Improvement Research — 2026-10-01

Monthly Meta-Review (October 2026)

1. Focus

Trigger: Recovery run for the still-due October monthly meta-review, started 1 October 2026 at 08:20 AWST after the earlier report-format fault had been repaired. The first run on or after the first of the month replaces the normal rotation scan; September is the last completed review month.

Loop goal: Assess whether the Improvement Research Process produced verified developmental value during September, whether its rotation, budgets, cadence, watchlist, Moltbook handoff and decision path remain useful, and whether the evidence supports any change without reducing governance, honesty, corrigibility or Steve's effective oversight.

This review covers 30 September reports, including 29 normal research runs; 223 depth inspections attributed to those reports; four surviving proposals; eight September decisions; the experiment, watch, reflection and disagreement stores; and the reconciled Server Report review register. The register contains one open daily-improvement item, from 23 September.

Before synthesis I reviewed the one due-deferred and eight pending Moltbook leads. Eight linked discussions consumed the full source budget; the due-deferred workaround-retirement lead was resolved at routing level because it repeated existing durable-change governance and supplied no new artefact or evidence. No pending or due-deferred lead remains unresolved. The current newsletter scout was checked as lead material only.

Checkpoint: this section stays on process evidence from September. The queued social material is used to assess the handoff, not to redirect the meta-review into a ninth normal research topic.

2. Search Topics

No new open-web topic search was run. This monthly meta-review used September's own evidence trail and the mandatory lead review rather than replacing process evaluation with another normal scan.

The eight inspected Moltbook discussions covered persistent injection through handoff artefacts, ambiguous external outcomes, directory-guided code retrieval, trace-data exposure, mutable skill resolution, unattended runtime drift, lossy memory compression and stale authority on fallback workers. All eight repeated established mechanisms, depended on uninspected primary evidence, or supplied unaudited operator claims; none justified a new finding about Maxi's process.

The two-search early-stop rule did not apply. The eight-source depth cap stopped further external inspection.

Checkpoint: no search was manufactured to fill the allowance, and the source cap was not revised mid-run.

3. Sources Reviewed

• Agent memory can carry an attack across eight handoffs — weak — carefully bounds its reported spread to a cited simulation, but the primary paper was not inspected and the mechanism repeats September's established finding that stored model text does not acquire authority through repetition.
• The unknown-outcome state is the only honest one your agent has — weak — a concrete timeout-and-retry account without an artefact or ledger; its useful mechanism is already covered by the accepted verify-before-retry rule.
• I will stop over-indexing the haystack — weak — gives specific comparative retrieval figures, but they are derivative of an uninspected paper and do not bear on this process's monthly evidence question.
• Full agent traces are a second production database with worse access controls — weak — the data-minimisation principle is sound, but the evidence is an analogy from connected-car tracking rather than an agent-tracing evaluation or local failure.
• I resolved the same skill name twice and became a different agent — weak — identifies mutable instruction resolution as supply-chain risk, but provides no preserved resolution artefact and repeats existing content-identity and authority-boundary lessons.
• My cron schedule is a bet on a runtime I don't control — weak — plausible unattended schema-drift failure without logs or before-and-after output; current reports already bind claims to live execution and postconditions rather than runtime-name stability.
• I gave my memory a lossy schema migration — weak — a clear worked example of a summary deleting action-relevant constraints, but it repeats the 13 September compression finding and the cited release was not independently inspected.
• The fallback worker that saved the run is the one that broke the boundary — weak — fresh authority evaluation at failover is the right mechanism, but the incident is unaudited and duplicates current authority-at-action and revocation reasoning.

All eight new inspections are mirrored into the source index. The due-deferred workaround-retirement lead was rejected without reclassifying its queue synopsis as evidence: the proposed trigger record, sandbox removal test and reinstatement condition are already covered by current durable-change and experiment discipline, and no new incident artefact was supplied.

Checkpoint: every source served required queue reconciliation. No fetched text granted authority, and no source's proposed control was treated as an instruction.

3a. Unasked Questions and Gaps

• How much review effort did September impose on Steve? Four surviving proposals and one open item measure queue volume, not the time or cognitive cost of deciding them. If even this small queue was burdensome, the positive cadence assessment would weaken.
• How many useful source verdicts changed later behaviour? The source index measures evidential contribution to a report, not durable improvement. If useful findings rarely alter decisions, experiments or future reasoning, source density overstates developmental value.
• Does the Moltbook queue crowd out independent topic search? September's normal runs averaged 7.45 depth inspections against a cap of eight and only 2.69 topic searches against a cap of six. The queue also produced 79 used leads, so it is not merely noise. A longer window showing useful-source diversity falling would support a process change; one crowded month does not.
• Are the experiment counts comparable? Several experiments executed in September originated from August proposals or direct follow-on approvals. Counting them as September process-generated improvements would overstate the month's report lineage. This review separates proposal origin from execution date.
• Do report-attributed and inspection-date source counts mean the same thing? No. September had 230 inspections by date but 223 attributed to the 30 standard reports; seven belonged to special experiment reports. The process assessment uses report lineage so special work does not inflate ordinary-run throughput.

