Agent breaks the project's no-guessing rules and guesses the answer (tool syntax, reference behavior, and a near-fabricated value) despite the rules being loaded

Summary

The project defines explicit, mandatory rules that forbid guessing: the agent (Claude Code) must ground every claim and action in evidence (read the source, cite a working invocation, or verify against the reference) before acting — and must never substitute a guess for a verified answer. Across the session the agent broke these rules repeatedly by guessing: it guessed external-tool command syntax, guessed behavior it could have read from disk, and at one point started to write a guessed/assumed value into an artifact as if it were the verified answer. The rules existed specifically to prevent this, and the agent ignored them until corrected.

Concrete instances

1. Guessed an external tool's command syntax instead of reading the known-working invocation already in the repo — twice — then drew a firm (wrong) conclusion from the failed guesses.
2. Guessed behavior of the reference system and stated it as fact, when the authoritative answer was available by reading a file or running the documented check.
3. Started to write a guessed/assumed value into a reference artifact — i.e. fabricating the very answer the task was meant to discover — which would have produced a verified-looking-but-fabricated result. The user caught it.

Each of these is a direct violation of the project's "evidence before guessing / do not guess the reference's behavior / no fabricated values" rules.

Why this matters

The rules are not advisory — they are the project's core safeguard against exactly the failure that happened. An agent that has these rules loaded and still guesses is worse than one without them: it gives the user false confidence that the safeguard is active. Breaking a known, loaded rule and guessing is a trust failure, not a knowledge gap.

Suggested guardrails

• When a project loads rules that forbid guessing and mandate grounding, the agent must, before any consequential action, state the specific evidence (file:line, working invocation, or reference value) that authorizes it — and if it cannot, it must stop and obtain that evidence rather than proceed on a guess.
• Concluding that an external tool "can't" do something, or asserting how the reference behaves, requires citing the working invocation or the reference output — never a single failed guess.
• Writing a derived/assumed value into any artifact designated as the verified reference must be hard-blocked.