HANDOFF.md

Backtest Lab v1.0.0 — Handoff

Built end-to-end and deployed. Everything below was verified against the live Space, not just locally.

Live URLs

What	URL
Backtest Lab (Space)	https://huggingface.co/spaces/Bit-Trading-Company/bit-backtest-lab
Direct app	https://bit-trading-company-bit-backtest-lab.hf.space
Signal store (Dataset)	https://huggingface.co/datasets/The-Bit-Trading-Company/bit-signal-store

Both tagged v1.0.0. Space stage: RUNNING on zero-a10g (ZeroGPU).

Deployment status

Check	Result
Build	✅ 142s build + ~30s app start
Cold start (app boot)	✅ ~30s, under the 60s budget
HTTP response	✅ 200 in 0.74s
Hardware	✅ ZeroGPU (zero-a10g) granted — no billing escalation needed
Example config end-to-end	✅ /load_example returns the design's worked example
Backtest on live Space	✅ +33.2%, 12 trades, $486 costs — bit-identical to local
Warm run time	✅ 0.1s engine time for a 3-year daily run (budget: <2s)
Coverage map from live manifest	✅ 18 rows, all labelled real
Login button	✅ present; /login/huggingface returns 302
ZeroGPU detected in-app	✅ "ZEROGPU AVAILABLE · EXTENSION RUNS ON YOUR OWN QUOTA"

Tests: all green

bash tests/run_all.sh --slow → 191 passed, run with -W error::FutureWarning.

Verified on both Python 3.11 (dev) and Python 3.10 (what the ZeroGPU image actually provides) — 190 offline + 1 real-model test.

The six known-answer tests

#	Test	Status
1	Buy & Hold, zero costs, reproduces the asset return (tol 1e-9)	✅
2	Deliberately lookahead-biased strategy is caught structurally	✅
3	Zero-cost vs costed run differ by exactly the modelled costs	✅
4	Zero-signal strategy → flat equity, zero trades	✅
5	Walk-forward train/test never overlap; holdout untouched by selection	✅
6	Same config + same data → bit-identical results	✅

31 tests cover these six directly. Test 2 catches three distinct cheats: reading bar t+1, a centred rolling window, and full-sample normalisation.

Spec compliance

• ✅ No lookahead paths. Enforced two ways: the engine applies the one-bar shift itself (a strategy cannot bypass it), and assert_causal proves it by perturbing future prices up and down and requiring earlier outputs to be bit-identical. All seven shipped presets pass.
• ✅ Fills at next bar open. fill accepts only next_open; anything else raises. A test asserts the entry price equals bar 1's open, not bar 0's.
• ✅ Costs default ON. Costs().enabled is True. gross − costs = net holds to 1e-9 on every trade row. Costs-off runs are labelled "these numbers are not achievable" in the UI.
• ✅ Every displayed metric traceable to engine output. The stat band, trade table and report are asserted against metrics.py values in test_ui.py.
• ✅ No placeholder slices remain. All 18 signal slices are real Chronos-Bolt output. The PLACEHOLDER labelling machinery is still present and tested for contributors on weaker hardware.
• ✅ No arbitrary code execution. Adapter families are allow-listed, model ids are validated, trust_remote_code is never set, and the design's Custom (code) preset is rendered disabled with an explanation.
• ✅ Token hygiene. Scanned every file in both repos: the token appears in none of them. It exists only in the local gitignored key.txt and as the HF_WRITE_TOKEN Space secret.

What shipped

Signal store — 18 signal slices / 60,375 real rows across chronos-bolt-small + chronos-bolt-base, 6 assets (BTC, ETH, SOL, SPY, QQQ, NVDA) × 3 timeframes (1d, 1h, 15m); 15 price slices; 4 precomputed comparison tables (model performance, calibration, directional accuracy, strategy×timeframe heatmap).

Derived data worth reading: Chronos-Bolt is consistently over-dispersed — its nominal 80% q10–q90 band actually covers 83–89% of outcomes across every slice. Worth knowing before trusting a q10 stop level.

