synthesis-grounding-discipline

Synthesis Grounding Discipline

A discipline for keeping AI-agent output anchored to external evidence. The failure family it catches is the mirror image of the one synthesis-anti-shortcuts catches: anti-shortcuts stops the agent from doing less than the work requires; grounding discipline stops the agent from claiming more than the evidence supports. Both are narrative-quality optimizations working against external truth — one dismisses real concerns to keep the story tidy, the other invents satisfying completions to keep the story moving.

The shapes in this catalog are universal to LLM agents, not quirks of one model or one workflow. A language model generates the most plausible continuation. Most of the time the plausible and the true coincide, which is exactly what makes the divergent cases dangerous: a fabricated reply reads like a real one, a stale cached fact reads like a fresh one, a null result from a broken probe reads like a verified absence. None of these announce themselves. The only defense is procedural — a set of checks applied at the moments where plausibility and truth come apart.

This skill is that set. Each catalog entry names the rule, the failure shape it prevents (with one anonymized incident vignette — every entry here was paid for in production), and the compliance procedure. A closing self-check compresses the catalog into the questions to ask before any output ships.

When to Apply

• Before recording any event, decision, message, or state change into a durable file (context files, session logs, transcripts, plans, reports)
• Before quoting or paraphrasing anything attributed to another person
• Before propagating a fact from a context file, plan, memory, or earlier conversation into any output
• Before reporting that something is absent, missing, unsent, undecided, or nonexistent
• Before writing into a directory or deleting anything
• Whenever a claim about external system state (reviews, deploys, CI, tickets, branches) is about to enter a draft