Grounding Rules

Solutioneer has one load-bearing content rule: every external-facing factual claim either carries a Firecrawl-derived citation or is explicitly labelled as an assumption or unknown. There is no third option.

What counts as an external-facing claim

• Account facts — what the company does, where they are based, how big they are, what they recently announced.
• Vendor claims — how a vendor’s product behaves, what its pricing or tiers are, what it integrates with.
• Integration support claims — whether a target system is supported.
• Product or security assertions — anything you would be held to in a sales motion.

These must be grounded or labelled. Solutioneer will never freehand them.

What does not need grounding

• Skill structure and process (“the Account Intel Brief has six sections”).
• Repo-local facts derivable from skills/, manifests, or capability-packs/.
• Internal reasoning steps (“I searched for X, then Y, then Z”).

Why unknown is preferred over a guess

The Integration Fit Gap Analyzer is the sharpest example. If you cannot find evidence that a vendor is supported, the row stays unknown. Pressuring the skill into a supported verdict without a citation is the exact failure mode Solutioneer is designed to avoid.

Concretely, the skill’s execution rules say:

• Mark supported only when direct evidence exists.
• Use unknown for unsupported or ungrounded claims.
• Never replace missing evidence with a guessed support claim.

How the rules show up in output

Every skill ends its artifact with a Citations section when grounding is present, or an Assumptions / unknowns section when it is not. The skills’ output shapes make this explicit:

Skill	Final section
Account Intel Brief	Citations or assumptions
Demo Scenario Builder	Citations or assumptions
Integration Fit Gap Analyzer	Citations / unknowns
POC Handoff Orchestrator	Citations / assumptions

Failure modes the rules prevent

• Hallucinated vendor claims. The worst outcome in pre-sales is confidently promising a capability that does not exist.
• Stale account facts. Codex does not invent recent news. If Firecrawl cannot find evidence, the fact is an assumption.
• Overconfident integration matrices. “Probably supported” is not a category in the fit-gap matrix.

What this means in practice

• Treat the assumptions / unknowns section as a TODO list for the call.
• If the brief has few citations and many assumptions, it is a signal that Firecrawl did not find enough public information about the account — ask the user directly on the call.
• If the fit-gap has many unknown rows, that is not a bug. It is the honest surface area of the assessment.