v2.1.0 — Research Workflow 2.0

🧠 v2.1.0 — Research Workflow 2.0

Open-ended discovery, made tradable

• Structured discovery cards now turn messy research input into a durable hypothesis format: phenomenon, why alpha may exist, tradable universe, required data, signal construction, timestamp lag, holding period, failure risks, and source citations.
• The bridge discovery prompt now asks Claude Code / Codex to read beyond papers and news: earnings calls, company releases, Reddit, X, forums, GitHub, industry reports, sell-side notes, SEC filings, and regulatory filings.
• Each discovered family can carry source credibility metadata and a compiled signal recipe, while staying backward-compatible with older discovered-family payloads.

Idea-to-signal and research memory

• Added a deterministic Idea to Signal Compiler layer: universe, feature, rank rule, lag, hold, portfolio, formula, and rebalance rule are attached to every proposed strategy.
• Every completed experiment now stores a Research Workflow Audit: novelty check, point-in-time data contract, experiment registry 2.0 record, feature-store metadata, strategy-library card, paper-trading queue, memory graph, and live research feed events.
• Novelty checks flag known-factor overlap, momentum disguise, repeated tests, pool correlation, and similar historical failures.

Serious validation and production realism

• Added workflow reports for walk-forward validation, regime analysis, alpha decay, capacity, execution simulation, and benchmark comparisons against random rank, SPY, equal weight, sector-neutral momentum, and low volatility.
• The risk desk now includes explicit checks for source credibility, point-in-time data discipline, capacity/liquidity, and execution stress alongside the existing deflated Sharpe, PBO, costs, drawdown, and alpha-pool gates.
• Point-in-time policy is first-class: source availability lag, revision handling, as-of joins, universe membership, and leak checks are preserved with the experiment.

Human review mode and UI

• Added optional Human review before backtest in Settings. When enabled, the desk pauses after proposal so the boss can approve, reject, or edit the hypothesis before formal testing.
• The Fund & Research Board now summarizes the latest workflow audit: discovery, compiled signal, credibility, novelty, walk-forward pass rate, capacity, execution, and paper-trading status.
• Experiment detail pages now expose the full audit: discovery card, citations, validation windows, operations assumptions, feature-store record, point-in-time layer, baselines, and research feed.

Verification

• npm.cmd test passes: 28 tests.
• npm.cmd run build passes.

Historical simulations only · no brokerage connection · not investment advice. Market quotes and research sources are for personal/educational use.

v2.0.0 — Real markets & the full roadmap

📈 v2.0.0 — Real markets, the whole roadmap, and a game worth staying for

Real market data

• 20 years of real US daily data bundled (32 large caps + SPY, adjusted closes via the keyless Yahoo chart API; refresh anytime with node scripts/fetch-market-data.mjs).
• A real cross-sectional backtester: per-family price signals (momentum, reversal, low-vol, 52w-high, trend, seasonality, pairs, lead-lag, vol-managed, quality proxy), long/short rank portfolios with costs, in-sample/out-of-sample splits, real equity curves through 2008/2020/2022 — and no lookahead by construction.
• Real mode is the default; the deterministic mock simulator remains a setting.

Roadmap — all shipped

• Thompson-sampling direction bandit: explore / refine / repair / recombine, posteriors derived from history, every pick narrated in the idea's reasoning trace.
• Pool-level ΔSharpe reward (AlphaGen): experiments store daily return series and are scored by how much the candidate pool improves; pool correlations are now real Pearson correlations.
• MAP-Elites niche archive (QuantEvolve): family × horizon × risk grid steering exploration toward open niches.
• CSCV PBO: the desk's probability of backtest overfitting, live on the board.
• CLI research brain: hypotheses can come from Claude Code / Codex through the local bridge, validated against the knowledge base.

The game layer

• Boss XP & bilingual titles (Intern Boss → 量化教父) with level-up toasts.
• 16 achievements with a trophy wall, virtual fund NAV in the HUD.
• Fund & Research Board (click the meeting table): pool equity sparkline, niche grid, bandit posteriors, PBO.
• Confetti on promotions, rare office events between runs.

