Contact

What keeps running after the room has decided

Last updated 8 September 2026

AGI²⁺ began by improving the reasoning that happens before a commitment. The runtime extends that across time. It holds what the organisation believed and why, what it committed to, what should have happened by now, what actually happened, and whether the mechanisms behind the original decision still hold.

Two loops are described on this site. The Decision Loop runs inside one decision and stops at Learning. This one does not stop.

The AGI2+ persistent institutional runtimeA closed loop of twelve stations. World, then Observation, then Institutional State, which humans and machine work on together, then Decision and Commitment, Action, and World Changes. The loop turns and continues through Monitor, Verify and Falsify, Mechanism Drift and Adaptation, Replan, Update State, and Re-anchor. Re-anchor returns to Institutional State, so the runtime does not terminate.A PERSISTENT RUNTIME › NOT A PIPELINE THAT TERMINATESHUMAN + MACHINEWORLDOBSERVATIONINSTITUTIONALSTATEDECISION /COMMITMENTACTIONWORLD CHANGESMONITORVERIFY / FALSIFYMECHANISM DRIFT /ADAPTATIONREPLANUPDATE STATERE-ANCHOR↺ RE-ANCHOR TO THE INSTITUTIONAL STATE The AGI2+ persistent institutional runtimeA closed loop of twelve stations. World, then Observation, then Institutional State, which humans and machine work on together, then Decision and Commitment, Action, and World Changes. The loop turns and continues through Monitor, Verify and Falsify, Mechanism Drift and Adaptation, Replan, Update State, and Re-anchor. Re-anchor returns to Institutional State, so the runtime does not terminate. Shown as a single column for narrow screens.A PERSISTENT RUNTIMEWORLDOBSERVATIONINSTITUTIONAL STATEDECISION / COMMITMENTACTIONWORLD CHANGESMONITORVERIFY / FALSIFYMECHANISM DRIFT /ADAPTATIONREPLANUPDATE STATERE-ANCHORHUMAN +MACHINE↺
Observation enters the institutional state, humans and machine work on it together, and re-anchoring returns the loop to it.

The Persistent Institutional Loop

Fourteen stations, and the last one returns to the first

  1. State

    What the organisation holds as true, owed and unresolved.

    Gate: the state is written down rather than assumed.

  2. Objective

    What the state is being reasoned about, and by when.

    Gate: the objective names a decision, not a topic.

  3. Hypotheses

    The live explanations, including the strongest rival.

    Gate: a rival explanation exists and is testable.

  4. Commitments

    What was agreed, who owns it, and what it predicts.

    Gate: every commitment has an owner and a date.

  5. Due / overdue

    What falls due next, and what has passed without happening.

    Gate: overdue items stay visible.

  6. Monitor

    The signposts, watched at their thresholds rather than at review time.

    Gate: each signpost has a source and a threshold.

  7. Verify

    What should be observable by now, set against what is.

    Gate: expected and actual are recorded separately.

  8. Falsify

    The search for the observation that would show a hypothesis wrong.

    Gate: the falsifier was named before anyone went looking.

  9. Mechanism-drift check

    Whether the causal mechanism the decision relied on still operates.

    Gate: drift is separated from new information.

  10. Adaptation check

    Whether the actors have changed behaviour in response to pressure.

    Gate: a failed forecast is read four ways before one reading is chosen.

  11. Replan

    What changes now: the commitment, the owner, the date, or the decision itself.

    Gate: the change is authorised by the people who hold the authority.

  12. Score

    An internal assessment of how well the state is holding up.

    Gate: nothing here becomes a public number.

  13. Update state

    Beliefs, evidence, commitments and open uncertainty are revised in place.

    Gate: what was dropped is recorded with its reason.

  14. Re-anchor

    The revised state becomes the authoritative starting point for the next cycle.

    Gate: the next cycle begins here, not from a fresh prompt.

The fourteenth station returns to the first. What crosses that boundary is what the next cycle inherits: beliefs, evidence, unresolved uncertainty, commitments, deadlines, falsifiers, signposts, mechanism assumptions, observed adaptations and the original rationale. Not a summary of the file.

Memory and state

Why remembering is not enough

Memory answers

What happened. A model with memory can tell you what was said in March.

Institutional state answers

What remains true. What did we commit to? What should have happened by now? What actually happened? What has been falsified? What has changed? What is overdue? What must happen next?

Memory preserves the past. State governs what happens next. The difference is not the size of a context window. Institutional state is explicit and revisable, it names where each part came from, and the next cycle must start from it.

Commitments carried forward

Signposts are where this starts: an observable indicator, its source, its threshold, and the action when the threshold is crossed. The architecture carries them forward with the commitments they belong to.

For a consequential decision the runtime holds seven things per commitment:

  • Commitment. What was agreed or decided.
  • Owner. Who or what is responsible for it.
  • Due. When it should be fulfilled or reviewed.
  • Expected state. What should be observable by that point.
  • Actual state. What is observed.
  • Overdue. What should already have happened and has not.
  • Trigger. The observation that forces reconsideration or action.

None of this is project management under another name. A task tracker asks whether the work was done. The runtime asks whether the belief behind it still stands.

What happens when reality differs from the forecast

Falsify

Collecting supporting evidence is the easy half. For a load-bearing hypothesis the runtime asks which observation would show it wrong, records that while the hypothesis is still comfortable, and goes looking. Different from ordinary monitoring.

Mechanism drift

A forecast can fail while the original causal reasoning was reasonable, because the mechanism connecting cause to expected outcome has changed. So the question is whether that mechanism still operates as it did. New information arriving is not the same thing.

Adaptation

Governments, competitors, markets and adversaries respond to pressure and to each other. A failed prediction is poor evidence that the mechanism never existed. The runtime separates four readings:

  • Hypothesis failure
  • Mechanism persistence
  • External counterforce
  • Actor adaptation

Re-anchor

The Decision Loop ends at Learning: what did this episode teach. Re-anchoring asks the next question. What is now the authoritative state the following cycle begins from?

Where this stands

This page describes an architecture. AGI²⁺ v26.0 is built around a testable hypothesis: that long-horizon intelligence requires more than memory, and needs persistent institutional state, explicit commitments, falsification, and the ability to detect mechanism drift and actor adaptation. No benchmark result has been established for it.

SCORE, at station 12, is internal state assessment. The runtime may score its own state. It does not score people, and it does not collapse a consequential decision into a single visible number.

The method has been demonstrated on real decisions. The system that runs it is being built through engagements, and the long-horizon stations are the part being specified rather than the part in service.

The first operational application

The AGI²⁺ pre-commitment review is where the architecture meets one real decision: the frame checked, the cruxes named, the decision structure proposed, the signposts written. Those signposts are the seed of everything above.

The review