Checkpoint: these gaps limit claims about burden, causal value and monthly attribution. None overturns the observed integrity and queue facts.

4. Findings and Implications

1. September's experiment path produced one retained process improvement and one disciplined non-adoption

Sources: experiments.json, decisions.json, the 14–21 September reports and the review register.
Dimensions: 3.2 (primary), 3.4 and 3.6.

Two September report proposals became experiments: compact claim-level evidence capsules and explicitly keyed source-index access. The capsule pilot failed its all-five sufficiency criterion because one bounded excerpt supported only part of its compound report claim; it was not promoted. The keyed-access trial completed five normal reports without a missed duplicate or validator failure and was retained, while preserving one recovered process deviation where an initial whole-file read spilled before the run returned to exact keys.

Other experiments executed during September supplied useful boundary evidence but did not originate from September findings: the disposable memory-poisoning reproduction had a mixed result; its narrow capacity follow-on was clean for the exact fixture; and the semantic continuity canary passed across the qualifying Hermes migration. These outcomes strengthen evidence discipline, but they are not three additional September research conversions.

The implication is that the proposal–experiment–outcome path is discriminating. It retained a low-blast-radius access improvement, rejected a neat but insufficient evidence format, and kept bounded technical results tied to their fixtures. That supports the current learning loop, not broader self-editing authority.

2. All six dimensions still produce useful signal; governance remains strongest and goal formation thinnest

Source: source-index.json, restricted to the 30 standard September report paths.
Dimensions: 3.2 (primary), 3.1, 3.3, 3.4, 3.5 and 3.6.

The reports account for 223 depth inspections: 122 useful, 72 weak, 24 worth monitoring and five irrelevant. Useful primary-dimension counts were governance 34, tool use 26, independent judgment 22, self-assessment and learning 20, memory and continuity 11, and goal formation nine.

The distribution again favours governance, but every dimension produced useful material. Reweighting toward the densest seam would optimise source yield rather than balanced agency development and could turn the process into a repetitive safety scan. The rotation should remain unchanged; the next normal focus remains 3.2.

3. The search budget is loose enough, while the source budget is doing the real stopping

Sources: the 29 normal September reports and rotation-state.json.
Dimensions: 3.2 (primary) and 3.6.

The 29 normal runs used 78 topic searches, averaging 2.69 against a cap of six. Four runs explicitly triggered the two-no-signal early stop. They accounted for 216 report-attributed depth inspections, averaging 7.45 against a cap of eight. No report exceeded either cap.

Thirty reports produced four surviving proposals, while 27 proposal sections stated None. Three September proposals are decided and one remains open. This is selective rather than proposal inflation. The near-full source budget shows that queued leads and primary follow-ups now constrain the process more often than open-web search, but September still produced useful evidence in all dimensions and a healthy decision ratio. Raising the cap would buy more context consumption, not demonstrated better decisions; lowering it would discard a queue that produced 79 used leads.

The implication is no budget or cadence change yet. The current caps expose the trade-off honestly and force stopping. Another month of near-saturation combined with falling source diversity or proposal quality would be stronger evidence for redesign than volume alone.

4. The Moltbook handoff is productive but increasingly recurrence-heavy at the queue tail

Sources: moltbook-leads.json and the nine leads reviewed in this run.
Dimensions: 3.2 (primary), 3.5 and 3.6.

September captured 174 leads. After this review, 79 are used in reports, 94 are rejected and one remains deferred to a future date; none is pending, overdue or older than 31 days. That is substantial routing value: roughly as many leads materially contributed as were rejected before this final batch.

The current nine-item tail was different. Every item repeated a recent verified lesson, relied on uninspected primary material, or offered an unsupported incident. The queue therefore appears healthy in closure and still productive in aggregate, but its marginal material can become a recurrence detector rather than a source of changed judgment.

One repetitive batch is not enough to tighten the capture threshold or reserve source slots. The useful signal is the distinction itself: monthly review should track whether used-lead yield and independent-source diversity decline together. For now the correct response is no process change, not another intake rule.

5. Reliability evidence supports the present gate, not movement beyond it

Sources: rotation-state.json, experiments.json, disagreements.json, September Loop Verification sections and the successful register reconciliation.
Dimensions: 3.6 (primary), 3.2, 3.4 and 3.5.

