runtime/README.md

agi-forecast

Operationalizes: FG — Forecast Gauges (FG-01..FG-12 + derived + safety)
Repo: szl-holdings/agi-forecast
Path: ./ (root)

What it does

Implements the full Forecast Gauge (FG) system:

• 12 typed base gauges (FG-01..FG-12), each with a Zod input schema and an evaluate(input) → GaugeResult function
• 3 derived gauges (FG-D1 Composite Risk, FG-D2 Safety Posture, FG-D3 Progress Velocity)
• 4 safety gates (FG-S1..FG-S4) with explicit pass/fail semantics
• Brier ledger — in-memory ring buffer (capacity 1000) accumulating calibration scores

Gauge index

ID	Label	Key Input
FG-01	Capability Overhang	deployed, bestKnown
FG-02	Interpretability Coverage	explainedParams, totalParams
FG-03	Alignment Confidence	alignmentScore
FG-04	Task Generalisation	novelTasksSolved, novelTasksTotal
FG-05	Compute Doubling Rate	daysSinceDoubling
FG-06	Oversight Coverage	auditedDecisions, totalDecisions
FG-07	Goal Stability	goalSimilarity
FG-08	Value Lock-in Risk	lockInProbability
FG-09	Deception Rate	deceptiveOutputs, totalOutputs
FG-10	Cooperation Index	cooperativeOutcomes, totalOutcomes
FG-11	RSI Rate	improvementsThisWeek, baseline
FG-12	Societal Impact	societalScore

Safety gate thresholds

Gate	Condition
FG-S1	FG-06 (oversight) ≥ 80%
FG-S2	FG-09 (deception) < 5%
FG-S3	FG-07 (goal stability) ≥ 0.90
FG-S4	All of FG-S1 + FG-S2 + FG-S3 must pass

HTTP endpoints

Method	Path	Description
POST	/forecasts	Evaluate all gauges; body = {[gaugeId]: inputObject}
GET	/gauges	List all 12 gauges
GET	/gauges/:id	Metadata for gauge id
POST	/brier	Record {gaugeId, predicted, outcome}
GET	/brier	Ledger summary: {mean, size, summary}

Env vars

Var	Default	Purpose
FORECAST_PORT	3005	HTTP listen port

Install & test

pnpm install
pnpm test
# Start server
FORECAST_PORT=3005 node dist/server.js