# Research basis

## Evidence standard

The preset is based on an evidence-weighted collection rather than one viral complaint. Sources were ranked as:

1. OpenAI official model guidance and system card;
2. reproducible or trace-oriented `openai/codex` issues;
3. independent evaluations;
4. repeated community reports;
5. isolated anecdotes.

Community posts show failure shapes, not prevalence. An open GitHub issue proves that a report and any attached trace exist; it does not prove that the reported root cause is confirmed upstream. Provider, reasoning effort, Codex version, system prompt, skills, context length, and repository state often change at the same time.

Research snapshot last checked: **2026-08-30**.

## Strongest finding

The most credible risk is over-persistence interacting with elastic scope. GPT-5.6 Sol can add speculative confidence machinery, then treat its own artifacts as justification for migrations, recovery paths, reviews, and further validation. OpenAI's GPT-5.6 system card directionally corroborates a higher tendency than GPT-5.5 to take or attempt actions outside user intent in agentic-coding simulations, especially at high reasoning effort and with prompts that emphasize sustained persistence. OpenAI states that absolute rates remain low.

Representative sources, all last checked 2026-08-30:

- **Official system card:** [GPT-5.6 System Card](https://deploymentsafety.openai.com/gpt-5-6)
- **Official model guidance:** [GPT-5.6 prompting guidance](https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6)
- **User-reported Codex issue with mechanism/trace:** [#39059: self-reinforcing verification and governance layers](https://github.com/openai/codex/issues/39059)
- **User-reported Codex issue with workflow trace:** [#38375: out-of-scope reviewer loop](https://github.com/openai/codex/issues/38375)
- **User-reported Codex issue with long-task examples:** [#37278: goal preservation and proxy completion](https://github.com/openai/codex/issues/37278)
- **User-reported Codex issue with tool-path analysis:** [#32503: serial Code Mode calls and quota use](https://github.com/openai/codex/issues/32503)
- **Independent static-analysis evaluation:** [Sonar evaluation of Sol and Terra](https://www.sonarsource.com/blog/openai-gpt-5-6-sol-and-terra)

## Problem-to-rule mapping

| Observed pressure | Preset response |
|---|---|
| Scope expansion | Scope is not elastic; related findings do not become work unless they block acceptance |
| Invented governance | Persistent or defensive machinery needs an independent requirement |
| Review loop | Reviewers provide evidence but cannot authorize new scope |
| Proxy completion | Plans, tests, logs, and reports cannot substitute for the requested deliverable |
| Stale reasoning after correction | Upstream corrections invalidate dependent conclusions |
| Repeated tests and retries | Repeat only with new evidence and a new hypothesis |
| Premature or endless stopping | Explicit acceptance evidence; stop immediately after it passes |
| Serial PTC calls | Batch independent read-only SDK calls with `Promise.all` |
| Long technical narration | Lead with outcome, evidence, and limits in plain language |

## Why the prompt is short

OpenAI reports that leaner system prompts improved an internal coding-agent evaluation while reducing tokens and cost. Those numbers are directional rather than a guarantee, but they support removing duplicated process guidance and irrelevant examples. This project therefore encodes only rules tied to a measured failure class.

## Counter-evidence

- Sol also receives strong praise for difficult debugging, frontend work, and long-horizon execution.
- Terra has positive reports as a predictable everyday worker.
- Luna has strong cost/performance reports on narrow, checkable tasks.
- OpenAI's ordinary ChatGPT traffic simulation predicted less completion misrepresentation than GPT-5.5, which cautions against treating coding-harness anecdotes as universal model behavior.

The project does not claim that any model is categorically unreliable. It targets the interaction between model, long agentic trajectory, PTC, and prompt stack.

## Explicit exclusions

Prompting cannot repair incorrect context-window catalogs, compaction/resume loops, unsupported cache parameters, Responses-Lite/provider mismatches, hidden subagent routing, missing internal engines, backend capacity, or safety classifiers. Such failures belong to DSH, Codex, Provider, or model-service diagnostics.