Live camera feed ✕ stop

Person A

Watching for motion…

The camera is driving the math. A small pixel-diff motion detector runs on each frame (locally, no upload) and finds the moving region in the scene. That horizontal position becomes the starting zone for the heatmap below — so when you move left or right in front of the camera, the forecast actually shifts. Counterfactuals (Walking forward / Stopping / Turning back) then evolve from where Person A is right now. Production replaces this naive pixel-diff with the full perception layer (DINOv2 / V-JEPA 2 + live_scenario.py); shape of the answer is the same, accuracy is much higher.

Active world: Stock market — Fed decisions

If the Fed cuts rates, where does the index go?

Every cell is one possible future. The darker the cell, the more likely. Click any cell for the exact probability + a plain-English read of that specific outcome.

Today's index: 3,200. Rows = % change from today.

Probability of each outcome click a cell for detail · click a column for that time-slice's full distribution

Preview scenario

Preview only — switch back to Stock market or Live perception to see a real forecast.

Likelihood: rare most likely

Day 15 — distribution across all outcomes at this time. ✕

worst0%best

Most likely the index drifts close to flat over the next 30 days, with probability widening as the horizon stretches. Single-scenario projection from the stock_market engine fixture (day-1 anchor: 3,200). Not a live calibration — this scenario hasn't accumulated historical validations yet. Don't trade on this.

What would change this picture?

How I came to this answer

Six steps from your inputs to a real answer — the master plan §7 spine.

1

Read what you brought in.

Input: Stock market scenario · counterfactual: No Fed action.

2

Figure out what you're asking.

Forecast the distribution of the index's % change over the next 30 days, conditional on the selected counterfactual.

scenario: stock_market · target: distribution over market.index_value

3

Pick the rules of this world.

One operator: market kinematics at full weight.

operator_graph: [ kinematic w=1.0 order=0 ]

4

Run the simulation.

Project step-by-step; each step's distribution becomes the prior for the next.

engine: QueryEngine · world_line: 8c4d2147…

5

How sure am I?

The heatmap intensity at each cell is a deterministic Gaussian density given the step's mean and uncertainty. Sigma grows like √t. The scenario hasn't accumulated calibration data yet — these are shape-of-the-answer bands, not calibrated probabilities.

calibrated: null · reason: cold-start, no validations

6

Plain answer.

Under no Fed action, the most-likely outcome stays close to flat over 30 days.

Why this lives in a drawer: transparency matters, but the answer should come first. ChatGPT shows you the answer; Omytea shows you the answer and, on demand, every step.

Technical details

Raw payload from a real warmed demo session.

0. Input

source: "scenario:stock_market"

1. NL → parser

"如果美联储降息"   // canonical Chinese phrasing the regex parser matched
matched_pattern: "rate_cut" · parser_confidence: 0.85

Honest disclosure. Today's hand-rolled nl_parser matches a small Chinese pattern set; the English UI translates to the canonical trigger. The v4.1 M5 OmyteaCompiler-LLM slot is the designed path to an English-native compiler (no model downloaded in v4.1).

2. BeliefProgram (intent IR)

intent_kind: "rate_cut"
scenario_id: "stock_market"
target_kind: "distribution"
distribution_variables: ["market.index_value"]

3. OperatorGraph

entries: ["kinematic" w=1.0 order=0]
graph_id: "scenario:stock_market"

4-6. dsl.Query · engine_result · calibrated

engine_result.value: {"market.index_value": [(3200.0, 1.0)]}
world_line_id: "8c4d2147-9fc5-4e69-af2c-effa18ace1d1"
calibrated: null   // cold-start, no validations

Engine returns a single-point distribution today. The heatmap expands it into a deterministic Gaussian over the row variable with sigma growing as √t — a shape placeholder, not a live forecast. When the transition operator returns a multi-point distribution per tick, this view shows that real data instead.

Backing routes: POST /api/query · GET /api/snapshot · GET /api/analytics/calibration. Camera / file inputs: POST /api/live/start + POST /api/stream/frame (live_scenario.py) — wired in backend, this static demo only shows the live preview locally. Cycle plan: WORK_PLAN_V41.md.