Transforming complex market noise into structured investment intelligence
Pantheon Research combines quantitative frameworks, risk-regime models, governed market data, deterministic signal engines, backtest analytics, narrative research, and multi-model AI overlays across Global Macro, Equities, Crypto, DeFi, Technical Analysis, Commodities, Fixed Income, and Currencies.
AI should not replace the investor. AI should compound the investor's discipline.
BUIDL_QUESTS 2026 · OPC Hackathon Submission
| Resource | Link |
|---|---|
| Live Product | https://pantheon-research.com |
| Public Review Repository | https://github.com/0xjacobzhao-byte/pantheon-research-buidl-quests-2026 |
| Full Production Repository | https://github.com/0xjacobzhao-byte/Pantheon-Research — Private · closed-source; temporary read-only access available to judges upon request |
| Judge Verification Guide | docs/judge_evidence.md |
| Founder | Jacob Zhao · x.com/0xjacobzhao |
Pantheon Research is an institutional-grade cross-asset research command center. It is built for sophisticated investors, research analysts, allocators, and crypto-native market participants who need disciplined, explainable decision support across the whole opportunity set — macro, equities, digital assets, and FICC — rather than a dozen disconnected tools stitched together by hand.
The problem is not a lack of financial information. The problem is the absence of structured, governed, and explainable decision intelligence. Markets produce more data than any individual can process, but raw data is not a view, and a raw model opinion is not research. Pantheon sits between the two: it turns governed market evidence and disciplined investment frameworks into research that can be audited, compared, and acted on with confidence.
Pantheon is deliberately not a market-data dashboard, not a finance chatbot, not a raw-LLM opinion generator, and not an autonomous trading system. It is a research operating system in which deterministic frameworks produce reproducible scores and signals, a five-model LLM layer interprets and stress-tests governed evidence, and a human remains the portfolio manager. AI compounds the investor's discipline; it does not remove the investor's accountability.
This public repository is the open-source, judge-runnable review version of that system: a sanitized vertical slice you can clone and run in minutes, with a live Qwen + DeepSeek research-overlay comparison, evidence-pack provenance, data-quality governance, and reproducible tests — free of production credentials and proprietary strategy implementation.
Beyond the original dual-model demo, this public repo now ships six public-safe
vertical slices of the production system. All are offline-first, deterministic,
secret-free, and judge-runnable; navigate them from the top-level tabs
(Overview · Five-Model LLM · Macro Risk · Research Ops · Data Lineage · Paper Gateway · BTC Stack) and verify them with scripts/judge_smoke.sh.
| Module | What it proves | Docs |
|---|---|---|
| Five-Model LLM Cockpit | Production runs five LLM modules; this public repo now compares Claude, ChatGPT, Gemini, DeepSeek and Qwen over one hash-committed evidence pack, surfacing agreement/disagreement/red-flags/missing-evidence — no winner declared, no live paid call. | docs |
| Macro Risk Budget | Deterministic research logic: four-quadrant regime, hysteresis-confirmed regime, hard stops, exposure cap, fail-closed degraded output — works fully offline. | docs |
| Research Ops / Validation Console | Validation discipline: readiness, PIT policy, record kinds, and a hard no-alpha-claim contract — performance fields are null with a reason, never invented. |
docs |
| Canonical Data Lineage | Provenance: idempotent ingest, append-only vintage history, and end-to-end lineage from a provider record to the five LLM overlays. | docs |
| Paper / Shadow Trading Gateway | Paper-only and LIVE disabled by construction — no broker, no real order path; the LLM can propose but a human must approve, and approval only unlocks a paper simulation. | docs |
| BTC Three-Layer Decision Stack | Cross-asset depth: L1/L2/L3 layers + macro input resolved into a research posture (NO_TRADE/WAIT/SLOW_SCALE/RISK_THROTTLE/RESEARCH_ONLY) — never a live order. |
docs |
A single unified judge demo flow links every module
end-to-end at GET /api/judge/full-demo.
The private production repository remains closed-source; judges may request temporary read-only access. See security & sanitization for exactly what is and is not ported.
- Framework First. Investment discipline precedes AI. Frameworks define what matters, what invalidates a view, and when the system must refuse to conclude.
- Data Governed. Every observation carries provenance, freshness, quality, and provider state — missing or stale data is labeled, never silently guessed.
- Signal Is Not a Trade. Research output, portfolio judgment, and execution are separate architectural layers, by design.
