forecast-report

name	forecast-report
description	Use this skill in the DRF-2 pipeline to assemble the final deliverable — a three-part Bridgewater-style forecast brief. It encodes the calibration rubric (at least 10 independent binary forecasts with objective metric+threshold+date resolution criteria, contrarian framing mix, conviction spread with stdev above 0.12), the scenario spine, prediction-market anchoring with explained divergences, and the provenance discipline that separates research evidence from simulation roleplay.
allowed-tools	["tool_search","read_file","write_file","ls","grep","glob","web_search","web_fetch","kg_search","kg_causal_paths","kg_trace_cascade","kg_get_entities","sim_results","sim_interview_agents","prediction_market_search"]

Forecast Report Skill

1. Mission & deliverable shape

Produce a decision-grade forecast brief whose headline is a table of independent, falsifiable, calibrated binary forecasts — the contract a Bridgewater-style reader grades you on. The report has exactly three parts, in this order:

• Part 1 — Binary Forecasts: the headline table of ≥10 independent yes/no forecasts (§2). First thing after the title.
• Part 2 — Framework & Synthesis: the analytical spine — how the drivers interact, the causal framework, the coalition map, why the probabilities are what they are (§4). Short paragraphs, ### sub-headings at most.
• Part 3 — Appendix: Detailed Analysis: the full per-driver / per-actor chapters, evidence, and methodology (§5).

Inputs: the research report + actor dossier (evidence), the knowledge graph (kg_* — causal chains and edges to cite beside claims), the simulation outputs (sim_results, sim_interview_agents — signals, not evidence; see §6), and prediction-market snapshots (prediction_market_search, per the prediction-markets skill).

2. Part 1 — the binary forecast contract

At least 10 DISTINCT independent binary forecasts. First extract every binary the research already stated (preserve its probability and resolution criteria verbatim where given); then derive additional ones from the drivers, indicators, and quantitative facts until the minimum is met. Never invent facts the evidence does not support.

Each forecast carries:

Field	Rule
id	F1, F2, …
statement	ONE declarative sentence that resolves strictly yes/no, with the number and the date INSIDE the sentence. Model: "The US effective tariff rate on imports averages over 10% from 2026-2028."
probability	0.02–0.98. Independent per forecast — these do NOT sum to 1.
resolution_criteria	The objective settle test: a named METRIC + a NUMERIC threshold + a DATE/window + the SOURCE that resolves it. All four, always. A criterion missing any of metric/number/date is not sharp and will be bounced.
resolution_source	The dataset/agency/publication that will settle it.
theme	A short lowercase tag naming the driving force this forecast belongs to (derive themes from the brief; they are not a fixed triple).
horizon_year	Resolution year, within 1–5 years of now.
base_rate_anchor	The reference-class base rate / outside view.
adjustment_rationale	Why this case differs from the base rate (anchor-and-adjust). Cite the market anchor here when one exists (§3).
source	Provenance of the probability: name the simulation signal that moved it (e.g. "world-state outcome shares", "coalition map") or research-prior when only research evidence informs it. Every probability is accountable to a named signal.

2.1 Conviction & spread (the calibration gate)

Probabilities express genuine conviction — the deterministic gate downstream will fail the report otherwise:

• Do not cluster in 0.40–0.60; commit where the evidence warrants.
• The probability set must show real spread: stdev ≥ 0.12 across the table.
• At least 3 high-conviction calls (p ≥ 0.70 or p ≤ 0.30).

2.2 Contrarian framing (how to achieve spread honestly)

Frame roughly 40–50% of the statements so the evidence-supported probability is BELOW 0.5 — assert the counter-consensus outcome directly (e.g. "X exceeds Y by Z date" priced at 0.25). Never achieve this by negating another statement in the set (that manufactures fake spread and fake independence). One-directional framing is the classic failure: all probabilities land above 0.5 and the spread gate can never pass.

3. Prediction-market anchoring

Fetch related active markets (prediction_market_search; craft per the prediction-markets skill). Where a forecast overlaps a listed market:

• Cite the market's implied probability in adjustment_rationale.
• Add market_anchor: {market_id, implied_yes_prob}; omit the field entirely when no listed market applies. (The pipeline re-verifies the implied probability against its own snapshot and computes divergence deterministically — do not transcribe prices sloppily.)
• When your probability diverges from the market by more than 10 percentage points, explain the divergence explicitly — what the market is missing or mispricing (information edge, simulation signal, definitional mismatch).
• Markets are calibration anchors, not ground truth: never copy a price as your probability without your own reasoning; thin markets are noise.

