bdistill-predict

Structured Prediction Assembly

Build predictions through a disciplined 4-phase pipeline: decompose the question, extract evidence, challenge your own reasoning, then predict with calibrated confidence. Every prediction produces a portable card with full provenance.

When to use

Make a structured forecast with evidence, not a vibes-based answer
Binary YES/NO questions with probability (prediction markets, Polymarket, Kalshi)
Scenario analysis ("what happens if the Fed cuts?", "will this merger close?")
Ground predictions in your extracted KB + current web data
Track accuracy over time with Brier scores and a resolution ledger

Input contract

required:
  question: string   # The prediction question ("Will the Fed cut rates before July 2026?")
  domain: string     # Domain slug ("macro-rates", "us-politics", "ag-commodities")
optional:
  binary: bool       # YES/NO mode with probability output (default: false)
  grounded: bool     # Recall KB entries + search web for current data (default: false)
  deep: bool         # 36 knowledge probes before extraction, ~25 min (default: false)
  medium_depth: bool # 12-18 knowledge probes, ~12 min (default: false)
  market_price: float  # Market-implied probability 0-1, enables edge calculation (default: null)
  timeframe: enum[1d, 1w, 1m, 1q, 1y]  # Resolution horizon (default: inferred from question)
output:
  card_id: string
  prediction: string           # "YES"/"NO" or directional statement
  probability: float           # 0.0-1.0 (binary mode only)
  confidence_boundary: [float, float]  # [low, high]
  evidence: array              # Tagged [kb]/[web]/[model]
  failure_modes: array         # What could make this prediction wrong
  card_path: string            # Path to saved prediction card

Output contract

format: JSON prediction card
location: data/predictions/cards/{card_id}.json
fields:
  card_id: string              # UUID or slug
  question: string
  prediction: string           # "YES"/"NO" or directional text
  probability: float           # 0.0-1.0 for binary, null for directional
  confidence_boundary: [float, float]
  evidence:
    type: array
    items:
      text: string
      tag: enum[kb, web, model]
      source: string           # URL, KB entry ID, or "model-knowledge"
  failure_modes: array[string]
  domain: string
  timeframe: string
  market_price: float | null
  edge: float | null           # probability - market_price (binary only)
  created_at: string           # ISO 8601 UTC
also_generates:
  - data/predictions/cards/{card_id}.html  # Shareable HTML card

With bdistill MCP (full power)

This is the primary workflow. The MCP server manages state, phases, and export.

START: Call bdistill_predict_start with question, domain, and optional flags (binary, grounded, deep, medium_depth, market_price, timeframe).

DECOMPOSE: Break the question into 4-6 sub-questions. Return a JSON array:

[
  {"text": "What does the latest dot plot signal?", "category": "event"},
  {"text": "Is PCE trending toward 2% target?", "category": "knowledge"},
  {"text": "What is the labor market signaling?", "category": "knowledge"},
  {"text": "How does a rate cut propagate to asset prices?", "category": "causal"}
]

Categories: rule (IF-THEN logic), event (scheduled/observable), knowledge (facts/data), causal (mechanism/transmission). Call bdistill_predict_respond with your decomposition.

DEEP PROBES (if deep or medium_depth): The server generates knowledge probes automatically. Answer each probe with domain-specific detail. Call bdistill_predict_respond after each answer. Deep mode runs 36 probes; medium_depth runs 12-18.
EXTRACT: For each sub-question, gather detailed evidence. If grounded mode is on, search the web and recall KB entries. Tag every piece of evidence:
- [kb] — from your extracted knowledge base
- [web] — from web search during this session
- [model] — from model training knowledge Call bdistill_predict_respond after each extraction.
CHALLENGE: Critique your own evidence. Structure as:
- Weak points: Which evidence is thin or stale?
- Missing angles: What did you not investigate?
- Key assumptions: What must hold true for your prediction?
- ADJUSTMENT: A number from -0.30 to 0.00 representing how much to discount your initial probability estimate. Typical range: -0.05 to -0.15. Call bdistill_predict_respond with your critique.