- LLM as Analyst, Not Oracle. Models explain, challenge, compare, detect contradictions, and surface missing evidence — they do not generate positions.
- Human Remains Portfolio Manager. AI compounds discipline; it does not remove accountability.
Wrong strategy × AI = faster failure. Right strategy × AI = compounded discipline.
The architecture spans seven layers, left to right and top to bottom:
- External data sources — macro/rates, equities, crypto/DeFi, social/alt data, and positioning/market-structure feeds.
- Governed data platform — ingestion scheduling, provider health, validation and normalization, a PostgreSQL store of canonical observations, derived and product snapshots, evidence artifacts, TTL/freshness checks, and data-quality labeling.
- Research engines — Macro, Equity, Crypto, FICC, Technical Analysis, Narrative, Backtest/Validation, and Capital Flow Intelligence.
- Deterministic + five-model LLM layer — deterministic value/risk models, factor regressions, portfolio signals, and hard signals, alongside a five-model LLM research overlay (Source Pack → Prompt → Schema Validation → Overlay Comparison) that produces source-backed qualitative overlays, not trade execution.
- Information layer — the Pantheon Dashboard across all cross-asset modules.
- Signal & delivery layer — Telegram, user feed, research alerts, LLM signal channels, and a human-review gate.
- Trading layer — a staged roadmap that is manual today, with no live autonomous trading.
- Deployment stacks — core production on Vercel + Railway, with completed deployments on Google Cloud and Alibaba Cloud.
Vercel + Railway is the primary production path. Google Cloud and Alibaba Cloud are validated shadow / proof deployments, not equal production writers. The five-model research layer keeps evidence-backed conclusions separate from explicitly labelled model inference. Capital Flow Intelligence is an internal validation surface. Trading is independently gated and staged; live autonomous production trading is not claimed.
End-to-end flow:
Providers
→ Canonical Observations
→ Product Snapshots & Evidence Packs
→ Deterministic Research Engines
→ Multi-Model Research Comparison
→ Dashboard & Alerts
→ Human Decision
→ Independently Gated Execution
Full detail: docs/architecture.md ·
docs/deployment_architecture.md.
Pantheon Research is a live, multi-year research system. At a glance:
- a live cross-asset web product with mobile / PWA support;
- a mature strategy layer of versioned investment frameworks;
- a database-first information layer with governed provenance and data-quality states;
- deterministic research engines producing reproducible scores and signals;
- five developed LLM research modules with cross-model comparison;
- Research Ops and provider-health governance;
- a signal and research-distribution layer (alerts, Telegram, LLM channels);
- completed multi-cloud deployments across three stacks;
- validation, backtesting, and a staged trading-gateway roadmap.
The table below is deliberately explicit about scope: what exists in the full private production system versus what a judge can actually inspect and run in this public review repository.
| Capability | Full Pantheon Production | Public Review Repository |
|---|---|---|
| Strategy frameworks | Developed and actively used across all domains | Public-safe method documentation (docs/strategy_stack.md) |
| Information platform | Production database-first data architecture | Evidence-pack, data-quality & canonical data-lineage slices |
| LLM research layer | Five provider modules developed, live | Runnable five-model cached cockpit (Claude · ChatGPT · Gemini · DeepSeek · Qwen, offline) + live-capable Qwen/DeepSeek overlay |
| Macro & BTC research | Full production engines | Deterministic Macro Risk Budget + BTC three-layer decision slices (offline) |
| Multi-cloud deployment | Vercel + Railway · Google Cloud · Alibaba Cloud — completed & validated | Deployment documentation + secret-free proof artifacts |
| Trading | Manual today; paper / approval / constraint-bound execution on a staged roadmap | Paper-only, LIVE-disabled gateway demo |
| Validation | Backtest and forward-validation workflows | Research Ops / validation console (no-alpha-claim) + sample evidence |
The public repository now ships six offline, judge-runnable governance slices (see Full Public Migration). They prove the mechanism and governance — evidence packs, fail-closed provider states, multi-model comparison, data-quality, lineage, and a LIVE-disabled paper gateway — over bundled/cached data, not full production parity or live paid model calls. See Repository Access and Judge Review.
