Provenance Methodology

Provenance Methodology

Every behavioral claim you write MUST have a citation. No exceptions.

The Citation Rule

A "behavioral claim" is any assertion about what the target system does, how it responds, what data it accepts or produces, what errors it raises, what limits it enforces, or how it transitions between states.

Cite as you go. Do NOT batch citations at the end of your analysis. Every time you write a behavioral claim, the very next thing you write is the citation.

Citation Format

Inline HTML comment, immediately after the claim:

- Sessions expire after 30 minutes of inactivity
  <!-- cite: source=official-docs, ref=https://docs.example.com/sessions#timeout, confidence=confirmed, agent=doc-researcher, corroborated_by=runtime-observation -->

Required Fields

Optional Fields

Placement Rules

• The <!-- cite: --> comment MUST appear on the line immediately following the claim it supports, or on the same line after the claim text.
• If a single claim is supported by multiple independent sources, use a single citation with corroborated_by listing the additional sources.
• If a paragraph contains multiple claims, each claim gets its own citation. Break compound sentences into separate cited items.
• Block-level claims (tables, code blocks, decision trees) place the citation comment immediately after the closing block.

Source Types (Strongest to Weakest)

Agent Source Type Guide

Confidence Levels

confirmed

Two or more independent sources agree, OR a single runtime observation with reproducible steps.

Use when:

• Official docs state X AND source code confirms X
• Runtime observation shows X AND SDK client handles X
• Two independent community sources describe X consistently
• A single reproducible runtime observation (documented input, steps, output)

inferred

One authoritative source, no contradictions.

Use when:

• Official docs state X but no second source confirms it
• Source code clearly implements X but no docs mention it
• A single well-regarded community source describes X, consistent with known behaviors

assumed

Convention, pattern matching, or reasoning. No direct source.

Use when:

• Following a common convention (e.g., "REST API probably returns JSON") but no source confirms
• A pattern in one module is assumed to apply in another
• Evidence is partial and the claim is the most plausible interpretation
• Filling a spec gap where behavior must exist for the system to function

How to Determine Confidence

digraph confidence_determination {
    rankdir=TB;

    "Determine confidence level" [shape=doublecircle];
    "How many independent sources?" [shape=diamond];
    "Is the source authoritative?" [shape=diamond];
    "Any contradicting evidence?" [shape=diamond];
    "Use confirmed" [shape=box];
    "Use inferred" [shape=box];
    "Use assumed" [shape=box];
    "STOP: Record the contradiction" [shape=octagon, style=filled, fillcolor=red, fontcolor=white];

    "Determine confidence level" -> "How many independent sources?";
    "How many independent sources?" -> "Any contradicting evidence?" [label="2 or more"];
    "How many independent sources?" -> "Is the source authoritative?" [label="1"];
    "How many independent sources?" -> "Use assumed" [label="0"];
    "Is the source authoritative?" -> "Use inferred" [label="yes"];
    "Is the source authoritative?" -> "Use assumed" [label="no"];
    "Any contradicting evidence?" -> "STOP: Record the contradiction" [label="yes"];
    "Any contradicting evidence?" -> "Use confirmed" [label="no"];
}

Authoritative sources: official docs, source code, runtime observation, specific community content. Not authoritative: your own reasoning (always assumed unless corroborated).

Confidence Escalation

Confidence can be upgraded but never downgraded without new contradicting evidence.

When escalating, update the <!-- cite: --> comment with the new confidence level and add the confirming source to corroborated_by.

Handling Contradictions

digraph contradiction_handling {
    rankdir=TB;

    "New source contradicts existing claim" [shape=ellipse];
    "Record both claims with citations" [shape=box];
    "Does runtime observation contradict docs?" [shape=diamond];
    "Does source code contradict docs?" [shape=diamond];
    "Higher-ranked source wins" [shape=box];
    "Runtime observation wins (docs may be stale)" [shape=box];
    "Source code wins for implementation details" [shape=box];
    "Preserve all conflicts in citations for audit" [shape=box];

    "New source contradicts existing claim" -> "Record both claims with citations";
    "Record both claims with citations" -> "Does runtime observation contradict docs?";
    "Does runtime observation contradict docs?" -> "Runtime observation wins (docs may be stale)" [label="yes"];
    "Does runtime observation contradict docs?" -> "Does source code contradict docs?" [label="no"];
    "Does source code contradict docs?" -> "Source code wins for implementation details" [label="yes"];
    "Does source code contradict docs?" -> "Higher-ranked source wins" [label="no"];
    "Higher-ranked source wins" -> "Preserve all conflicts in citations for audit";
    "Runtime observation wins (docs may be stale)" -> "Preserve all conflicts in citations for audit";
    "Source code wins for implementation details" -> "Preserve all conflicts in citations for audit";
}

Common Mistakes

Citing everything as "confirmed" because you feel confident. confirmed requires 2+ independent sources. Your confidence in your own analysis is not a second source. One source = inferred. Period.

Using source=inferred for everything you cannot directly cite. source=inferred means the claim was derived by reasoning. If you read it in docs, source=official-docs. If you saw it in code, source=source-code. Source type = where the evidence came from. Confidence = how sure you are.

Omitting the ref field or using vague references. "ref=docs" is useless. "ref=source code" is useless. Use specific references:

• ref=https://docs.example.com/auth#tokens (specific URL)
• ref=workspace/raw/source/analysis/chunk-023.md:45 (file:line)
• ref=provenance/sessions/cli-explorer-abc123.jsonl:2026-03-01T10:30:00Z (session:timestamp)

Not recording assumed claims. Unrecorded assumptions look like confirmed facts to downstream agents and the implementer. Always cite assumed claims with confidence=assumed. List them in an Assumptions section. This is honesty, not weakness.

Batching citations at the end. You forget sources. You guess at confidence. Citations become decoration instead of evidence. Cite as you go.

Session Capture

After each subagent completes, the orchestrator:

1. Copies the session JSONL to workspace/provenance/sessions/{agent-name}-{agent-id}.jsonl
2. Records the agent run in workspace.json
3. Appends to the audit log at workspace/provenance/audit-log.md
4. Creates a git commit containing all output files, session log, and updated metadata

The session log contains every tool call, search result, and file reading that the agent performed. It is the deepest layer of provenance evidence.