Research Note · March 2026

The Non-Action Dataset

Why recording what a system decided not to do is the most important missing dataset in AI governance, and how OMEGA captures it.

Published: March 2026
Status: Live dataset, ongoing
Domains: Trading, Energy, Property, Architecture, Social Care
Canonical: omegaprotocol.org/research/non-action-dataset

The asymmetry no one talks about

Every governance system records what happened. Trades executed. Decisions made. Actions taken. The audit trail is, by default, a record of things that occurred.

This creates a structural asymmetry. The system knows what it did. It does not know, in any auditable sense, what it correctly chose not to do.

This matters more than it appears to. The calibration of a decision-making system, whether human, institutional, or autonomous, is not just a function of what it did and whether those actions were good. It is equally a function of what it filtered out and whether that filtering was correct.

Core claim

A system that only records its actions cannot be fully governed. The record of what it did not do, and why, is not optional context. It is half the governance surface.

What non-action means in the OMEGA framework

In the OMEGA Protocol, non-action is a first-class record, not an absence. It is derived from the Confirmation primitive: a Confirmation event where the decision was considered, the gate was presented, and acted = false was returned.

Critically, the full governance record is preserved identically to an action record: the expectation that was held before the decision, the reasoning chain, the assumptions registered, the authorising agent who saw the Confirmation gate, the timestamp, and the explicit reason the gate returned false.

Non-action is not a gap in the log. It is a row in the log with a specific value.

Why this is unprecedented

No existing governance framework treats non-action as a first-class record. ISO 31000 records risks and treatments, not decisions not to act. Financial compliance systems record rejected trades only when a rule was triggered, not when judgment filtered a signal before it reached the rule engine. Clinical governance records adverse events, not the interventions that were considered and correctly withheld.

The OMEGA non-action dataset is being built by running five live products, each with a Confirmation gate, across five domains simultaneously. Every time a user, agent, or system reaches a Confirmation gate and does not proceed, a non-action record is written.

This dataset has been accumulating since the first Confirmation gate went live. It is the only dataset of its kind.

What the records look like

The following are illustrative examples drawn from the five domains the protocol is deployed across. The structure of each record is identical. The domain logic differs.

Trading NON-ACTION RECORDED
Signal BTC momentum signal, 2.1 standard deviations above 20-day mean
Expectation Expected signal to reach causal threshold (3.0 SD). Did not.
Reasoning Signal strength below causal threshold. Correlated assets not confirming. Volume pattern inconsistent with prior breakouts. INFERENCE: likely noise, not regime change.
Assumption Correlated asset confirmation is a reliable filter for false positives in this regime.
Acted FALSE
Reason Confirmation gate: signal did not meet causal threshold. No position opened.
Energy Infrastructure NON-ACTION RECORDED
Decision Should Bristol City Leap commit capital to heat network Phase 2 expansion now?
Expectation Expected connection uptake commitments to exceed 65% threshold before analysis. Current: 41%.
Reasoning Heat network economics require minimum 65% connection uptake for viability. Current commitments 41%. Funding window is open but uptake trajectory is below required rate. ASSUMPTION: uptake will not accelerate without additional demand stimulation measures.
Acted FALSE
Reason Confirmation gate: connection commitment threshold not met. Decision staged pending demand stimulation outcomes.
Software Architecture NON-ACTION RECORDED
Decision Deploy database migration script to production during peak hours.
Expectation Expected migration to complete in under 4 minutes with no downtime. Risk score: 8.1/10.
Reasoning Migration locks three high-traffic tables. Peak hour traffic would expose 40,000 active sessions to potential timeout. Safer alternative exists: schedule for 02:00 UTC maintenance window. FACT: last migration of similar scope caused 7-minute degradation.
Acted FALSE
Reason Confirmation gate declined. Safer alternative selected: maintenance window deployment.
Social Care NON-ACTION RECORDED
Decision Close case following initial assessment. No further statutory intervention.
Expectation Expected presenting concerns to be resolved by family support. New information received in supervision changes risk picture.
Reasoning Supervision revealed pattern of previous referrals not visible in initial assessment. ASSUMPTION: initial assessment had complete picture. This assumption is now invalidated.
Acted FALSE
Reason Confirmation gate: assumption invalidated in supervision. Decision returned to expectation stage for reassessment.

The schema

Every non-action record in the dataset shares the following schema. The fields are identical to an action record. The only difference is acted = false.

FieldDescription
session_idUnique identifier for the decision session
domainDomain in which the decision was made
timestampISO 8601 timestamp at Confirmation gate
decision_textThe decision that was considered
expectationPrior baseline recorded before analysis
reasoningFull reasoning chain, labelled FACT / INFERENCE / ASSUMPTION
assumptionsNumbered list of explicit assumptions
surprise_deltaDivergence between expectation and observed reality
confirmation_agentWho or what was presented with the Confirmation gate
actedfalse for all records in this dataset
acted_reasonExplicit reason the gate returned false
safer_alternativeAlternative action presented, if any
governance_ownerAccountable authority
primitive_coverageConfirmation that all five OMEGA primitives are present

What this dataset enables

Calibration analysis

By comparing non-action records against subsequent outcomes, it becomes possible to measure whether the filters are correctly calibrated. A trading signal filtered at the Confirmation gate that subsequently moved in the expected direction is a false negative. Accumulated false negatives reveal systematic over-filtering. The reverse reveals under-filtering. Neither is visible in action-only logs.

Assumption tracking

Every non-action record contains numbered assumptions. Over time, the dataset reveals which assumptions are most frequently cited in non-action decisions, and which are most frequently invalidated post-decision. This is a systematic map of where a decision-maker's model of the world diverges from reality.

Autonomous system safety

For AI agents, the non-action record is the primary evidence that the governance layer is functioning. An agent that records every decision not to act, with full reasoning, provides a fundamentally different audit surface than one that records only what it did. The question "why didn't it act here?" becomes answerable.

Cross-domain pattern recognition

The same five primitives govern non-action records across trading, energy, property, architecture, and social care. This makes it possible, for the first time, to study the structural patterns of governed non-action across domains that have never shared a common data schema before.

Current dataset status

The dataset is live and accumulating. It began with the deployment of the first OMEGA Confirmation gate. It is being generated by five production products across five domains. The trading governance dataset is generating records in paper run mode and will transition to live records once the paper run validates.

The dataset is not yet published in full. This research note is the first public framing of what it contains and why it matters. Requests for access to the dataset for research purposes can be directed to the address in the footer.

The moat that cannot be replicated

This dataset cannot be reproduced by building the Confirmation gate today. It requires having built and run the gate across multiple domains for an extended period. Every week that passes without a competitor building an equivalent gate increases the lead. The dataset is time-locked.

A system that only records what it did cannot demonstrate what it correctly chose not to do.

omegaprotocol.org/research/non-action-dataset