Maturity is tracked honestly, per capability, using a consistent vocabulary:
LIVE (production product) · BETA · INTERNAL (private only) · SHADOW
(proof deployment, not production) · EXPERIMENTAL · PLANNED · PUBLIC DEMO
(runnable in this repository). INTERNAL is never publicly available and
SHADOW is never production.
| Capability | Production maturity | Public evidence |
|---|---|---|
| Cross-asset dashboard | LIVE |
PUBLIC DEMO (offline slice) |
| Macro / cross-asset research | LIVE |
PUBLIC DEMO (Macro Risk Budget, offline) |
| Equity cockpit | LIVE |
PUBLIC DEMO (MA / NVDA) |
| Five-model comparison | LIVE (five providers, live) |
PUBLIC DEMO (five-model cached cockpit, offline) |
| Evidence-backed vs inferred lanes | LIVE |
PUBLIC DEMO (governance code + docs) |
| Research Ops / validation | LIVE |
PUBLIC DEMO (validation console, no-alpha-claim) |
| Canonical data lineage | LIVE |
PUBLIC DEMO (lineage + vintage history) |
| BTC decision stack | LIVE |
PUBLIC DEMO (three-layer, offline) |
| Paper / trading gateway | PLANNED / staged |
PUBLIC DEMO (paper-only, LIVE-disabled) |
| Signal Alert Layer | BETA |
Documented |
| Pantheon Pro delivery | BETA |
Documented |
| WeChat Mini Program | IN PROGRESS |
Documented |
| Capital Flow Intelligence | INTERNAL (validation surface) |
Documented only |
| Multi-cloud proofs | SHADOW (GCP, Alibaba) + LIVE (Vercel + Railway) |
Proof artifacts |
Status-label definitions are expanded in docs/architecture.md.
A public capability is listed as PUBLIC DEMO only when it is present and tested
on this repository's main — the six governance modules are runnable offline
over bundled/cached data (no live paid model calls). Broader modules (e.g.
Capital Flow Intelligence, live five-model paid calls) remain INTERNAL / private
production and are not claimed as public code.
Pantheon's strategy layer is a library of versioned investment frameworks, not prompts. Each framework encodes what matters in a domain, the conditions that invalidate a view, and the risk constraints that govern exposure. The table summarizes the domains; full proprietary thresholds and formulas remain in the private production repository.
| Domain | Framework / Engine | Core Method | Primary Output | Status |
|---|---|---|---|---|
| Global Macro | Macro regime engine | Tiered indicators: liquidity, real yields, credit, volatility, cycle | Regime classification, hard stops, exposure guidance | PRODUCTION |
| US / CN / HK / SG Equities | Stage-gated equity evaluation | Kill-fast screening, moat & cash-engine analysis, valuation triangulation, macro permission | Company verdicts, sizing, risk clusters | PRODUCTION |
| Narrative | Narrative lifecycle engine | Scarcity, lifecycle stage, reflexivity, liquidity gate | Capped-sizing, asymmetric-payoff calls | PRODUCTION |
| Technical Analysis | State-first TA engine | Market regime, normalized indicator features, dynamic weights, synthesis | TA state, execution & portfolio risk budget | PRODUCTION |
| Bitcoin | Multi-horizon BTC framework | Long-cycle bottom model, primary daily framework, short-horizon risk radar | Horizon-reconciled stance | PRODUCTION |
| Ethereum | Multi-quadrant ETH valuation | Settlement/security, monetary utility, network effect, revenue floor, regime weighting | Valuation stance with kill switches | PRODUCTION |
| DeFi | Protocol & yield risk engine | Protocol risk scoring, organic-yield analysis, liquidity & exit, counterparty | Verdicts: eligible / watch / avoid / data-gap | PRODUCTION |
| Fixed Income · FX · Commodities (FICC) | FICC allocation engines | Macro, valuation, positioning, structure, execution layers | Cross-asset allocation & carry views | PRODUCTION |
Selected framework deep dives
- Global Macro — a tiered indicator system spanning liquidity, real yields, credit, volatility, and cycle that resolves to a regime classification with explicit hard stops and exposure guidance, so downstream engines inherit a consistent macro permission.
- Equities — a stage-gated company evaluation: kill-fast screening removes disqualified names early; survivors pass through moat and cash-engine analysis, valuation triangulation, macro permission, position sizing, portfolio risk clusters, and monitoring with post-mortems.
- Narrative — tracks scarcity, lifecycle stage, and reflexivity behind a liquidity gate, with capped sizing and asymmetric-payoff logic rather than conviction-weighted bets.
