Tools Are Not the Workflow
Claude, MCP, AKShare — these are tools. Tools alone don’t produce research. Stringing them into a disciplined pipeline that converts raw signals into actionable analysis — that’s where the value is.
This article walks through the five stages of KSINQ’s research workflow. Each stage has a defined input, output, and quality gate. Fail the gate, don’t advance. No exceptions.
Stage 1 — Signal Detection
The signal detection layer continuously scans multiple data sources through MCP connections — market prices, news flow, QDII premiums, freight rate indices, macro indicators. → For how MCP connects these sources, see MCP for Investment Research
When a pre-defined threshold triggers, the system generates a Signal Alert. Two examples: a QDII premium exceeding its 90th percentile historical level; a divergence between physical and paper commodity prices exceeding two standard deviations.
Signal Alerts are not research conclusions. They’re starting points. An alert says “something is unusual” — it doesn’t say “what it means.”
Gate: Specificity. Every alert must identify the exact data point, the breached threshold, and historical context for the breach. “Market volatility is increasing” — that’s not an alert, it doesn’t pass.
A Dify workflow orchestrates this stage, running scheduled queries and applying threshold logic. Output is a prioritized signal list that arrives each morning before the researcher starts work.
Stage 2 — Context Assembly
Once a signal clears the gate, the next step is assembling analytical context around it. Claude queries relevant data sources automatically through MCP. → MCP connection architecture details in MCP for Investment Research
A concrete example: the signal is an unusual spike in a QDII US-index ETF premium. Context assembly pulls four types of data — historical premium series for that fund and comparables, underlying index performance and NAV, recent SAFE statements or QDII quota data, and sell-side commentary on QDII flows. Four dimensions, from different sources, assembled in a single pass.
Gate: Completeness. Assembled context must cover the signal from at least three independent angles — price data, fundamental data, qualitative information. If an angle is missing (say, no recent regulatory commentary exists), the gap gets explicitly flagged, not papered over. Better to label “insufficient information” than to pretend completeness.
Stage 3 — Analytical Reasoning
With context in place, core analysis begins. Multiple models collaborate at this stage, each handling the task types they’re best at. → Which model does what and how routing works, see Multi-Model Orchestration
The output of this stage: a thesis with a clear stance, an evidence chain with each link marked as fact or assumption (A→B→C), and the points where our analysis diverges from market consensus. In Howard Marks’ framing — second-level thinking. Not “what happened,” but “what expectation is the market pricing, and why we think that expectation is biased.”
Gate: Falsification criterion. The thesis must include a defined, observable, time-bounded condition under which it gets abandoned. If you can’t articulate “under what circumstances I’d be wrong,” the thesis doesn’t advance. This is Popper’s falsificationism hardcoded into the workflow.
Stage 4 — Three-Lens Review
This is the methodological core. Each thesis gets interrogated from three independent perspectives. AI assists the first two; human judgment governs all three.
Fundamental lens: Does the thesis hold up? Is the evidence chain logically consistent? Are divergence points from consensus clearly stated? Are assumptions explicitly marked? Claude does the initial screen and flags weaknesses, but the final call is human.
Risk lens: What happens if the thesis fails? Probability and impact of the core thesis collapsing? Does the falsification criterion cover the key risks? Do the assumptions survive macro shocks? Claude’s job here is constructing the adversarial case — the strongest argument against the thesis — then placing it side by side with the thesis for the analyst.
Market structure lens: Are timing and conditions right? Liquidity, cross-asset correlations, policy lags, catalyst windows. This lens is mostly human-driven — “market feel” isn’t a model strength yet, and I’m not sure that changes soon.
Gate: Unanimous passage. All three lenses must pass. Analytically elegant but unacceptable risk — doesn’t pass. Risk is controlled but market conditions don’t cooperate — doesn’t pass. Only theses that survive all three independent challenges enter the final research output.
Stage 5 — Output, Monitoring, and Prior Updates
Theses that clear the three-lens review get rendered into a standardized research memo — Executive Summary, Thesis, Evidence Chain, Supply Chain Signal, Risk Assessment (including falsification criteria), View Summary, Timeline. Fixed format, every time.
But publishing the memo isn’t the end. The workflow does two more things simultaneously:
Reverse-injection into signal detection. If the falsification criterion involves a specific data threshold (say, “copper drops below $8,000/ton”), that threshold gets automatically fed back into Stage 1’s monitoring layer. If there’s a catalyst date, a calendar alert is created. A thesis isn’t a static document filed away — it’s a live hypothesis under continuous surveillance.
Bayesian prior updates. Each completed research cycle adjusts the prior on the relevant theme. After a round of copper supply analysis, the model’s baseline assumptions about copper get updated. Next time a copper-related signal fires, analysis starts from a higher baseline that incorporates the last cycle’s findings. Over time, this compounds into an evolving knowledge base — each research cycle raises the starting point for the next.