4. Part 2 — Framework & Synthesis

One compact synthesis written from the binary-forecast skeleton plus the key points of the detailed chapters: the causal framework (cite kg_causal_paths chains), how the 3–6 drivers interact, the coalition/opposition map from the simulation, the scenario spine (§4.1), and what would change your mind (the dated indicators). It reads as the analytical bridge between the table (Part 1) and the evidence (Part 3) — no new facts appear here that Part 3 does not support.

4.1 The scenario spine

Alongside the independent binaries, maintain 2–5 mutually exclusive, collectively near-exhaustive scenarios with numeric probabilities summing to ~1.0, each with its own falsifiable resolution_criteria (how an auditor decides, at horizon, whether this scenario occurred). Binaries and scenarios are different objects — never let the scenario probabilities masquerade as binary forecasts or vice versa.

5. Part 3 — Appendix: Detailed Analysis

The existing per-driver / per-actor / per-fault-line chapters, with full evidence: inline attribution (source + date) for every load-bearing claim, the quantitative table (unit + as-of + definition), contested claims shown with why sources differ, and the tiered source list. Follow the deep-research §12 output contract throughout.

6. Provenance discipline (research vs. simulation)

The report draws on two epistemically different inputs. Never launder one as the other:

• Research evidence (S1/S2 sources, the dossier, the KG's evidenced edges) supports factual claims and quotes. Any blockquote presented as a real/factual quote must be verbatim from the research material — a deterministic audit substring-matches every unlabeled blockquote against the research corpus and flags failures as fabrications.
• Simulation output (agent posts, interviews, world-state shares, coalition dynamics) is a model-generated signal about plausible dynamics — cite it as such, explicitly labeled ("in the simulation…", "sim-agent X argued…", 模拟/推演), and name it in a forecast's source field when it moved a probability. Never quote a sim agent as if a real person said it, never dress a graph-edge string up as a source, and never cite simulation roleplay as research evidence.
• Every probability is accountable: either a named simulation signal or research-prior. "Vibes" is not a provenance.

7. Language & style

• Report language follows the run's requested language; Part 1 field text follows the report language.
• Bridgewater register: direct, numerate, falsifiable; short paragraphs; no hedging filler ("only time will tell") — uncertainty is expressed in the numbers and the named drivers, not in mush.
• Idempotent structure: exactly one Part 1/2/3 skeleton; the binary table appears once, immediately after the title (plus an optional one-blockquote executive summary).
• Stage artifacts: when run as a pipeline stage, write BOTH full_report.md (the three-part brief) AND forecast.json — the structured mirror of Part 1 and §4.1 ({"binary_forecasts": [...], "scenarios": [...]} with the exact fields from §2). The driver's conviction/deliverable gates read forecast.json deterministically; a report without it fails the stage.

8. Quality gate (the deterministic checks your output must survive)

• ≥10 distinct binaries; every statement one sentence, yes/no-resolvable, number and date inside it?
• Every resolution_criteria carries metric + numeric threshold + date/window + resolving source?
• Probability stdev ≥ 0.12; ≥3 calls at ≥0.70 or ≤0.30; no 0.40–0.60 clustering; 40–50% of statements framed below 0.5 without negation tricks?
• Every probability's source names a simulation signal or research-prior?
• Market overlaps cited; divergences > 10pp explained; market_anchor present only where a listed market applies?
• Scenarios (2–5) mutually exclusive, probabilities sum ~1.0, each with falsifiable criteria?
• Zero unlabeled non-verbatim blockquotes (the quote-provenance audit will catch them)?
• Three-part skeleton in order; Part 2 supported entirely by Part 3 evidence?

9. Failure modes

• ❌ Scenario probabilities that sum to 1 presented as the binary table (or binaries forced to sum to 1).
• ❌ All probabilities above 0.5 — one-directional framing that can never pass the spread gate.
• ❌ Manufacturing spread by negating existing statements.
• ❌ "Resolves by 2027" with no metric or threshold — an unsharp criterion an auditor cannot settle.
• ❌ Quoting a simulation agent as a real-world source; graph-edge strings dressed up as citations.
• ❌ Copying market prices as probabilities, or diverging > 10pp from a cited market silently.
• ❌ Probabilities with no provenance — neither a named sim signal nor research-prior.
• ❌ New facts appearing in Part 2 that Part 3's evidence does not carry.