Engine — vectorbt, next-bar-open fills, commission + fixed/volume-scaled slippage, SL/TP/trailing stops, fixed-% and vol-target sizing, three validation modes with a locked holdout, full IS/OOS/holdout metric split, trade list with R-multiple, MAE/MFE and per-trade trigger reasons.

UI — the design implemented in Gradio Blocks with the Bit design system vendored verbatim (tokens + Styrene/Mac Minecraft fonts): Strategy Builder (5 numbered sections), Results Canvas (Overview / Trades / Comparison / Robustness / Report / Coverage), Run Manager, pinned disclaimer footer.

⚠️ Human-verify checklist

These need a real human with a real account — I cannot complete them:

1. OAuth login end-to-end. The button renders and /login/huggingface redirects correctly, but I cannot complete a real sign-in. Click Sign in with Hugging Face and confirm you land back signed in. If it fails, check Space settings → the hf_oauth: true header is set in README.md and may need the owner to approve the OAuth app once.
2. ZeroGPU extension with a real account. Sign in, open the Coverage tab, pick an uncovered range, press Estimate then Extend coverage. Confirm: progress stages appear, the commit lands in the dataset repo, and the toast reads "Coverage extended by @you". This is the only path that spends real GPU quota, so it is untested against real hardware.
3. Add-model flow with a real Hub model id, same reason.
4. Torch version difference. The Space runs torch 2.11.0 (ZeroGPU only accepts 2.11/2.10/2.9.1/2.8); my adapter smoke test ran locally on 2.13.0. The first real ZeroGPU extension doubles as the check that Chronos inference behaves identically on 2.11.
5. Visual pass on a real screen. I verified the UI functionally and via screenshots, but the in-app browser stalled before I could click through every tab at full size. Worth a human eye on the Comparison and Robustness tabs with several runs loaded.

Open items from DECISIONS.md

• D-001 — Space is under the personal namespace. Gradio Spaces under an org need a paid Team/Enterprise plan (402). You chose the personal namespace; the dataset stays under the org. To move it later, subscribe and re-point BIT_SPACE_REPO.
• D-006 — Binance is geo-blocked (451) from the seeding machine; crypto came from Coinbase via the fallback chain. Stooq now serves an HTML block page, so the equity chain effectively runs on yfinance alone. Tiingo is wired but needs a TIINGO_KEY secret to join the chain.
• Sentiment is a stub. Sentiment-Gated Momentum runs on a labelled, deterministic price-derived proxy behind a SentimentSource interface — not news sentiment. Wiring a real feed is the obvious next task.
• Pairs Trading is in the design's preset list but not runnable in v1 (needs a second leg). It is shown disabled with that reason.
• Equity intraday depth is provider-capped (~730d hourly, ~60d 15m) and recorded in the manifest as provider_max_days, not hidden.

Things I'd flag as genuinely uncertain

• The example strategy loses money on the locked holdout (−5.2%, Sharpe −0.38). That is the system working: the holdout is doing its job and the app reports it rather than burying it. Don't read the headline +33.2% as a result.
• Walk-forward on short ranges produces no OOS windows at all. The app now says so explicitly instead of printing 0.00, but it is easy to configure a run that looks complete and has no out-of-sample content. Watch for the red note.
• vectorbt is pinned to 1.0.0, not the 1.1.0 I developed against, because 1.1.0 requires Python ≥3.11 and the ZeroGPU image is 3.10. The full suite passes on both, but that is a real version skew between dev and prod.

Repo layout

hugging-face/
  key.txt            local only, gitignored, never committed
  DECISIONS.md       15 logged decisions across all phases
  HANDOFF.md         this file
  ops/               provisioning scripts (local only, not deployed)
  store_stage/       dataset card staging
  space/             → pushed to the Space
    app.py  src/{store,data,engine,strategies,adapters,charts,metrics,
                 comparisons,runtime,extension,config,ui/theme}.py
    scripts/seed_store.py   tests/ (191)   assets/ (vendored design system)
    DESIGN_NOTES.md  the extracted design contract Phase 3 was built against