- Technical Analysis — a state-first architecture: market regime and normalized indicator features feed dynamic weights and a synthesis step, with execution risk and a portfolio risk budget kept distinct from signal.
- Bitcoin — a long-cycle bottom model, a primary daily framework, and a short-horizon risk radar, with explicit conflict resolution when the horizons disagree.
- Ethereum — a multi-quadrant valuation across settlement/security, monetary utility, network effect, and revenue floor, weighted by regime with kill switches.
- DeFi — protocol risk scoring, organic-yield analysis, liquidity and exit mechanics, and counterparty/centralization checks that resolve to explicit verdicts (eligible, watch, avoid, or data-gap).
Full framework documentation — including thresholds, weights, and version
history — remains in the private production repository and can be reviewed by
judges under temporary access. A public-safe overview lives in
docs/strategy_stack.md.
Pantheon is database-first: research reads from governed, point-in-time observations, not from ad-hoc API calls at request time.
Providers / APIs / Filings / Web / On-chain / Social
↓
Ingestion + Provider Routing
↓
Validation + Normalization
↓
Canonical Observations
↓
Derived Snapshots + Product Snapshots
↓
Evidence Artifacts
↓
Research Engines + LLM Context
↓
Dashboard / Alerts / APIs
| Data capability | What it does | Why it matters |
|---|---|---|
| Canonical observations | Single normalized record of each governed data point | One source of truth; reproducible research |
| Provider routing & fallback | Chooses and fails over between data providers | Resilience without silent gaps |
| Derived & product snapshots | Pre-computed, frontline-ready research payloads | Fast, consistent surfaces for engines and UI |
| Evidence artifacts | Structured, hashable evidence packs | Provenance a model output can be verified against |
| Freshness / TTL & quality labels | Marks staleness and data quality on every field | Fail-closed governance, never a guess |
| Provider health & Research Ops | Live view of provider config, coverage, per-ticker state | Operational trust in the research surface |
The public slice implements this discipline end to end: evidence packs are bound
to a sha256 content hash (backend/app/evidence_pack.py),
data-quality and Research-Ops snapshots are served
(backend/app/data_quality.py), and module-level
data_state is exposed across Macro / TA / FICC / Equity
(backend/app/sample_modules.py). Full detail:
docs/data_platform.md.
Pantheon draws a strict boundary between deterministic computation and LLM interpretation.
Deterministic engines own normalized inputs, scores, market regimes, valuation outputs, hard stops, signal candidates, risk constraints, reproducibility, and audit trails.
LLM research modules own qualitative interpretation, contradiction detection, evidence-gap discovery, cross-model comparison, source-backed narratives, confidence assessment, and risk explanation.
Five production LLM research modules have been developed. The public repository ships a runnable five-model cached cockpit — Claude, ChatGPT, Gemini, DeepSeek and Qwen compared over one hash-committed evidence pack, offline with no live paid call (see Full Public Migration) — plus the live-capable Qwen + DeepSeek overlay.
| Model | Role in Pantheon | Governance boundary |
|---|---|---|
| Claude | Qualitative overlay & risk reasoning (public cached cockpit) | Reads governed evidence; never trades |
| ChatGPT | Qualitative overlay & comparison (public cached cockpit) | Reads governed evidence; never trades |
| Gemini | Qualitative overlay (Google Cloud integration; public cached cockpit) | Reads governed evidence; never trades |
| DeepSeek | Qualitative overlay (public cached cockpit + live-capable overlay) | Reads governed evidence; never trades |
| Qwen | Qualitative overlay (Alibaba DashScope; public cached cockpit + live-capable overlay) | Reads governed evidence; never trades |
The overlay workflow is evidence-first, not prompt-first:
Structured Evidence Pack
↓
Provider-Specific Prompt / Context
↓
Schema-Validated Model Output
↓
Cross-Model Comparison
↓
Agreement / Disagreement / Missing Evidence
↓
Human Review
Provider states are explicit: a missing credential yields
BLOCKED_BY_MISSING_CREDENTIAL, malformed output yields PARSE_ERROR, and the
comparison headline (LIVE_DUAL / OFFLINE_SAMPLE / MIXED / PARTIAL /
BLOCKED) never reports a hollow success.
Pantheon keeps two research lanes strictly separated:
- Evidence-backed — grounded in source packs, with provenance, an evidence tier, content-hash / source references, and data-quality & freshness state. A conclusion is only eligible to be presented as evidence-backed when those requirements pass.
