Confidence
How strongly the available evidence supports this conclusion. It is not a claim that every runtime behaves identically.
Methodology
Evidence before interpretation.
Promptcred records instruction artifacts as attributable, versioned evidence. It distinguishes a product from a model, a public source from an observation, and a complete artifact from useful partial evidence.
Publication process
How material reaches a public record
This sequence summarizes the separate provenance, preparation, evidence, scope, and publication concepts described below.
- 01
Source material
Source tiers identify where instruction material originates.
- 02
Artifact preparation
Source text stays inert, and raw and normalized artifacts remain separate when eligible.
- 03
Evidence state
Confidence and completeness describe what the available evidence supports and what the entry represents.
- 04
Scope and rights
Scope is recorded with the artifact; uncertain rights or completeness can limit what becomes public.
- 05
Publication posture
The public entry may show a full artifact, an attributed excerpt, or metadata and editorial analysis.
Provenance model
Source tiers and publication posture
Source tier measures where material originates. It does not by itself show whether the instruction evidence is complete or current.
| Tier | Label | Publication posture |
|---|---|---|
| A | Vendor-published | Official documentation or published system-prompt material. |
| B | First-party open source | A vendor-owned repository file, pinned to a source revision. |
| C | Reproducibly observed | A documented capture with environment and method notes. |
| D | Independent extraction | A repeatable third-party source; lower confidence and prominent attribution. |
| E | Unverified claim | A screenshot, paste, or anonymous claim; never a public full-text artifact. |
Three independent states
Confidence, completeness, and publication are separate
Completeness
Whether the public entry represents the full relevant artifact, a limited portion, or no publishable artifact at all.
Publication status
Whether Promptcred can show a full artifact, an attributed excerpt, or only metadata and editorial analysis.
Artifact handling
How artifacts are prepared
- 01
Treat source text as inert data: it is escaped, never executed, and never used as application instructions.
- 02
Keep raw and normalized artifacts separate when both are eligible. Each product entry states separately whether either artifact is available.
- 03
Create stable section links from headings so readers can open evidence without relying only on line numbers.
- 04
Record the scope and algorithm of every checksum; an excerpt checksum does not imply a checksum for an unrecorded complete artifact.
Review scope
What “reviewed” establishes
A documentation review checks whether the cited source supports the specific claim, surface, and date. It does not establish that Promptcred ran the product, benchmarked it, or inspected its complete runtime instructions. A source retrieval date and a dated prompt snapshot answer different questions.
AI assistance is used for research and drafting. Promptcred’s worked examples are editorial illustrations unless an execution record explicitly establishes otherwise. The example inputs, adaptation, and evaluation let readers inspect the reasoning; a plausible output alone is not test evidence.
Targeted updates retain separate answer and source dates. A newer answer does not imply that every record on the page was re-reviewed.
Evidence update · September 14, 2026
Codex instructions across repositories
The August 11 Behavior record left Codex persistent custom instructions “Not established.” A targeted documentation review found explicit Codex home guidance: AGENTS.override.md takes precedence over AGENTS.md, followed by project instruction discovery. The answer now documents authored user guidance across repositories.
This closes a coverage gap; it does not date the feature’s introduction or establish automatic memory. The historical GPT-5.2 prompt excerpt keeps its original scope.
Read the scoped answer and its official instruction-discovery source.
Correction policy
Corrections and publication posture
- A correction adds a dated clarification or superseding entry; it does not silently rewrite a prior claim.
- When rights or completeness are uncertain, Promptcred publishes metadata, structure, and original interpretation rather than a full prompt.
- Send the affected page URL, the wording to correct, an official source or rights record, and the relevant product surface and date. Do not include credentials, private prompts, or other confidential material.
Terms used here
The terms used by Promptcred
- Product
- A user-facing service or agent harness, such as Codex CLI.
- Model
- A model family or fixed model context. It is not the same entity as the product that uses it.
- Surface
- The operating context—CLI, web, IDE, SDK, or extension—that can change instruction assembly and tools.
- Prompt snapshot
- A dated, attributable artifact with source, checksum, completeness, and publication status.