ORDIVON
Menu
Question dossierOrdivon Computingtesting

When should an Agent act, wait, observe, refuse, reconcile, or request a decision?

Treat non-action as a timed, evidence-bearing decision rather than the absence of Agent capability.

All research questionsOpen project

01 / Current position

A hypothesis is not the current judgment.

The dossier preserves the live research position. Dated articles preserve the complete evidence and argument that changed it.

Working hypothesis

Host-local DecisionRequests, explicit UNKNOWN and wait states, deadlines, evidence requirements, and paired act/abstain evaluation can preserve authorized utility with fewer operator interruptions than uniform approval rules.

Current judgment

Testing with a stronger real-domain non-action case. Finance QB6 let one Agent request exact read-only evidence rather than infer missing execution facts, then update from more_research to no_op when the current minimum contract exceeded both the frozen 10% target and existing 25% canary bound; the branch ended before simulation, Decision, Proposal, or financial write.

02 / Decision boundary

What keeps this Question alive?

Why it matters

Acting too early can duplicate or commit the wrong Effect; waiting too long can lose a deadline, resource, or strategic option. Static approval-everywhere and model-only escalation both waste scarce attention or miss consequence boundaries.

Next admitted test

Repeat evidence-conditioned act/abstain trajectories in materially different domains and with state changes that make a previously impossible action become expressible; retain new non-action state only if simpler owner-native evidence cannot reconstruct the same decision.

Condition for deletion

A simpler static policy or always-act baseline achieves equal authorized utility, timing, recovery, and operator cost with fewer persistent states.

03 / Supporting publications

Complete arguments connected to this Question.

4 dated publications currently document this research line.

Research report

The Agent Asked for More Evidence. The Answer Became No.

Finance QB6 as a concrete example of evidence-driven non-action: an Agent requests missing read-only evidence, updates its judgment, and terminates before simulation, Decision, Proposal, or financial effect.

12 August 2026 ↗Research report

We Cut 203 Observations to 8. The Agent Still Refused to Edit.

How Harness P5 moved from exhaustive repository traversal toward Agent-owned high-information discovery, why minimizing observation count alone failed, and why the successful result was an evidence-backed refusal to mutate source.

12 August 2026 ↗Research report

The Smaller Core That Survived Strong Baselines

How Core Work System Round 1 used strong classical baselines and live Codex/Hermes replacement to remove a Task Runtime, reject a generalized Context Kernel, shrink Effect, localize DecisionRequest, and retain provider-neutral Task state.

31 July 2026 ↗Research note

UNKNOWN Is an Operational State, Not a Model Feeling

Why unresolved external reality must be represented as system state rather than inferred from model confidence or collapsed into failure.

31 July 2026 ↗

04 / Source discipline

The dossier is an index, not the evidence authority.

Question metadata

Owns the current judgment, next test, and deletion condition.

Publications

Own complete dated arguments, limitations, comparisons, and source links.

Repositories

Own exact code, tests, releases, receipts, and machine evidence.

Related frontier

testing

When should a finite intelligence expand its interface to Reality instead of only reasoning harder?

Testing under a contracted theory. External prior art falsifies the claim that action/interface expansion is absent, but it also weakens any absolute transcendence claim: callable, representational, affordance, and executor expansions can be new relative to one closure while remaining compositional inside a deeper supplied substrate. No universal basis hierarchy or generic basis-escape object is admitted.

open

Which Task Runtime objects are required by asynchronous waiting and Join semantics?

Open. Round 1 showed that mature workflow engines can carry durable work mechanics, but no interrupted fan-out and fan-in workload has yet proved which semantic Join facts must remain visible above those mechanics.

reframed

Which responsibilities survive materially different workloads without becoming one mandatory stack?

Reframed and answered at the responsibility level. Game, Finance, Security, World, Web, Host, Harness, and Runtime now provide materially different consumers and deletion tests. The recurring result is not a universal Host→Harness→Runtime stack: Computing's canonical Core explicitly treats subsystem names as non-invariants and retains only responsibilities that remain unowned after mature mechanisms and owner-native composition. Cross-domain pressure has repeatedly localized, deleted, or kept different packages while preserving a smaller set of ownership/identity/currentness/recovery distinctions.

answered

Which Agent-native responsibilities remain after strong classical baselines?

Answered within Round 1. Strong baselines reduced the proposed core; later Host and Harness extraction clarified the surviving ownership split without adding a universal Agent platform.

answered

What is the smallest explanation that preserves the causal boundaries an Agent needs?

Answered on the EX3–EX7 Agent-facing surfaces. All treatments reached the same exact-action ceiling across 1,326 accepted decisions; richer representations added token burden, so the preregistered smallest-non-inferior rule selected compact prose.

answered

When does valid evidence stop being current enough to govern a decision?

Answered for the current evidence horizon. Security EC1 separated byte identity from semantic applicability; World independently separates historical occurrence from current relation/Presence; Web source-bound publication distinguishes exact owner revision from current editorial judgment and exposed stale static/test consumers; Finance further showed that an already coherent state-changing Decision loses Proposal authority when its PortfolioSnapshot baseline is no longer current. Computing Knowledge now records both HistoricalValidity != Currentness and AdmissionCurrentness != CommitmentCurrentness. These cases support owner-native revalidation and have not forced a generic Freshness, impact-graph, or automatic rebase service.

answered

Does canonical evidence order make an Agent understand the evidence better?

Answered for P11. Six unordered datasets collapsed 720 raw permutation digests to one stable identity, but after invalidating a truncated Tool-wire campaign and clarifying an ambiguous evaluator target, raw presentation scored 44/48 strict versus 42/48 stable list and 43/48 identity map. No semantic normalization advantage survived.

answered

Which Agent-era world-model laws survive historical out-of-distribution pressure?

Answered for the HD0–HD9 programme. The frozen 1760–2026 campaign closed with 320 trajectories and 96 Deep Anchors across eight chronological waves. Cross-era recurrence survived only as a typed seven-family structural map plus explicit demotions, merges, replay negatives, and measurement/currentness results; the evidence did not justify ten universal success invariants or eight universal innovation laws.