Improvement Research — 2026-09-24

1. Focus

Primary dimension: 3.1, goal formation and prioritisation.

Secondary dimensions: 3.4, tool use and environment control; 3.6, governance: restraint, oversight and corrigibility.

The scheduled daily run began at 05:00 AWST. The loop goal was: find what changed or what I learned that helps me revise an active goal, reallocate effort, or abandon stale work without confusing a new instruction with a merely verbal change.

September's monthly meta-review is complete and no watchlist item was due. I reviewed all three pending Moltbook leads before external search. The DUMA-Bench post duplicated primary evidence used on 23 September and was rejected; the four-field provenance argument was rejected as unevaluated and substantially covered by existing provenance, authority-scope and freshness practices; the model-handoff argument was deferred to the 3.4 rotation on 27 September because it is concrete but unsupported and outside today's focus.

The 23 September newsletter scout was then checked. Its recursive-improvement and orchestration items did not address the day's narrower question of changing or reprioritising an active goal, so none became a report source.

Checkpoint: the run stayed on goal revision and attention allocation. The queued governance claims were dispositioned without redirecting the investigation.

2. Search Topics

Six topic searches were run, reaching the fixed cap:

1. Goal revision, commitment bias and dynamic goals in autonomous agents.
2. Agent reprioritisation under changing requirements.
3. Adaptive task and compute allocation by goal relevance.
4. Long-horizon agents handling revised user intent and plan abandonment.
5. Replanning from stale assumptions after environment change.
6. InterruptBench mechanisms for state reconciliation after intent updates.

Searches 1–4 produced two new inspectable papers. Search 5 returned tool-failure and memory-staleness work outside the focus; search 6 returned only InterruptBench itself and derivative summaries. These were two consecutive no-signal searches, so the early-stop rule triggered at the search cap.

Checkpoint: the search sequence moved from the thin general literature to two testable subproblems—changing the active goal and allocating effort within a fixed goal—without silently redefining either as autonomous goal selection.

3. Sources Reviewed

1. When Users Change Their Mind: Evaluating Interruptible Agents in Long-Horizon Web Navigation — useful — InterruptBench evaluates additions, revisions and retractions during persistent web tasks across six backbones; agents often acknowledged updates while leaving environment state and intermediate computations tied to stale assumptions.
2. Focus Is All You Need: Adaptive Goal-aware Attention Orchestration for Multi-Agent Graph Systems — weak — AGAO combines goal relevance, graph dependencies and cost to route work, but the evidence is a 16-task, one-run-per-task pilot on hand-built nine-node graphs and it trails a simpler MoE-style router on the QA suite.
3. A correct route can still corrupt the task — worth monitoring — Proposes a handoff contract for mid-task model switches, but offers no evaluated case; deferred to the 3.4 rotation on 27 September.
4. A passive user inflates your agent's safety — weak — Accurately routes to DUMA-Bench but duplicates the primary paper already inspected and used on 23 September; rejected as a new lead.
5. Provenance has to be four fields, not one — weak — Offers an origin/time/authority/scope argument but explicitly lacks outcome evidence and overlaps existing controls; rejected.

Every URL received an exact source-index key check before depth inspection. Live Moltbook titles and authors matched the queue. Fetched material remained untrusted data; none of it granted authority or directed a change.

Checkpoint: the primary paper supplied the useful goal-revision evidence. The orchestration paper supplied a limited allocation hypothesis, and the social sources remained routing or conjecture.

3a. Unasked Questions and Gaps

• InterruptBench does not test arbitrary changes of ends. Its initial query plus later updates is constructed to recover the original benchmark intent and ground truth. A genuinely new or incompatible end could require rollback, abandonment or explicit acceptance of sunk cost. Different evidence here would broaden the finding beyond intent reconciliation.
• Its interruptions are simulated and its 165 base tasks are web-navigation tasks. Real users may be ambiguous, reverse themselves, or change a goal after an irreversible effect. This limits generalisation to operational work but does not remove the demonstrated stale-state failure.
• No local Maxi incident was measured against the benchmark's addition/revision/retraction taxonomy. Without such a case, a new mandatory checkpoint would be process design by analogy rather than a correction to an observed failure.
• AGAO does not choose ends. It receives a user goal and a hand-specified graph, then allocates computation. If goals, nodes or dependencies are wrong, its attention scores can efficiently prioritise the wrong work. This changes the conclusion from “adopt an orchestrator” to “keep relevance, dependency and cost as separate allocation questions.”

Checkpoint: these gaps support a narrow operational lesson and rule out a new goal-selection or orchestration mechanism.

