SimpleFunctions Thesis Builder

When to use

• The user has a hypothesis they want to structure (not just research a topic)
• An agent needs to attach a directional view to specific markets with falsification conditions
• A thesis from prior conversation needs to be evolved with new signals

Prerequisites

npm install -g @spfunctions/cli
sf login

Thesis lifecycle requires authentication — theses are user-scoped.

The 8-verb lifecycle

create  →  context  →  signal  →  evaluate  →  augment  →  fork  →  publish  →  unpublish

Each verb is a CLI command and an HTTP endpoint.

Instructions

Step 1: Create the root thesis

sf create "<thesis statement in one sentence>"

Example:

sf create "Fed cuts rates by July driven by jobs softening + CPI ≤ 2.5%"

The server runs an LLM classifier that:

• Rejects shell-escape garbage and obvious junk
• Expands the thesis into a causal tree (root claim → drivers → markers → markets)
• Assigns initial confidence

Returns the new thesis id. Capture this — every subsequent command needs it.

Step 2: Pull the generated tree

sf context <thesis-id> --json

Returns:

• Full causal tree (root + child nodes)
• Current confidence
• Related markets the LLM mapped to leaf nodes
• Implied edges (your thesis price vs. market price)
• Any pending signals

Inspect the tree. Common issues:

• Nodes too broad ("Fed dovish" instead of "FOMC voting 7-3 for cuts")
• Missing falsification conditions
• Wrong markets attached

If the tree is wrong, use Step 4 (augment) to add nodes, or fork (Step 5b) to start a variant.

Step 3: Inject a signal

When new information arrives — a news article, a Powell speech, a CPI release — feed it:

sf signal <thesis-id> "Powell speech: 'job market is softening more than expected'"

The server:

• Appends the signal to the thesis timeline
• Schedules the next monitor cycle
• Updates affected node confidences

Step 4: Evaluate

sf evaluate <thesis-id>

Triggers a deep evaluation (uses heavy model — costs more, slower). Returns:

• Updated confidences per node
• Confidence delta from prior evaluation
• Position recommendations
• Kill condition check

Run evaluate periodically (daily or after major signals), not on every minor update. Use signal for routine updates; evaluate is the expensive consolidation step.

Step 5: Augment (add new nodes)

sf augment <thesis-id>

LLM proposes new tree nodes (e.g., "Add: SEP dot plot at FOMC meeting"). You accept or reject each. Append-only — existing nodes are never deleted.

Step 5b: Fork (alternative branch)

To explore a contrarian or refined version:

sf thesis fork <thesis-id> "<variant>"
# or via MCP: fork_thesis tool

Creates a new thesis id with the same parent, independent confidence and signals.

Step 6: What-if simulation

sf whatif <thesis-id> --json

Returns implied prices if you assume specific node confidences. Useful for stress-testing the position.

Step 7: Publish

When the thesis is mature and you want it public:

sf publish <thesis-id> --slug fed-rate-cut-july

Makes the thesis live at simplefunctions.dev/thesis/fed-rate-cut-july. The slug is normalized (/[^a-z0-9-]/g strip, max 60 chars).

To unpublish:

sf unpublish <thesis-id>

Producing a useful tree

A good causal tree:

n1  Fed cuts by July                 65%
├── n1.1  Jobs report softens        75%   [market: KXNFP-26JUN-T100K]
├── n1.2  CPI ≤ 2.5%                 50%   [market: KXCPIYOY-26JUN-T2.5]
└── n1.3  No external shock          85%
n2  Yield curve responds             70%   [market: KX2Y10Y-26JUL-T0]
n3  Equity flows in                  55%   [market: KXSPX-26JUL-T5500]

Each leaf node should have:

• A specific, observable trigger (not "things go well")
• A confidence estimate
• An attached prediction-market contract (so you can measure your edge)

If a leaf has no market attached, either find one with sf scan --keyword "<leaf trigger>" or note that the leaf is unmeasurable.

Kill conditions

Always define what would invalidate the thesis. The heartbeat will check these:

sf heartbeat <thesis-id> --closed-loop-entry --closed-loop-exit --json

Configure heartbeat fields:

• killConditions: list of conditions (signals or market price thresholds)
• cadence: how often to re-check (default: 30m for active, 24h for monitor)
• evaluation_tier: which model to use (cheap for routine, heavy for evaluation)

See sf-thesis-monitor skill for full monitoring setup.

Errors and recovery

• Thesis creation rejected: LLM classifier thinks the input is garbage. Re-write as a clean directional claim.
• Empty tree after create: thesis too vague. Try with more specific markers ("Fed cuts" → "FOMC votes for ≥25bp cut by July 30").
• No markets attached to leaves: the leaf is unmeasurable. Either find a market or accept it's a qualitative input.

Related skills

• sf-research — topic research that informs thesis design
• sf-thesis-monitor — heartbeat + kill condition setup
• sf-pre-trade-check — adversarial check before opening positions linked to thesis

Reference docs

• <https://docs.simplefunctions.dev/build/thesis-lifecycle>
• <https://docs.simplefunctions.dev/api-reference/thesis>
• <https://docs.simplefunctions.dev/concepts/heartbeat>
• <https://docs.simplefunctions.dev/reference/concepts>