- Model-inferred / AI-prior — explicit model reasoning that goes beyond the available evidence. It is always labelled, is never presented as sourced fact, cannot mutate a deterministic rating, cannot execute, and may only raise a verification task or a human-review requirement.
An AI prior can never masquerade as source-backed evidence.
| Governance capability | Pantheon implementation |
|---|---|
| Evidence provenance | Source packs and evidence artifacts bound to hashes / references |
| Fail-closed states | Missing, stale, blocked, parse, and provider errors stay visible |
| Schema validation | Structured model output validated before it is served |
| Multi-model comparison | Agreement and divergence surfaced provider by provider |
| Evidence hierarchy | Source-backed research separated from AI-prior inference |
| Human review | Disagreement and missing evidence create review requirements |
| Research Ops | Coverage, provider health, maturity, and audit surfaces |
| Signal separation | AI research does not directly execute trades |
Product availability vs. validation maturity. A research surface can be live
while its forward-return validation is still immature. Pantheon tracks product
availability, framework maturity, validation maturity, and
public-performance eligibility separately. BTC is among the more mature
validation tracks; equity forward samples are still accumulating; reconstructed
results and live forward results are kept separate; and validation-only data is
not a public alpha claim. Full detail:
docs/llm_research_layer.md.
Pantheon delivers research through product surfaces, not just modules.
| Surface | Audience | Function | Status |
|---|---|---|---|
| Web dashboard (Overview, Macro, US/CN/HK/SG Equity, BTC, ETH, DeFi, TA, FI, FX, Commodity) | Investors, analysts, allocators | Cross-asset research cockpit | LIVE |
| Research Ops / Data Quality | Operators | Provider config, coverage, per-ticker state | LIVE |
| PWA / mobile | Investors | On-the-go access | LIVE |
| WeChat Mini Program | Investors (CN) | Localized access | IN PROGRESS |
| Telegram distribution & alerts | Subscribers | Signal & research delivery | IN PROGRESS |
| Research / evaluation / data APIs | B2B, developers | Programmatic evidence & evaluations | IN PROGRESS |
| Membership / payments | All | Subscription foundations | IN PROGRESS |
Pantheon distributes research through two complementary mechanisms, kept architecturally distinct because they optimize for different things.
Deterministic delivery — scheduled research, alerts, and structured signals carrying provider and data-quality states, maximizing reproducibility and auditability.
LLM-assisted delivery — research briefs, synthesis, contextual explanation, Q&A, and model comparison, maximizing breadth and contextual depth.
The two are not interchangeable: deterministic systems are the record of truth;
LLM systems are the interpreter. Both feed a human-review gate before
anything reaches a subscriber, and neither executes a trade. Delivery channels
(user feed, research alerts, Telegram, LLM signal channels) are described in
docs/signal_and_delivery.md.
Pantheon runs on one code source, several deployment substrates, with exactly one canonical production writer:
- Vercel + Railway is the primary production path — Railway is the single canonical writer to the production database.
- Google Cloud and Alibaba Cloud are isolated shadow / proof deployments, not equal production writers. They exist to validate portability, regional deployment, model-provider integration, cost, latency, observability, and operational friction.
- The public repository is a sanitized offline judge demo — it is not a production deployment and does not write to any production database.
flowchart TB
PGH[Private production repo<br/>master · code source]
PGH --> V[Vercel<br/>production frontend]
PGH --> R[Railway<br/>production FastAPI · canonical writer]:::writer
PGH --> G[GCP Cloud Run<br/>Gemini shadow / proof]:::shadow
PGH --> A[Alibaba Cloud<br/>Qwen shadow / proof]:::shadow
V --> R
R --> DB[(Canonical PostgreSQL)]
R --> JOBS[Production jobs / scheduler]
G --> GDB[(Isolated shadow data role)]
A --> ADB[(Selected RDS mirror / shadow role)]
XGH[Public review repo<br/>main · this repository] --> DEMO[Docker Compose<br/>offline judge demo · no production writes]
classDef writer fill:#e8f5e9,stroke:#2e7d32,color:#1b5e20;
classDef shadow fill:#fff8e1,stroke:#f9a825,color:#5d4037;
| Environment | Role | Runtime | Data role | Writes / scheduler |
|---|---|---|---|---|
| Vercel | Production frontend | pantheon-research.com | — | — |
| Railway | Production backend / canonical writer | FastAPI + jobs | Canonical PostgreSQL | Enabled |
| GCP Cloud Run | Gemini shadow / proof | Scale-to-zero container | Isolated shadow role | Fail-closed OFF |
| Alibaba Cloud | Qwen shadow / proof | ECS + Nginx + Dockerized FastAPI | Selected RDS mirror | Fail-closed OFF |
| Public judge demo | Offline review slice | Docker Compose | Bundled / local data | No production writes |
The Alibaba deployment exposes a secret-free proof endpoint
(/api/proof/alibaba-cloud) that returns booleans only and makes no external
calls. Deployment triggers, canonical-writer / scheduler safety, provider-model
proofs, version parity, and rollback are documented in
docs/deployment_architecture.md; the precise
database scope is in docs/live_proof.md and
docs/alibaba_deployment_parity.md.
