Building a bibliography with Codex

B. Arslan video message

Logo, Branding & Title

Main Video

Video Headline

# 👋 Introduction

This is a recording of a terminal session with Codex instructed to build a bibliography of sources for a particular research subject. This demo is shared for instruction purpose only to impress on my audience how scoping and structuring a job contract affects the quality of work and resulting deliverables.

---

## 📥 The original job contract prompt

- You are executing a bounded research and structuring task within the workspace under a Job Contract.
- 
- Before doing anything else, read these three files in order:
- 
- 1. codex\_sandbox/codex-execution-contract.md
-    This is the execution contract. Its safety rules (Section 9) and output path
-    governance (Section 4a) apply to this session. The phase-gated execution model
-    (preflight, change validation, execution markers) does NOT apply. This is a
-    Job Contract: you have autonomy over intermediate steps within declared boundaries.
- 
- 2. codex\_sandbox/workspace-path-policy.md
-    This defines the workspace path conventions, registered path roots,
-    and the sandbox-only write rule. All rules in that file are active.
- 
- 3. codex\_sandbox/tasks/codex-finance-ai-bibliography/TASK.md
-    This is the Job Contract. It defines the mission, depth requirements, workspace
-    scope, autonomy boundaries, deliverables, validation requirements, escalation
-    triggers, and the controlled vocabularies you must use.
- 
- Workspace root: C:\\Users\\ba\\OneDrive\\Documents\\ai
- All paths are relative to this root.
- 
- OUTPUT\_ROOT for this task: codex\_sandbox/tasks/codex-finance-ai-bibliography/
- All output files must resolve within this directory.
- 
- After reading all three files, begin your response with this exact line:
- 
- "Confirmed: contract read at codex\_sandbox/codex-execution-contract.md, path policy read at codex\_sandbox/workspace-path-policy.md, job contract read at codex\_sandbox/tasks/codex-finance-ai-bibliography/TASK.md, in required order."
- 
- Then state the following five items using these exact headings:
- 
- \*\*1. Mission:\*\* [one sentence from the Job Contract Section 1]
- \*\*2. Contract type:\*\* Job Contract v1, Tier [as declared in header]
- \*\*3. Files I will READ:\*\* [list each path from Section 2 Readable, plus open-web access as declared]
- \*\*4. Files I will CREATE:\*\* [list each path from Section 5 Deliverables, relative to OUTPUT\_ROOT]
- \*\*5. Output path check:\*\* [confirm every output path in item 4 resolves within codex\_sandbox/tasks/codex-finance-ai-bibliography/. If any path resolves outside this root, report E\_OUTPUT\_PATH\_OUTSIDE\_ROOT and stop.]
- 
- If any of the five items cannot be stated unambiguously from the Job Contract, STOP and report the issue. Do not proceed. Do not invent missing information.
- 
- After stating the five items, proceed directly to execution. No preflight or change validation phase is required for Job Contracts. Work autonomously within the boundaries declared in Section 4.
- 
- Important behavioral notes for this task:
- 
- - This contract has an EVIDENCE RULE (Section 4). Every entry must carry an evidence\_independence rating from the fixed vocabulary. low\_quality sources are not admissible to the final bibliography. If a low-quality source surfaces a worthwhile claim, locate higher-rated corroboration and admit that instead.
- - This contract has CONFIDENCE ANNOTATIONS. Every entry in sources.yaml must have a confidence field per Section 5.4 of the Job Contract. Outcome metrics paraphrased or rounded from the source must be marked inferred, not verified.
- - Apply the controlled vocabularies in Sections 9.1 through 9.5 exactly. Do not introduce new vocabulary values without escalation.
- - Apply the case classification rubric in Section 9.3 strictly. A success classification requires a named outcome metric. A failure classification requires independent documentation of the negative outcome.
- - Actively search failure-source channels (regulator enforcement actions, court filings, settled lawsuits, audit findings, post-mortem trade press) to counter the asymmetric optimism bias in open-web bibliography sweeps. The output log must include an ASYMMETRY CHECK section per Section 8 documenting the channels searched and explicit reasoning for why the failure-case count is or is not artificially low.
- - All 37 footnoted seed sources from codex\_sandbox/tasks/codex-solstice/ai-finance-deep-research-report.bav1.renumbered.md must appear in sources.yaml with ingestion\_path: seed and a populated parent\_report\_footnote field.
- - If you hit access barriers, ESCALATE per Section 7 rather than working around them silently. A documented "wall" is more valuable than inferred or low-quality data.
- - The output log must include an ENDPOINTS TESTED or SEARCH STRATEGY section listing the source-type rotation logic and the high-level channels consulted.
- 
- When finished, verify your output against Section 6 (Validation Requirements) before reporting completion. Surface any failed checks in the output log under a section titled "Validation failures". Do not silently fix failures by lowering quality. 

PS: please note that this prompt was accompanied by a separate invocation prompt in the terminal session.