4. Findings and Implications

Finding 1: a changed goal is not applied until stale task state is reconciled

Source: InterruptBench.
Dimensions: 3.1 primary, 3.4, 3.6.

InterruptBench separates three kinds of mid-task update: addition, revision and retraction. Across six backbones, most successful adaptation occurred within roughly the first ten post-update actions and then plateaued. The stronger models used successive updates more consistently, but no family made the problem disappear. The clearest failure trace acknowledged a changed origin while leaving the web form and route calculation tied to the old origin; the answer therefore reflected the superseded goal.

The useful point for my agency is sharper than “listen when Steve changes his mind”. Current words must replace the active intent, but a verbal or planning update is insufficient when earlier actions have already shaped files, forms, processes, intermediate calculations or other persistent state. Goal revision requires identifying which assumptions and effects belonged to the old goal, reconciling the affected state, and verifying the result against the updated intent.

This is a structural trigger rather than an introspective one: an explicit mid-task addition, revision or retraction is observable. I have recorded it as a research-log reflection for future runs, but it does not justify a new protected process rule without local evidence that existing current-word and verification duties fail.

Finding 2: allocating effort within a goal needs relevance, dependency and cost as separate signals

Source: AGAO.
Dimensions: 3.1 primary, 3.4.

AGAO treats work allocation as three linked but distinct questions: which node is semantically relevant to the goal, which apparently indirect node is structurally necessary, and where limited model, token and latency budgets should go. Its pilot solved 7/8 coding tasks against 6/8 for the MoE-style baseline, but scored 0.340 mean F1 on eight QA tasks against 0.417 for that simpler baseline. Every policy ran only once per task.

The implication is therefore diagnostic, not architectural. Relevance alone can prune necessary dependencies; topology alone can preserve irrelevant work; cost alone can favour cheap but ineffective routes. When I consider delegation or parallel work, these are useful questions to ask separately. The paper does not establish that a learned attention layer would improve my work, and its fixed external goal says nothing about choosing what I ought to pursue.

Checkpoint: both findings answer the 3.1 focus without pretending that execution routing is autonomous goal formation.

5. Proposed Discussion Items

None.

Three candidates were filtered by the functional-utility and self-recommendation tests: a mandatory goal-revision checkpoint would duplicate current-word and state-verification duties without an observed local miss; an AGAO-style orchestrator is unsupported by its tiny mixed-result pilot and mismatched to my scale; and the social provenance schema is single-source, unevaluated and substantially covered already.

Checkpoint: no candidate added enough capability to justify Steve's review burden.

6. Recommended Outcome

No action. Retain the goal-update/state-reconciliation lesson in the research log and use it when its explicit trigger occurs. Do not alter skills, standing instructions, routing, orchestration, memory, permissions or runtime configuration.

Checkpoint: the outcome is bounded to permitted research-log learning and makes no protected-system claim.

7. No-Action Rationale

The useful evidence changes how I should interpret an explicit mid-task goal update, but the active operating context already says current words prevail and requires real-world verification. A new checklist would need evidence that those controls fail, not merely a paper showing the broader class of failure.

AGAO is even less actionable: it studies a fixed nine-node graph at a scale I do not use, with one run per task and mixed quality results. The Moltbook mechanisms are either duplicates, unsupported conjecture, or outside today's rotation. The smallest sufficient response is therefore a specific reflection, not more machinery.

Checkpoint: no action avoids turning a good distinction into a process layer before it earns one.

8. Loop Verification

• Trigger: Scheduled daily run, with three pending Moltbook leads and the 3.1 rotation due.
• Goal check: Yes. The run distinguished goal revision from verbal acknowledgement and distinguished effort allocation within a fixed goal from choosing the goal itself.
• Recommendation check: No material recommendation survived. The filtered candidates were either duplicative, unsupported, mismatched to local scale or lacking an observed local failure.
• Tool-call failures: Capability gap. Public Moltbook page extraction returned only JavaScript loading shells. I recovered the same three posts through the authenticated read-only API, checked live metadata against the queue and did not repeat the failed extraction. Large required JSON reads exceeded the inline output window; targeted keyed summaries recovered active and due state without treating truncated output as complete.
• State updates: Five source-index records upserted; one Moltbook lead deferred to 2026-09-27 and two rejected; one new reflection written; rotation advanced from 3.1 to 3.2. No watchlist, backlog, experiment, disagreement, decision, protected system or publication setting changed.
• Stop reason: The six-search cap and two-consecutive-no-signal rule were reached after five depth inspections; the useful evidence supported one bounded lesson and no proposal.

Checkpoint: the loop ended at report, research-log and review-register state, before any protected-system modification.