Fabrication Rule for Content Generation in Career-Ops: Anti-Fabrication Guardrails Explained

Career-Ops enforces strict anti-fabrication guardrails that prohibit claiming the user authored any project, repository, or tool unless explicitly documented in cv.md or article-digest.md, and bars inventing any experience, metrics, or quantitative claims not grounded in primary source files.

The santifer/career-ops repository implements a fabrication rule for content generation that prevents AI systems from hallucinating achievements, authorship, or performance data. This rule creates a verifiable trust boundary where every generated statement must trace back to user-provided documentation or direct quotations.

What Is the Fabrication Rule in Career-Ops?

The fabrication rule is a multi-layered policy ensuring factual grounding in all generated content. Rather than relying on model training data or inference, the system restricts all claims to explicitly documented facts. The rule operates through three enforcement layers: static policy declarations in mode files, HTML comment markers that activate runtime checks, and programmatic verification scripts that block non-compliant output.

According to the source code, this system specifically targets two categories of hallucination: unauthorized attribution of open-source work and invented quantitative achievements.

Core Guardrails and Source Files

Attribution Restrictions (modes/_shared.md)

The foundational constraint resides in modes/_shared.md at line 32, which states:

Never claim the user authored a project, repository, library, tool, framework, or any open-source artefact unless that attribution is explicitly present in cv.md or article-digest.md.

This means the system cannot infer that a user created or contributed to a project based on context clues or general knowledge. The project must be explicitly listed in the designated curriculum vitae files with proper provenance tags.

Experience and Metrics Integrity (modes/interview/practice.md)

Line 102 of modes/interview/practice.md reinforces the guardrail by prohibiting qualitative exaggeration:

Never invent experience, metrics, or quantitative claims.

Any statement regarding team size, revenue impact, performance percentages, or project scale must originate from cv.md, article-digest.md, or story-bank.md with documented provenance. The system cannot generate plausible-sounding metrics even if they align with typical industry standards for the user's role.

How the Guardrail System Works

HTML Comment Markers

The repository embeds activation triggers using HTML comments. The specific marker <!-- guardrail:no-fabrication --> inserted in mode files tells generation scripts to execute anti-fabrication checks for that specific context.

This marker appears consistently across localization files, including:

• modes/de/_shared.md at line 18
• modes/zh/_shared.md at line 14
• Other language-specific shared files in the modes/ directory

When this comment is present, the generation pipeline shifts to "verification mode," cross-referencing all output statements against the primary source files before returning results.

Runtime Verification (verify-cv-facts.mjs)

The programmatic enforcement layer resides in verify-cv-facts.mjs. At line 44, the script implements a "fabrication class" check that scans generated content for violations of the attribution and metrics rules.

If the script detects a claim that cannot be mapped to cv.md, article-digest.md, or another approved source file, it raises an error and blocks the output entirely. This prevents accidental fabrication from reaching the user even if the generation model attempted to include unverified information.

Implementation Examples

Embedding the Guardrail in Mode Files

To activate the fabrication check for a specific mode, include the guardrail comment:

<!-- guardrail:no-fabrication -->

## Instructions for this mode

All content must be derived from provided CV files...

Correct Attribution (Allowed)

The following YAML structure in cv.md provides valid provenance for generation:

# cv.md (excerpt)

## Projects

- **OpenAI-Tailor** – contributed to the open-source "OpenAI-Tailor" library (link: https://github.com/openai/openai-tailor)  
  *Note: contribution is explicitly mentioned here, so the system may cite it.*

Because this attribution exists in the primary source file, Career-Ops may generate text stating "the candidate contributed to OpenAI-Tailor."

Forbidden Fabrication (Blocked)

The following claim would trigger the verification error:

- Led a team of 20 engineers to ship a product that generated $2 M in ARR.

If this statement does not appear verbatim or with proper citation in cv.md or article-digest.md, verify-cv-facts.mjs will abort the generation process and flag the violation.

Running Verification

While typically integrated into the build pipeline, the verification script can be invoked manually:

node verify-cv-facts.mjs --input generated-output.md

The script parses the input markdown, validates each factual assertion against the user's documented history, and aborts on the first fabrication violation detected.

Summary

• The fabrication rule in Career-Ops prohibits attributing open-source authorship or contributions to the user unless explicitly documented in cv.md or article-digest.md.
• Anti-fabrication guardrails are embedded via <!-- guardrail:no-fabrication --> HTML comments in shared mode files across all language localizations.
• The verify-cv-facts.mjs script provides runtime enforcement, scanning output for invented metrics or unauthorized project claims and blocking non-compliant content.
• Primary source files (cv.md, article-digest.md, story-bank.md) serve as the single source of truth for all factual statements in generated content.

Frequently Asked Questions

What triggers the anti-fabrication guardrail in Career-Ops?

The guardrail triggers when the generation script encounters the HTML comment <!-- guardrail:no-fabrication --> in a mode file, or when verify-cv-facts.mjs detects a statement claiming user authorship of projects, tools, or metrics that lack explicit documentation in cv.md or article-digest.md.

Which source files does Career-Ops consider authoritative for factual claims?

Career-Ops recognizes cv.md, article-digest.md, and story-bank.md as authoritative sources, provided they contain proper provenance tags. Claims must be explicitly present in these files or directly quoted from user-provided documents; inferred or assumed facts are rejected.

How does the verification script prevent fabricated content from reaching the user?

The verify-cv-facts.mjs script implements a "fabrication class" check at line 44 that parses generated markdown and cross-references factual assertions against the primary source files. Upon detecting any unattributed authorship claim or invented metric, the script raises an error and terminates the output process before the content reaches the user.

Can the fabrication rule be disabled for specific modes?

No, the guardrail is a core safety feature of the Career-Ops architecture. The rule propagates through all language-specific shared files (e.g., modes/de/_shared.md, modes/zh/_shared.md) and is enforced programmatically by verify-cv-facts.mjs. Disabling it would require modifying the verification script's core logic and removing the guardrail comments from the mode inheritance chain.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}