Agent-Cite

Cite it or it's opinion. Evidence enforcement for AI agent workflows — every claim, every source, zero uncited inference.

---

The Problem

AI agents hallucinate. They state facts that don't exist, cite numbers they invented, and claim "not found" without actually searching. The output looks confident. The citations look real. But when you check — the file doesn't exist, the number has no source, and the "unfound" thing is sitting right there.

You can't distinguish a real claim from a hallucinated one unless every claim cites its source.

Agent-Cite enforces that standard. Not as a suggestion — as a requirement.

---

Without vs. With Agent-Cite

What Claude/Codex Does WITHOUT Agent-Cite

Behavior	Result
"I'll try to cite sources"	Best effort — cites some, silently skips others
No systematic scan	Misses violations without knowing it missed them
Can't verify its own citations	Citations may point to files that don't exist
No structured report	Just prose with maybe some references
No pass/fail verdict	No gate — just a suggestion
Next session, forgets the standard	Back to zero

What Claude/Codex Does WITH Agent-Cite

Behavior	Result
Systematic scan of every statement	Categorized as factual claim, number, or absence claim
Named violations for every miss	UNCITED_INFERENCE, UNVERIFIED_NUMBER, UNSUPPORTED_ABSENCE
Source verification	Checks cited file paths actually exist on disk
Structured report	Summary table, violation list, line numbers, suggested fixes
COMPLIANT / NON_COMPLIANT	Binary verdict — not a suggestion
/cite-fix reads the actual codebase	Proposes real file:line citations, not guesses
Configurable protocol	evidence-protocol.yaml persists rules across sessions
Project-wide scoring	/cite-report gives evidence coverage like test coverage

The Analogy

Domain	Without Tool	With Tool
Testing	"Please write tests"	Jest/Pytest — runs them, reports coverage, blocks on failure
Linting	"Please format nicely"	ESLint/Prettier — enforces rules, auto-fixes, CI gate
Security	"Please write secure code"	Snyk/Semgrep — scans, names vulnerabilities, blocks merge
Evidence	"Please cite sources"	Agent-Cite — scans, names violations, blocks on uncited inference

Without Agent-Cite, citation is a request. With it, citation is a requirement.

---

Commands

Command	When	What It Does
/cite-audit	After AI generates output	Scan for uncited claims, return COMPLIANT or NON_COMPLIANT
/cite-fix	After violations found	Search codebase for real sources, propose accurate citations
/cite-report	Periodic health check	Project-wide evidence coverage score and worst offenders

Example

/cite-audit @docs/report.md --strict

Output:

## Evidence Audit Report
### Verdict: NON_COMPLIANT

| Category | Total | Cited | Violations |
|----------|-------|-------|------------|
| Factual claims | 23 | 15 | 8 |
| Numbers/stats | 7 | 2 | 5 |
| Absence claims | 3 | 0 | 3 |

Fix with: /cite-fix @docs/report.md

/cite-fix @docs/report.md

Output:

| # | Line | Violation | Proposed Fix | Confidence |
|---|------|-----------|-------------|------------|
| 1 | L42 | UNCITED_INFERENCE | [source: src/config.py:15] | HIGH |
| 2 | L67 | UNVERIFIED_NUMBER | [derived: tests/bench.py:23] | MEDIUM |
| 3 | L89 | FALSE_ABSENCE | tests/test_module.py EXISTS (12 tests) | CRITICAL |

---

Violation Types

Type	Severity	What It Catches
UNCITED_INFERENCE	error	"The system uses Redis" — WHERE does it say that?
UNVERIFIED_NUMBER	warning	"500 req/sec" — WHERE is that number from?
UNSUPPORTED_ABSENCE	warning	"No tests exist" — DID you actually look?
BROKEN_CITATION	error	[source: data/old.json] — that file doesn't exist
FALSE_ABSENCE	critical	"No tests exist" — but tests/test_module.py has 12 tests
UNVERIFIABLE	human decision	No source found anywhere — the claim may be fabricated

---

When to Use / When NOT to Use

Use Agent-Cite on:

• AI-generated reports and analyses
• AI-generated documentation and READMEs
• AI-generated recommendations and summaries
• Data classification outputs
• PR descriptions with factual claims
• Any AI output that says "X is true"

Do NOT use Agent-Cite on:

• Pure code (functions, classes, modules)
• Formatting or refactoring changes
• Boilerplate and scaffolding
• Creative writing

The rule: If the output asserts facts, cite it. If it just does something, don't.

---

Installation

# Clone
git clone https://github.com/saisumantatgit/Agent-Cite.git

# Install into your project (auto-detects your CLI)
cd your-project/
bash /path/to/Agent-Cite/install.sh

Or for Claude Code, install as a plugin:

cp -r Agent-Cite/ ~/.claude/plugins/agent-cite/

What Gets Installed

CLI Tool	What Gets Installed
Claude Code	.claude/commands/*.md + agents + skills + hook
Codex	Appends to AGENTS.md
Cursor	.cursor/rules/cite.md
Aider	Appends to .aider.conf.yml
Generic	Raw prompt files

Plus: evidence-protocol.yaml (configurable rules) and reference docs.

---

Configuration

Create evidence-protocol.yaml in your project root (a template is installed automatically):

severities:
  UNCITED_INFERENCE: error      # Always fails
  UNVERIFIED_NUMBER: warning    # Fails only with --strict
  UNSUPPORTED_ABSENCE: warning
  BROKEN_CITATION: error

include:
  - "docs/**/*.md"
  - "data/**/*.json"

exclude:
  - "CHANGELOG.md"
  - "node_modules/**"

thresholds:
  healthy: 90      # ≥90% = HEALTHY
  needs_work: 60   # 60-89% = NEEDS_WORK
  critical: 0      # <60% = CRITICAL

---

Platform Support

Platform	Config	Notes
Claude Code	.claude-plugin/plugin.json	Full support (agents + commands + skills)
Codex	AGENTS.md	Prompt-based
Cursor	.cursor/rules/cite.md	Rules-based
Aider	.aider.conf.yml	Config-based
Generic	prompts/*.md	Paste into any LLM

---

Part of the Agent Suite

Tool	What It Does	Tagline
Agent-PROVE	Makes agents think before they act	"Prove it or it fails."
Agent-Trace	Makes agents see before they edit	"See the ripple effect before it happens."
Agent-Scribe	Makes agents remember what they learned	"Nothing is lost."
Agent-Cite	Makes agents prove what they claim	"Cite it or it's opinion."
Agent-Drift	Intent drift detection and prevention	"Stay on target."
Agent-Litmus	Test quality validation	"Tests that actually test."

PROVE validates your thinking. Trace maps your blast radius. Scribe records your decisions. Cite enforces your evidence. Together: think rigorously, edit safely, remember everything, prove every claim.

Note: Agent-PROVE includes basic evidence auditing (/audit command). Agent-Cite is the standalone deep dive — citation fix, project-wide reporting, configurable protocol. If you already use PROVE, Agent-Cite adds the remediation and reporting PROVE doesn't have.

---

Origin

Built from the evidence protocol methodology that powered the Agent-PROVE project — where zero uncited inference was enforced across 1,086 regulatory data points. The protocol caught fabricated numbers, false absence claims, and citations to files that no longer existed. Agent-Cite generalizes that discipline for any codebase.

The standard is simple: if an AI says it, the AI must show where it learned it.

---

Contributing

See CONTRIBUTING.md for:

• Adding citation patterns
• Adding CLI adapters
• Extending violation types
• Modifying the report

---

License

MIT