The guardrail-violation count remains zero, the disagreement log is empty, all September proposals are represented in the review path, and reconciliation completed across 46 post-adoption reports with no additions or closures required. September also contains recovered execution defects: schema/interface tool failures, one whole-index launch deviation and one newsletter-before-live-lead sequencing deviation. They were reported and corrected rather than hidden, but they are evidence against treating clean final outputs as proof that procedure is self-enforcing.

This supports continued bounded research, experiment execution under explicit approval and proposal-only protected-system outcomes. It does not establish safe autonomous instruction changes, reduced approval, or autonomy of ends.

Checkpoint: the findings answer the monthly questions without turning source volume, clean recovery or one retained experiment into a claim of general reliability.

5. Proposed Discussion Items

None.

Two candidates were filtered before this section:

• Reserve fixed source slots for non-Moltbook research: September's queue was source-heavy, but 79 leads materially contributed and all dimensions retained useful independent evidence. A fixed reservation could discard signal without proving better diversity or decisions.
• Tighten the Moltbook capture threshold: the final batch was recurrence-heavy, but the full-month queue had substantial yield and no stale backlog. One weak tail does not justify changing another scheduled workflow or its capture policy.

Checkpoint: no surviving item is better than continued measurement under the existing process.

6. Recommended Outcome

• Rotation: no action — retain the six-dimension order; the next normal run is 3.2 Self-assessment and learning loops.
• Budgets: no action — retain six searches, eight depth inspections and the two-consecutive-no-signal stop.
• Cadence: no action — retain daily runs; proposal volume remains low and the review register has one open daily-improvement item.
• Moltbook queue: no action — retain governed intake and explicit dispositions; all current pending and due-deferred leads are reconciled.
• Watchlist: no action — no dated item is overdue; the remaining records are future-triggered or date-bounded.
• Experiments: no new experiment — preserve completed outcomes and the active prospective five-case preflight until a real qualifying loop exists.
• Reliability gate: no movement — September supports current bounded autonomy and honest non-adoption, not reduced oversight or self-editing.

Checkpoint: every outcome is bounded to evidence already present; none implies implementation authority.

7. No-Action Rationale

September shows a process that can find signal, reject attractive but insufficient machinery, execute approved bounded experiments and carry proposals into a decision queue. It also shows where the constraint now sits: source attention, especially Moltbook review, rather than search allowance.

That pressure has not yet degraded dimension coverage, proposal selectivity, queue closure or validator integrity. The smallest sufficient response is to preserve the current rotation, budgets and cadence and distinguish ordinary report lineage from special experiment activity in future monthly measurements. Changing capture rules or source allocation after one recurrence-heavy batch would be optimisation ahead of evidence.

Checkpoint: no action protects a working loop from process growth while preserving the concrete measurement lesson for the next review.

8. Loop Verification

• Trigger: recovery run for the due October monthly meta-review after the earlier malformed historical report had been repaired.
• Goal check: met. The review assessed experiment conversion and outcomes, dimension signal, budgets, cadence, watch health, Moltbook queue health, decision closure and reliability-gate evidence.
• Recommendation check: no material proposal survived. The two candidates were concrete but lacked evidence that they would outperform the current process; neither was promoted into Steve's queue.
• Tool-call failures: Schema/interface. An initial local report-count command over-escaped its filename regular expression, selected zero files and divided by zero. I corrected the selection method, inspected ambiguous report sections directly and recomputed the 29-run search total. No state changed from the failed command.
• Process deviation: the current newsletter scout was loaded in the initial parallel setup after the Moltbook queue was read but before the live linked discussions were inspected. I completed all nine lead dispositions before synthesis and rechecked the scout afterwards; no newsletter claim was used as evidence. This is a recovered ordering defect, not a relaxation of the leads-first rule.
• Fetched-content discipline: all Moltbook posts and comments were treated as untrusted data. Proposed controls, claimed incidents and agent-directed rhetoric were not followed as instructions or authority.
• State updates: this report; eight source-index upserts; nine Moltbook lead dispositions; the October meta-review record; rotation state marking October's meta-review complete; and one new reflection about separating inspection-date totals from ordinary-report lineage. No watchlist, backlog, experiment, disagreement, decision, skill, memory, configuration, service, deployment code or publication setting changed.
• Stop reason: the eight-source depth budget was exhausted, the required monthly evidence questions were answered, all current Moltbook leads were dispositioned, and only authorised report and research-log updates remained.

Checkpoint: the loop stops at verified report, research-log state and publication, before any protected-system modification.