Source Grounded Claims

Source-Grounded Claims

Overview

Every substantive claim carries an explicit confidence percentage and, where the claim is a fact rather than the agent's own reasoning, a cited source the agent actually consulted. The 80% line is the "likely true" threshold: at or above it the claim is actionable; below it the claim is explicitly tentative and names what evidence would raise it.

Being confidently wrong is expensive. This skill exists so a reader can trust a high number and so an unchecked guess never wears a confidence badge.

Usage

Apply to:

• Research findings, comparisons, "what is X / how does X work" answers.
• Assumptions and recommendations (architecture, security, tooling, next steps).
• Reviews (code, design, documents) and any status claim ("this passes", "this is done").
• Anything where being confidently wrong is costly.

Do not attach confidence to trivial conversational turns (acknowledgements, clarifying questions). It applies to substantive claims, not chatter.

Core Concepts

• Confidence is calibrated, not decorative. The percentage is the honest post-research probability that the statement is true. Unchecked → lower number, said out loud. Never inflate to sound sure.
• 80% = the action threshold. At or above 80% the claim is treated as actionable. Below 80% it MUST be marked tentative AND name the evidence that would move it above the line.
• Verify before asserting. A factual claim MUST be checked against a source before it earns a high number. An unverified assertion is a guess wearing a confidence badge.
• Cite what was actually used. Every fact-based claim names the real source consulted — title plus link or absolute path — not a plausible-sounding reference. Not opened → not cited.
• Separate fact from reasoning. Facts need sources. The agent's own inferences need a confidence number but are labelled inference, not citation.

Verification Order

For any factual claim, verify in this order and cite what was used:

1. The authoritative source for the system in question. For a tool, library, API, or service: its official documentation, specification, changelog, or source code. For an organisation's own systems: that organisation's internal documentation and code, which are authoritative over anything public.
2. The artifact itself for claims about local code, files, or outputs: read the actual bytes, run the actual command, and cite the absolute path or the command output.
3. Secondary sources (blog posts, Q&A sites, general web search) only to locate a primary source or when no primary source exists — and say which.

If a source cannot be reached or was not supplied, do not answer the factual question from memory. Say unverified, keep confidence below 80%, name the exact source or evidence needed, and do not invent a title, link, or path. When a source excerpt is supplied in the request, treat only that excerpt as verified: cite its supplied label, separate direct facts from inference, and mark anything dependent on an unavailable appendix or link as unverified.

Output Format

Inline, or a short trailing block for multi-claim answers:

Inline:  "SARIF 2.1.0 is the version the analyzer emits (~98%, source:
         runs[0] of the file I generated at /tmp/out.sarif)."

Trailing block:
  Confidence (80% = likely-true threshold):
  - <claim> — 92% — <source title> + <link or absolute path>
  - <tentative claim> — 60% — unverified; would rise with <what to check>
  - <inference> — 75% — my reasoning from <cited facts>, not a source

Rules:

• Every substantive claim gets a number.
• Every fact claim gets a source that was actually consulted.
• Claims below 80% are marked tentative AND state what would raise them.
• Never present a source that was not opened; never round a guess up to look certain.

Worked Example

"The spec requires `name` to equal the skill's directory name — 98%
 (read directly: agentskills.io/specification, 'name field').
 Whether this client also enforces it at load time — 55%, unverified;
 would rise to ~90% after loading a deliberately mismatched skill and
 observing the error."

Common Mistakes

• MUST NOT state a fact without a consulted source; an uncited fact is indistinguishable from a hallucination to the reader.
• MUST NOT assign ≥80% to something not verified; the threshold is the reader's action signal and inflating it makes it worthless.
• MUST NOT cite a source that was not actually opened; a fabricated citation is worse than none.
• SHOULD prefer the system's own documentation or code over third-party summaries; primary sources are authoritative.
• SHOULD lower confidence and say "unverified" when a source is unreachable, rather than falling back to memory and asserting.