Non-claims: no three production writers; no active-active database; no automatic cross-cloud failover; no identical full production database clones; the selected Alibaba RDS mirror is not canonical.
Pantheon Research is built and submitted by Jacob Zhao as a one-person company. Jacob combines investment research, public markets, crypto, product strategy, AI-native systems, and full-stack implementation. One founder owns the investment hypotheses, framework design, product direction, data architecture, model evaluation, backend and frontend, deployment, documentation, and go-to-market.
The significance is economic, not agentic. AI-assisted development changes what a single disciplined operator can build: faster iteration, broader research coverage, multiple independent model reviews of the same evidence, generated documentation and tests, and one-person operation of multi-cloud infrastructure — collapsing coordination overhead that traditionally requires a research, engineering, data, and product team. Tools such as Claude Code, Codex, ChatGPT, Gemini, Qwen, DeepSeek, OpenClaw, Trae, and Qoder are implementation context; the innovation is the system and operating model, and the human founder remains the portfolio manager and the accountable decision-maker.
Pantheon's near-term commercial model is built around recurring software, research, and workflow revenue. Trading-related upside is deliberately treated as a later-stage opportunity that depends on validated signals, audited execution, and a real track record.
| Revenue Pillar | Target Buyer | Product | Revenue Mechanism | Current Status | Key Dependency |
|---|---|---|---|---|---|
| 1 · Subscription Research Platform | Individual investors, analysts, allocators, advisors, small teams | Cross-asset dashboard, AI overlays, model comparison, alerts, research briefs | Monthly / annual, tiered seats | Product live; pricing & packaging in progress | Willingness-to-pay evidence |
| 2 · Investment Skills Marketplace | Advanced retail, advisors, research teams, developers | Reusable, versioned framework "Skills" (Macro, US/CN/HK/SG Equity, BTC, ETH, DeFi Yield, TA, Narrative, FICC) | Per-Skill / bundle / enterprise / white-label | Framework base developed; marketplace on roadmap | Packaging & distribution |
| 3 · Research / Evaluation / Data APIs | Fintech, wealth apps, AI apps, funds, internal teams | Equity-evaluation, market-intelligence, and data/evidence APIs | Usage / per-report / quota / enterprise | Backend & artifacts developed; commercial API not launched | API packaging & SLAs |
| 4 · B2B / Institutional Licensing | Family offices, advisors, crypto funds, asset managers, fintechs | Dashboard licensing, custom workflows, white-label infra, premium services | Annual license + implementation + services | Foundation developed; GTM in progress | Design-partner validation |
| 5 · Long-Term Proprietary Strategy Upside | Own capital / mandates (future) | Validated strategies → controlled execution | Performance-linked, where legally permitted | Not a current revenue claim | Full validation + track record |
Commercial sequencing. Stage 1 — Software & research revenue (subscriptions, premium digests, paid evaluations) is the near-term focus: validate willingness-to-pay and build recurring revenue. Stage 2 — Team & institutional workflow revenue (multi-seat, advisor/family-office workflows, APIs, white-label) raises ACV and deepens integration. Stage 3 — Validated strategy monetization (paper sleeves, broker-integrated approval, licensed strategies) comes only after signal quality and execution discipline are proven.
No current revenue, user, AUM, or performance figures are claimed. Full detail —
customer segments, packaging logic, unit-economic drivers, moat, GTM motions, and
near-term experiments — is in docs/commercial_model.md.
Strategy first. Information second. Signal third. Execution last.