PREDICT: Final prediction as a JSON object:

{
  "prediction": "YES",
  "probability": 0.62,
  "confidence_boundary": [0.55, 0.70],
  "magnitude": "25bp cut, possibly 50bp",
  "assumptions": ["inflation continues declining", "no labor market shock"],
  "failure_modes": ["sticky services inflation", "geopolitical supply shock"],
  "timeframe": "1q"
}

Call bdistill_predict_respond with your prediction.

EXPORT: Call bdistill_predict_export to save the card (JSON + HTML). Offer to open the HTML card in the browser.

Standalone (no dependencies)

Same 4-phase pipeline. The agent manages state and writes output directly.

DECOMPOSE: Break the question into 4-6 sub-questions with categories. Write to a scratch variable or temporary file.
EXTRACT: For each sub-question, gather evidence. If grounded, use web search. Tag each piece [kb], [web], or [model].
CHALLENGE: Self-critique. Identify weak evidence, missing angles, key assumptions. Apply a probability adjustment (-0.30 to 0.00).
PREDICT: Assemble the final prediction card and write JSON to data/predictions/cards/{id}.json.

Card JSON schema for standalone mode:

{
  "card_id": "fed-rate-cut-jul-2026-a1b2c3",
  "question": "Will the Fed cut rates before July 2026?",
  "prediction": "YES",
  "probability": 0.62,
  "confidence_boundary": [0.55, 0.70],
  "evidence": [
    {"text": "March dot plot shows median expectation of 2 cuts in 2026", "tag": "web", "source": "https://..."},
    {"text": "PCE at 2.4% and declining", "tag": "model", "source": "model-knowledge"}
  ],
  "failure_modes": ["sticky services inflation", "geopolitical supply shock"],
  "assumptions": ["inflation continues declining", "no labor market shock"],
  "domain": "macro-rates",
  "timeframe": "1q",
  "market_price": 0.40,
  "edge": 0.22,
  "created_at": "2026-03-30T14:00:00Z"
}

Edge cases

Question too vague ("what happens next?"): Ask the user for specifics — timeframe, domain, binary or directional?
No KB entries for domain: Proceed with model-only and web evidence. Note lower confidence in the card and widen the confidence boundary.
market_price provided without binary: Ignore market_price, use directional mode. Warn the user that edge calculation requires binary mode.
Deep + grounded: Both flags compose. Deep probes run first to build context, then extraction uses KB + web. This is the highest-fidelity mode (~30 min).
Contradictory evidence: Surface the contradiction explicitly in the CHALLENGE phase. Do not paper over it — let the confidence boundary reflect the disagreement.

Example

Input: "Will the Fed cut rates before July 2026?" --binary --grounded --market_price 0.40

DECOMPOSE into 4 sub-questions:

What does the latest dot plot signal about rate expectations? (event)
Is PCE/CPI trending toward the 2% target? (knowledge)
What is the labor market signaling — cooling or resilient? (knowledge)
How are fed funds futures pricing cuts? (rule)

EXTRACT with [web] tags for current FRED data and dot plot, [model] for transmission mechanisms.

CHALLENGE: "Assuming inflation continues current trend, but services inflation is sticky at 3.1%. Labor market resilient — unemployment at 4.0%. Dot plot signals cuts but FOMC is data-dependent. ADJUSTMENT: -0.08"

PREDICT: YES, probability 0.62, confidence [0.55, 0.70], edge +0.22 vs market price of 0.40.

Composes with

bdistill-distill / bdistill-extract: Build a domain KB first for richer evidence when using grounded mode.
bdistill-calendar: Check upcoming events, use them as prediction scenarios.
Resolution workflow: Mark actual outcome on the card, compute Brier score (binary) or direction accuracy, re-extract on misses.
bdistill-predict-ledger: bdistill_predict_ledger returns your track record across all resolved predictions.
See references/prediction-card-schema.md for the full card format and resolution workflow.