# Model-behavior Chronicles: editorial sources and limits

Review date: September 5, 2026. This note records corrections to public claims and source boundaries. It is not an experiment, live model test, product benchmark or independent certification of historical events. Original publication dates and URLs are preserved.

## Grok and X Premium — originally April 29, 2026

The [original source note](/research/grok-sentinel-public-source-note-2026-04-29.md) says public thread retrieval was unavailable and requests an exact concession artifact before external-proof use. It supplies no stable thread link, dated interface capture or archived April documentation. Its earlier proof-note label and victory framing overstate its evidence.

The corrected [Chronicle](/chronicles/grok-sentinel-that-had-to-kneel) treats the dispute as operator-reported. The exact chatbot response remains independently unverified. Even an authenticated response would not itself establish product behavior, a company acknowledgment, a logged incident or a policy correction. Both retained images are conceptual artwork, not screenshots or company statements. The note's later roleplay loss remains part of the historical record.

The [official X Premium help page](https://help.x.com/en/using-x/x-premium) is the original linked reference. Its current wording cannot reconstruct what an April reader saw. The documentation complaint remains unresolved without dated evidence.

## Coding-agent comparison — originally February 22, 2026

The [original public source note](/research/anatomy-coding-agents-public-2026-02-22.md) links an unpinned community collection and repeats the publisher's architectural interpretation. No exact revision, file-level sampling manifest or provenance assessment accompanies it. Only six products received detailed sections despite the original seven-system promise. The seventh sample is unresolved; the corrected title and preview no longer promise it.

The [corrected comparison](/chronicles/anatomy-of-ai-coding-agents) separates attributed historical descriptions from unknown vendor capabilities and design hypotheses. It withdraws categorical industry-wide absence claims, unpinned versions and prices, and causal conclusions from warning text. Role instructions alone do not establish enforced permissions. No controlled results establish that Greyforge's arrangement performs better.

Official correction context read on September 5, 2026:

- [Cursor Plan Mode](https://cursor.com/docs/agent/plan-mode) describes planning before code changes.
- [Cursor Subagents](https://cursor.com/docs/subagents) describes delegated assistants with separate contexts.

These descriptions limit timeless claims of absent capabilities. They do not establish February behavior, verify the old community material or measure effectiveness. This note preserves only brief paraphrases, not copies of vendor pages or prompt material.

## Adversarial Research Protocol — originally February 27, 2026

The [public preview](/chronicles/adversarial-research-protocol) now describes a proposed decision method. Its historical VoiceOps report lacks a linked decision record or controlled efficacy comparison. The example is hypothetical. Existing claims of universal adoption and demonstrated organizational benefit are qualified. The maintained edition labels later design aids separately; they are not recovered historical incidents. The public stop-condition sentence has been corrected to require a finite budget rather than imply indefinite evidence-driven extension.

## Autonomy retrospective — originally April 2, 2026

The [retrospective](/chronicles/autonomy-engineered-laziness-accommodated) begins in early February, so its former four-month duration was unsupported. Product descriptions, inventory and roadmap statements are attributed to the historical publisher account. Unsupported savings, financial-quality, current-status and unattended-operation claims are removed or qualified. No attached inventory manifest, controlled labor-reduction study or autonomous-uptime record verifies those outcomes here.

## Retention boundary

The two original public source notes remain unchanged as historical records; this note corrects their use as evidence. No missing source artifact has been invented, and no new model interaction or operational reproduction was performed. The remaining historical evidence gaps are explicit, rather than treated as resolved by the editorial review.