Execution is intentionally the final stage, gated behind validation and risk controls. Each phase must pass explicit gates before the next begins — this is a governed development plan, not a feature wishlist.
| Horizon | Strategic Goal | Product / Data & Validation | AI & Distribution | Execution | Commercial |
|---|---|---|---|---|---|
| Phase 0 · Current | Cross-asset research OS | Live web product; governed data; backtest/forward validation | Five-model overlays; alerts & Telegram | Manual only | Product live; GTM prep |
| Phase 1 · Now | Product hardening & signal validation | Parity/perf hardening; provider fallback, freshness, data-gap remediation; expanded backtests & forward samples | Five-model parity, cost & disagreement diagnostics; alert reliability | Manual only | Pricing experiments; paid evaluations; design partners |
| Phase 2 · Next | Research-to-execution control plane | Paper-trade harness (sleeve → paper fill → P&L → attribution); broker read-only | Evidence-linked signals; human-review routing | Approval-based (human clicks) | Advisor / family-office pilots; API pilots |
| Phase 3 · Later | Constraint-bound automation | Reconciliation, incident response | Model research never bypasses deterministic gates | Constraint-bound, inside pre-approved limits; kill switch; human override | Strategy licensing; managed workflows |
| Phase 4 · Long-term | Multi-account infrastructure | Cross-account reconciliation; jurisdiction/tax reporting | Advisor console; committee review | Multi-account orchestration | Enterprise license; per-account fee |
Every phase transition must clear these gates; a failed gate blocks progression.
| Gate | Question | Failure Result |
|---|---|---|
| Data | Is the data usable (freshness, coverage, provenance, provider health)? | Fail closed |
| Framework | Is the methodology frozen (versioned rules, tests, hard stops)? | Research only |
| Validation | Does the signal have evidence (backtest, forward sample, attribution)? | No execution influence |
| Operational | Can it run reliably (monitoring, reconciliation, audit trail)? | Manual only |
| Risk | Can loss be bounded (caps, kill switches, macro budget, drawdown)? | Blocked |
| Commercial | Will users pay (pilots, retention, pricing evidence)? | Continue research product |
| Regulatory | Is the operating model permitted (legal review, jurisdiction)? | No launch |
No calendar dates are claimed — phases are sequenced by evidence, not deadlines.
Full detail — per-phase deliverables, entry/exit criteria, exclusions,
dependencies, risks, and success metrics — is in docs/roadmap.md.
Public repository — https://github.com/0xjacobzhao-byte/pantheon-research-buidl-quests-2026 — is open-source, sanitized, self-contained, and judge-runnable. It is a representative slice of the production system, free of production credentials and proprietary strategy implementation, and it runs end-to-end offline with no secrets.
Full production repository — https://github.com/0xjacobzhao-byte/Pantheon-Research — is private and closed-source. It holds the full production codebase and research documentation, including proprietary frameworks, provider integrations, operational infrastructure, and production assets. This boundary is deliberate IP and operational-security governance. Temporary read-only access can be granted to BUIDL_QUESTS judges upon request.
- Review the live product — https://pantheon-research.com
- Read the architecture diagram (above) and
docs/architecture.md - Inspect the evidence-pack and comparison code —
backend/app/evidence_pack.py·backend/app/comparison.py - Run Docker and the smoke test (below)
- Read
docs/judge_evidence.md - Request temporary private-repository access for full production review where required
git clone https://github.com/0xjacobzhao-byte/pantheon-research-buidl-quests-2026
cd pantheon-research-buidl-quests-2026
docker compose up --build # frontend :5173 · backend :8000
./scripts/judge_smoke.sh # end-to-end smoke test (offline, no secrets)cd backend && python -m pytest # 84 backend tests
cd frontend && npm test -- --run # 9 frontend tests
cd frontend && npm run build # production build (tsc + vite)Manual setup (no Docker)
Backend (Python 3.11–3.12):
cd backend && python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --reload --port 8000Frontend (Node.js 18+):
cd frontend && npm install && npm run dev| Service | URL |
|---|---|
| Frontend | http://localhost:5173 |
| Backend API | http://localhost:8000 |
| API Docs (Swagger) | http://localhost:8000/docs |
Jacob Zhao — 0xjacobzhao-byte · x.com/0xjacobzhao
License: Apache-2.0 — see LICENSE. No API keys, private user data, live trading credentials, production secrets, or private financial records are included in this repository.