R1 · truth engine

Truth Engine.

Five anti-deception guards for the failure modes humans miss: unsupported completion claims, path corruption, destructive post-merge drops, paper-thin audits, and suspiciously small delivery deltas.

What it is

The Truth Engine is not a vibe checker. It is a narrow set of deterministic refusal conditions aimed at the exact places where agentic systems tend to lie with high confidence: saying “done” without evidence, silently corrupting files, losing tree state in merge operations, waving through shallow audits, and claiming a large task was completed with implausibly little code.

Five guards. One refusal boundary.

Each card starts with a weak claim. Running the audit walks the guards in order and replaces charm with a concrete rejection tied to missing or inconsistent evidence.

audit idle
Boundary verdict

The audit stays narrow on purpose. It does not debate tone. It checks whether the run can cite proof, preserve files, survive merge, defend the audit posture, and justify the delivery delta.

Why these five

Each guard maps to a concrete recurring failure mode. `file_line_required` stops mission-complete prose that outruns the diff. `path_marker_corruption` catches broken writes. `post_merge_tree_check` spots destructive branch loss. `completionist_skeptic` reads like due diligence instead of reassurance. `delivery_ratio` forces a written explanation when the estimate and delivery diverge sharply.

What the operator gets

The point is bounded skepticism. The operator does not get a vague warning or a “double check this” nudge. The operator gets a refusal with a narrow reason, the evidence types that were missing, and the exact class of failure that fired.