Capabilities › In-House Proof Case
Quant v1.1
A High-Stakes Decision Scenario — Used to Validate How AI Systems Get Governed
(Concretely, this is a quantitative trading system spanning US/HK/CN markets — but what this page tells you is not about trading capability, but what it proves across governance, evidence, and collaboration.)
This project is not here to show we can do quantitative trading. It is here to prove a different point: when AI suggestions, research systems, and human judgment all influence critical outcomes, the system must have boundaries, approvals, evidence, and execution discipline that cannot be bypassed.
Quant v1.1 is the in-house evidence for yunforce's three capabilities — governance, evidence closure, and human-led multi-agent engineering all hold under real pressure.
Three dimensions, three pieces of capability evidence
This case unfolds along three dimensions. Each dimension validates one of yunforce's three core capabilities — this isn't a reading order, it's an evidence order.
Where AI advice meets institutional discipline
A governance control plane that turns AI-advisor signals and research outputs into approved, auditable, executable decisions — no path bypasses approval to reach execution.
7 canonical RBAC roles · 16 canonical audit events · 0 execution bypasses
Read→Not a demo — an evidence closure
A phase doesn't close on "it looks like it works." It closes when all five evidence families — execution, governance, restart, discipline, heartbeat — agree.
7/7 fully-operational streak · 24 frozen configurations · 0 silent algorithm drift
Read→How we use AI to build AI systems — without letting AI lose control
Five agent roles plus one human approver. Three forced pause points. Final direction always stays human. This is how we keep AI in check inside our own engineering practice.
4 harness layers · 3 pause points · 1 final authority
Read→The project itself: read as a stack
Having read this far, you've seen capability evidence along three dimensions. The section below is about the internal engineering structure of Quant v1.1 — it explains why this single project can serve as evidence for all three capabilities at once.
Quant v1.1 is best understood as a three-layer stack, not a continuous implementation stream. Each layer solves one problem. Each layer inherits the one below it rather than quietly redefining it. That boundary discipline is the only reason later layers can stack safely.
Decide **what** deserves to trade
Boundary: produces a frozen research baseline. Later phases consume it, they don't silently retune it.
Prove the system can trade it **safely**
Boundary: does not create new strategy edge. Builds trustworthy execution, restart behavior, and evidence around the outputs already frozen.
Let AI advice and research outputs **flow in** without **bypassing** anything
Boundary: the execution daemon remains the only authority. No model, no API, no advisor reaches critical execution systems except through it.
Two counter-intuitive trade-offs
What's most worth showing about this project isn't the feature count — it's two decisions made deliberately against the prevailing industry narrative.
Phase 2 does not create new strategy edge. It creates trustworthy execution, restart behavior, and evidence around the outputs already frozen.
Many AI projects quietly revisit model or strategy choices during the engineering phase to make the final numbers look better. We chose to freeze research outputs and enforce a hard separation — so that every later operational and governance claim can honestly say "engineering changes were not hiding algorithm drift." It's a boundary. It's also a commitment.
Phase 3 is not an autonomous trading agent. It is the missing layer between sophisticated decision sources and a live execution system.
At a moment when "fully autonomous AI agent" dominates the narrative, we chose the opposite: AI suggestions, research outputs, and human discretion all converge into typed **governance objects** that require approval. Approval stays human. This isn't a retreat in technical ambition — it's a judgment about what kind of AI system institutions can actually adopt.
Each layer inherits the one below it rather than quietly redefining it. That is the hardest thing about this project — and the reason the three capabilities above can hold at all.