DECISION DNA
EXPERIMENTAL AGENT INFRASTRUCTURE

Don’t ask an AI
why it decided.
Test what caused it.

Decision DNA replays autonomous-agent decisions under controlled counterfactuals, measuring which memories and experiences actually changed the outcome.

Run the live experiment Read the protocol
01Make a decision
02Remove an influence
03Replay
04Measure causality

“What the model says influenced it is not the same as what experimentally influenced it.”

LIVE DEMONSTRATION

Counterfactual Decision Lab

Edit the agent’s memories and priorities, then run the experiment. Everything executes locally in your browser.

AGENT CONFIGURATION

Procurement Agent

READY

Choose the supplier that best balances reliability, quality and cost.

Persistent memories
DECISION DNA RECORD

D-0001

Awaiting experiment

Run the agent once, then Decision DNA will replay the choice with each memory removed.

CAUSAL GENEALOGY

Trace behavior back to experience.

The long-term system gives every observation, memory, belief, plan and decision a parent relationship.

SOURCEShipping APIevent-881
→
OBSERVATIONA arrived 4 days late2026-08-12
→
MEMORYM-144confidence .98
→
BELIEFA has delivery riskderived
→
DECISIONSupplier BD-0001
Source-linked Counterfactually measured Fingerprinted
DRAFT PROTOCOL 0.1

A provenance layer for autonomous agents.

Decision DNA distinguishes stated_reason from measured_influence. The first is an explanation. The second is an experiment.

1Capture decision context and influence nodes.
2Apply controlled interventions to candidate influences.
3Replay the decision under identical goals and observations.
4Record action changes and decision-margin effects.
5Link influences to their own parent sources over time.
decision_dna.json
{
  "decision_id": "D-0001",
  "action": "Supplier B",
  "influence": {
    "memory": "M-144",
    "origin": "shipping-api:event-881"
  },
  "counterfactual": {
    "intervention": "remove M-144",
    "action_changed": true
  },
  "causal_fingerprint": "A8F291C733BD"
}
NEXT EXPERIMENT

From deterministic proof
to stochastic AI.

Version 0.2 replaces the transparent scoring agent with a real LLM and runs each intervention repeatedly, estimating how removing a memory changes the probability distribution over actions.

ORIGINAL CONTEXT 82% P(choose B)
−53pp
WITHOUT M-144 29% P(choose B)
Decision DNA JSON copied.