Answer
A model produces language for a person to interpret.
Annual report · 2026
The Machine Evidence Era.
AI spent decades learning to answer. Systems are now beginning to use tools and take actions. This report asks what evidence must survive when those actions matter.
Historical milestones are sourced. Forward-looking material is explicitly labelled emerging or scenario.
01Compute
02Models
03Agents
04Machine decisions
05Machine evidence
01 / 07
The path to action
Seventy years of systems learning to do more.
The history is not a straight line. These milestones mark changes in capability, deployment and governance—not a claim that one event created the modern field alone.
Artificial intelligence becomes a named research field
The Dartmouth Summer Research Project helped establish artificial intelligence as a field of study.
Primary source ↗Deep learning changes image recognition
A deep convolutional network achieved leading results in the ImageNet classification challenge.
Primary source ↗The Transformer architecture is proposed
The paper “Attention Is All You Need” introduced an architecture based on attention rather than recurrence.
Primary source ↗Conversational models reach a mass audience
ChatGPT launched as a research preview built around conversational interaction.
Primary source ↗AI risk management becomes more operational
NIST published its Generative AI Profile as a companion resource to the AI Risk Management Framework.
Primary source ↗Agent platforms connect models to tools
New agent APIs and SDKs made tool use, orchestration, tracing and guardrails explicit product primitives.
Primary source ↗Systems move from answers toward actions
Tool-using software and physical systems create a new requirement: preserve the external evidence available before an action.
Emerging phase · not a forecast date
What evidence will future machines rely on?
02 / 07
The next layer
AI no longer only produces an answer. It can call a tool, change a system and move something in the world.
Tool use
A system retrieves current information or changes external state.
Action
A software or physical decision creates consequences beyond the model output.
What evidence will future machines rely on?
03 / 07
The evidence gap
An action can outlive the context that produced it.
A live source changes. A response is overwritten. A policy evolves. Without a preserved boundary, reviewers may see the outcome but not the evidence available at the time.
Source
Where did the observation come from, and under what licence and cadence?
Time
When was it observed, received and used?
Boundary
What was checked, what remained unknown and who owned the decision?
What evidence will future machines rely on?
04 / 07
The evidence lifecycle
Evidence belongs before the decision.
The sequence matters. Evidence is not invented after the outcome. A caller applies its own policy and retains responsibility for the resulting action.
Named source
Observation
Evidence
Caller-owned policy
Advisory decision
Receipt
Verification
Replay when available
Target architecture. Current legacy receipt families and verification capabilities remain documented separately; a signature proves integrity and attribution, not objective truth.
What evidence will future machines rely on?
05 / 07
Where the question appears
Different industries. The same missing record.
The relevant sources and safety controls change by domain. The need to preserve attribution, freshness and integrity does not.
AI agents
Current context behind a tool call or automated workflow.
Robotics
Weather, route and hazard reports available before movement.
Logistics
Port, flight, route and disruption context behind dispatch.
Cybersecurity
Threat and vulnerability sources behind a response.
Financial services
Market, risk and regulatory inputs behind an automated decision.
Government
A reviewable record of current public-source context.
Space
Orbital, space-weather and positioning observations with time attached.
150 operating contexts
Explore the complete Project Atlas audience map.
What evidence will future machines rely on?
06 / 07
The live interface
Project Atlas makes the evidence layer visible.
Atlas plots current observations while keeping the source, observed time, freshness and response-integrity state attached. An Atlas observation is not automatically a receipt.
What evidence will future machines rely on?
07 / 07
The proposed infrastructure layer
This is where Dynamic Feed enters the story.
Dynamic Feed is building infrastructure for live, source-attributed machine evidence: observation interfaces, signed response surfaces, receipt paths and careful verification boundaries.
What evidence will future machines rely on?
Source notes