Polish

• Speech balloons push toward the emptier side and never cover colleagues (verified 10/10 clean samples); they ride along while characters walk, and text always stays inside the balloon art.

Historical simulations only · no brokerage connection · not investment advice. Market quotes are for personal/educational use.

v1.2.0 — 10x dialogue & CLI brains

🎭 v1.2.0 — ten times the dialogue, your CLIs as the brain

✨ New

• 150+ authored bilingual script templates (~460 lines) replace the small line bank. Conversations now condition on the live research context — experiment status, deflated-Sharpe survival odds, lineage generation, boss directives, even desk morale — and interpolate the real run numbers. Both English and 中文 are written natively, not translated.
• Claude Code / Codex CLI dialogue backends: run npm run dialogue-bridge and the characters' dialogue is generated through your already-authenticated CLI subscription (cheapest models, no API keys). Anthropic/OpenAI API-key backends remain.
• Bilingual research-memory lessons (no more English lessons inside Chinese chats).

🔧 Fixed

• Speech text now lives in per-balloon-type safe boxes — no more text spilling over cloud edges, spikes, or tails; CJK-aware bubble sizing.
• Ticker cards show two full lines instead of clipping mid-line.
• In-world leaderboard rows share the screen evenly.
• {family} placeholders show display names, not internal keys.

📦 Assets

• quant-research-lab-wallpaper.zip — rebuilt with everything above.

Historical simulations only · no brokerage connection · not investment advice.

v1.1.0 — 中文版 & review hardening

🌏 v1.1.0 — 中文版 & review hardening

New

• Full Chinese localization — flip the globe icon in the HUD (or Settings) and the whole game switches: HUD, boss bar, Love & Whip, modals, phase status, and complete Chinese dialogue banks for every conversation topic. 中文界面 + 中文角色对话。
• 中文 README: README.zh-CN.md
• Speech bubbles now scale with line length (large/XL balloons, up to 5 lines) and shift inward at the stage edges instead of clipping offscreen; in-world leaderboard rows no longer overlap.

Fixed (adversarial review pass, 8 confirmed findings)

• Agents no longer beeline through furniture after a rerouted walk (zone tracked at arrival, not departure)
• Conversations queue with priority preemption instead of silently dropping phase chats and love/whip reactions
• Early phases narrate the freshly proposed strategy, never the previous iteration's record
• Condenser line-swap race eliminated; Auto Run toggle no longer advances a phase when switching off
• Parameter mutation respects documented [min, max] bounds; boss-directive regexes no longer false-trigger on "sigma"/"gamma"/"152"

Assets

• quant-research-lab-wallpaper.zip — rebuilt with all of the above; drag onto Lively Wallpaper.

Historical simulations only · no brokerage connection · not investment advice.

v1.0.0 — The office opens

🏢 Quant Research Lab v1.0.0 — you're the BOSS now

The first full release of the autonomous anime quant office.

✨ Highlights

• Living office: six researchers walk between stations, hold multi-turn conversations about real experiment results, and gossip about desk lessons when idle.
• Honest research loop: 15 literature-grounded strategy families · UCB explore / lineage refine policy · deflated Sharpe ratio (Bailey–López de Prado) against the global trial registry · alpha-pool correlation gates · per-family alpha decay · research memory.
• Boss systems: directive bar (English / 中文) that steers the next hypothesis · Love & Whip with real mechanical consequences · idea-reasoning traces on every experiment.
• Dialogue brain: optional browser-direct condensation through Claude Haiku 4.5 or GPT-5.4-nano, with a free offline fallback that never goes quiet.
• Desktop wallpaper: chrome-free auto-running build with a draggable boss crown orb (Love / Whip / directive input right on the desktop).

📦 Assets

• quant-research-lab-wallpaper.zip — drag onto Lively Wallpaper and the lab lives on your desktop.

🔬 Research notes

The design synthesis (RD-Agent(Q), QuantEvolve, AlphaGen/AlphaAgent, FinMem/FinCon, Harvey–Liu–Zhu, McLean–Pontiff) with exact formulas and the roadmap lives in work/RESEARCH_DESIGN_DOC.md.

Historical simulations only · no brokerage connection · not investment advice.