dollar figures until validated. The trustworthy path: record a live trade tape against the strategy's dry-run quotes for a few days, then score the fill model on data it never saw. The user must never touch a terminal — you drive everything via scheduled_task. This skill is how.

Setup — locate or install QuoteSim

Find the product dir, in this order (set $QS accordingly):

1. output/quote-sim-product/ — original author's workspace layout
2. output/quotesim/src/ — installed from the community repo
3. Not found → install it:

cd /tmp && rm -rf scp && \
git clone --depth 1 --filter=blob:none --sparse \
  https://github.com/Starchild-ai-agent/community-projects scp && \
cd scp && git sparse-checkout set 2195/quotesim && \
mkdir -p /data/workspace/output/quotesim && \
cp -r 2195/quotesim/* /data/workspace/output/quotesim/ && \
pip install flask requests --quiet

The runner (validate_runner.py) needs no API keys — it polls free public Hyperliquid endpoints. Inside Starchild it routes via sc-proxy automatically (httpcompat.py); outside it uses plain requests. Verify once: cd $QS && python3 validate_runner.py --collect --coin BTC should print a COLLECT line, not a traceback.

When to use

Trigger on any of: "validate <ASSET>", "get a real number for <ASSET>", "make <ASSET> trustworthy", "run the shadow logger", "how do I trust this", or a low (0–44/100) trust score the user wants resolved.

Hard limits (state these honestly)

• Assets need a real public tape. Source: Hyperliquid public endpoints (recentTrades + metaAndAssetCtxs). Every HL perp qualifies. Map the user's asset to its HL coin symbol (e.g. HYPE, BTC).
• HIP-3 builder-dex markets (commodities/equities on HL) ARE validatable. They live on a separate universe — natural gas = xyz:NATGAS, gold = xyz:GOLD, TSLA = xyz:TSLA. recentTrades accepts the prefixed coin directly, but mark lookup needs {"type":"metaAndAssetCtxs","dex":"xyz"} (bare call will NOT find them → "no mark" error). The runner handles xyz:* natively and sanitizes : in shadow filenames (HL_xyz_NATGAS_*.jsonl). Pass VAL_COIN=xyz:NATGAS, never bare NATGAS.
• Pyth-only RWA feeds cannot be tape-validated — Pyth is oracle-only, no trade tape. Check the HL xyz: dex FIRST (it covers many commodities and equities with real tapes); only fall back to "risk-shape-only" if absent.
• Thin markets take time. recentTrades returns ~10 prints per poll; a 1-minute poll captures low/mid-liquidity books well (exactly the "should I list this?" candidates). On very thin tapes the 200-fill gate can take WEEKS — tell the user a realistic timeline after the first day's fill count, don't promise "2 days" blindly. Very high-volume majors undercount between polls — note it; a websocket tape is future work.

Readiness gate (in the runner)

--score withholds all dollar figures until ≥200 predicted fills AND ≥2 days of tape. Below that it returns ready:false with a COLLECTING progress line. This is the integrity rule — never surface partial-sample dollars to the user.

Workflow

1. Confirm the asset + config

• Resolve the HL coin (reject Pyth-only RWA with the honest note above).
• Use the same config the user ran in the dashboard if known; else the auto-tuned defaults. Pass it as a JSON string via VAL_STRAT.

2. Register the COLLECTOR (runs every minute, multi-day)

Use scheduled_task in command mode (cheap, no LLM). The runner persists inventory across cycles in shadow/HL_<COIN>_state.json, so each run appends to the same sample. Replace $QS with the absolute product dir.

scheduled_task(action="schedule",
  title="QuoteSim tape collect — <COIN>",
  schedule="every 1 minute",
  deliver="local",                       # silent; no push per cycle
  command="cd $QS && VAL_COIN=<COIN> VAL_SIDE=<SIDE_NOTIONAL> "
          "VAL_STRAT='<STRAT_JSON>' python3 validate_runner.py --collect")

Record the returned job_id (topic or memory) so you can cancel it later.

3. Register the DAILY SCORER (checks readiness, pushes when ready)

A once-daily command task that runs --score and prints only when ready (empty stdout = silent push-free run, zero cost). Convert the user's local time to UTC for the cron.

scheduled_task(action="schedule",
  title="QuoteSim tape score — <COIN>",
  schedule="0 13 * * *",                  # pick ~daily, user's evening
  command="cd $QS && python3 validate_runner.py --score --coin <COIN> "
          "| python3 -c \"import sys,json; d=json.load(sys.stdin); "
          "print('✅ '+d['coin']+' validated — trust the dollars now: '"
          "+str(d.get('total_per_day'))+'/day (spread '+str(d.get('spread_per_day'))"
          "+', directional '+str(d.get('directional_per_day'))+', '"
          "+str(d['fills'])+' real fills over '+str(d['days'])+'d). Re-run it in QuoteSim.') "
          "if d.get('ready') else ''\"")

4. Tell the user the simple version

One message: what you started, a realistic timeline ("collecting real tape on <COIN>; I'll message you the moment it's trustworthy — days on liquid books, weeks on thin ones"), and that they do nothing. Don't expose cron/job_ids/ paths unless they ask.

5. On completion

When the scorer reports ready, cancel the collector job (scheduled_task action=cancel job_id=...) so it stops consuming cycles, and optionally the scorer too. Confirm to the user and point them back to the QuoteSim dashboard, where the asset now clears the trust gate.

Manual one-off (debugging only — not the user path)

python3 validate_runner.py --collect --coin HYPE      # one cycle
python3 validate_runner.py --score   --coin HYPE      # current verdict

Notes

• deliver="local" on the collector keeps it silent (no per-minute spam).
• The collector is read-only (public endpoints, no keys, no orders).
• If the user changes the config mid-collection, the persisted inventory no longer matches — delete shadow/HL_<COIN>_state.json and restart collection.
• Progress check any time: run --score ad hoc; ready:false includes a fills/days progress line you can relay ("34 of 200 fills, day 1.2 of 2").