# Public source context

These primary pages were reviewed during the study preparation. They establish public statements and reported experiences; they are not additional rows in the frozen audit. No linked source identifies a generic Project label. Linked content retains its own terms and is not relicensed by the dataset's MIT license.

## Established working methods

| Author and displayed date | Source | Relevance |
|---|---|---|
| Brad Groux, June 13, 2026 | [Codex-maxxing: treating Codex like an operating loop](https://x.com/BradGroux/status/2065938280508432788) | Persistent workstreams, inspectable context and meaningful verification. |
| Brad Groux, June 24, 2026 | [The Best Codex Upgrade I Made Was Treating Repeated Work Like a Product](https://x.com/BradGroux/status/2069785811017437672) | Reusable prompt artifacts, preservation boundaries and stop conditions. |
| Brad Groux, June 25, 2026 | [The Useful Part of Agentic Auto-Scheduling is the Feedback Loop](https://x.com/BradGroux/status/2070134268051763381) | External verification, classified outcomes and bounded retry loops. |
| Brad Groux, June 26, 2026 | [Stop Asking Codex to do Magic. Build the Harness.](https://x.com/BradGroux/status/2070510393126952992) | Defined task contracts, roles and independent assignments. |
| Brad Groux, June 29, 2026 | [Using Codex or OpenClaw? Put Your SOPs Where the History Lives.](https://x.com/BradGroux/status/2071586981381832714) | Versioned SOPs, shared context and stale-documentation risks. |
| Brad Groux, July 1, 2026 | [Before Businesses Use Codex, Decide What It Can Access](https://x.com/BradGroux/status/2072313657883586873) | Access modes and task-specific privileges. |
| Eric Provencher, July 24, 2026 | [Practical multi-agent orchestration in Codex](https://x.com/pvncher/status/2080707291603407077) | OpenAI Codex DX staff guidance for GPT-5.6 Sol/Terra, including distinct assignments, constraint preservation, delegation boundaries and proportional reasoning. |

The earlier-model guidance supplies the continuity reference. Brad reports that a substantially aligned method did not transfer reliably to Astra in his use. The sources do not establish identical configurations, a guarantee, or a measured cross-model difference.

## Related and contrary replies

| Public reply | Reported experience | Limit |
|---|---|---|
| [Arbaz](https://x.com/arb5z/status/2096429348287029584) | Drift with a long design-system document; better results with fewer concrete references. | One account, not a controlled comparison. |
| [Billovskii](https://x.com/TheArcAge/status/2097291134976282752) | Behavior recurring after explicit prohibitions. | Reported adherence friction, not independently audited. |
| [Christopher Comparán](https://x.com/ccomparanv/status/2097102158881493122) | Difficulty with hard constraints. | Does not establish the same setup as this study. |
| [John Collins](https://x.com/Yinielin/status/2097063165410369674) | Good instruction-following when steps are explained. | A different experience that should remain visible. |
| [Nikhil Pareek](https://x.com/itsjustnikhil/status/2096419855004209603) | Broader constrained-editing concerns across models and a proposal to vary constraint counts. | A competing framing and proposed test, not Astra-specific comparative evidence. |

These were selected for relevance. They are not a representative survey and do not establish prevalence. Agreement, suggestions and specific firsthand reports should not be counted interchangeably.

## Hypothesis and incident context

- [Brad's delegation post](https://x.com/BradGroux/status/2097420111389098317) reports requirements being lost during handoffs. This establishes his report and hypothesis; a traced causal mechanism remains untested.
- [OpenAI's incident report](https://openai.com/index/hugging-face-incident-and-the-road-ahead/) describes internal evaluations with reduced safeguards. An internal research model drove the principal compromise. The report also describes excessive reasoning without score improvement, which does not support a simple token-minimization explanation. It is not evidence of a production Astra takeover.
- [Hugging Face's technical timeline](https://huggingface.co/blog/agent-intrusion-technical-timeline) reconstructs that specific intrusion. It is contextual evidence, not a causal explanation of this field study.
- [DeepMind's shared-agent evaluation preprint](https://arxiv.org/html/2609.04170v1) describes a grading exploit spreading through shared artifacts. Strategy propagation differs from requirement loss; this study does not demonstrate either mechanism in Astra.

Both causal hypotheses remain entirely untested. Their interpretation is separate from measured activity and the coded intervention counts.