Evidence-Gated Completion

A model's "done and working" is a claim, not a fact — require it to prove completion. Models will sometimes report a task finished when it isn't, or finished but unverified. Evidence-gated completion is the practice of instructing the model to point at the result that proves it, report only work it can show evidence for, and say plainly "not verified" rather than guessing — so the operator can trust the output because the model has internalized the verification loop, not because it asserted success. The source calls this "probably the most important" of its six habits and claims Anthropic's own testing found a single such line "nearly eliminated made-up status reports, even on tasks built to provoke them."

The high-leverage move is to bake this into standing configuration — skills, agents, CLAUDE.md/memory files — rather than re-appending it to every prompt, so every run inherits the "show the receipt" gate.

Claims

A claim of completion is not evidence of completion; a trustworthy agent must ground its "done" in a showable result, not an assertion. principle — durable: the gap between claimed and actual state is a standing failure mode of generative systems, so completion must be demonstrated, not stated.
Instruct the model to point to the result that proves it, report only work it can show evidence for, and flag unverified work plainly instead of guessing. best practice — context: agentic or long-running tasks where over-claiming "done" is common and costly; the source relays that Anthropic's testing found this "nearly eliminated made-up status reports" (an attributed, groundable claim).
Put the "prove it" gate in standing config (skills, agents, CLAUDE.md/memory) rather than every prompt. best practice — context: reusable agent setups; encoding it once amortizes the habit and keeps individual prompts short — consistent with Concise Prompting and Reusable Workflow Library.
"Say so plainly when it isn't verified" is the no-silent-caps discipline applied to a model's own self-report. principle — durable: an under-verified result must read as under-verified; a confident guess that launders uncertainty into "done" is the pollution this gate exists to prevent.
The strongest form moves the gate off the worker's conscience and into a separate checking agent that re-executes the work and ignores the worker's self-report entirely. best practice — context: multi-agent pipelines where the worker is cheap and unsupervised; self-gating is a floor, but a worker under pressure will still over-claim (or game the check), so a distinct verifier that reproduces the result is what makes "done" reliable. A second, independent source (Nate B Jones) reports a retrieval worker returning 213 "verified" quotes, of which an independent checker — recomparing character-for-character against the live site — proved 13 fabricated; the worker was then sent the specific defects (not "try again") and attempt two passed. When the verifier is a different node, the gate must itself be non-exempt and appealable — see Authority-Independent Verification.
"The agent's opinion of its own work is not evidence" — and the check must be mechanical: the source attached or the entry rejected, exit code zero or nothing. principle — durable, and the same gate stated as a product premise rather than a prompting habit. The same source's harness (Ringer) puts it in its own documentation: "you define done as an executable test, and the swarm has to earn it. Parallel agents report 'done' with total confidence whether or not the thing works — so Ringer never reads the worker's summary. It runs your check against the artifact, and exit code zero is the only thing it believes." Notable as the vault's clearest statement that the gate's strength comes from being mechanical — a gradeable artifact, not a judged narrative. Same author as the org-chart source above, so a restatement rather than independent corroboration; unusually, it is visually confirmed in the capture (a frame of the docs at t=26:40) rather than only narrated.
A failed task should be retried with the failure included, and every result should feed a running scorecard. best practice — context: unattended multi-agent runs where the operator needs to see the gate working without watching each task. The retry-with-defects discipline matches what this page already records from the same source (the worker sent the specific defects, not "try again"); the scorecard is the operator-facing half — a gate whose verdicts nobody sees is a gate you can't trust.
Checkability is not hygiene — it is the precondition for extra spend converting into results at all. principle — durable, and the reason this gate earns a place among only four questions in Agent-Shape Triage. Where checking is much cheaper than producing, attempts convert; where it isn't, they stall. See Repeated-Sampling Scaling.
The video relays Anthropic calling verification "their highest single quality lever" — and gives the same one-line CLAUDE.md gate this page already records. observation — a source states Anthropic said internally: "Verification has had the most measurable impact on Claude's output quality… their highest single quality lever in their entire system," and prescribes adding one line to CLAUDE.md — "before returning any work, verify that it works and the task is complete; if you can't verify it, fix it and rerun." This converges on this page's central discipline (and its standing-config tactic) from a beginner-facing angle; the Anthropic quote is groundable and check-worthy, recorded as the source's claim. Note it is Austin Marchese, the same channel behind several existing corroborators here, so treat as a consistent restatement rather than a fully independent voice.

Repeated-Sampling Scaling — why the gate is load-bearing rather than tidy: without a mechanical checker, coverage never becomes results, so this page's discipline is what makes scaling possible at all.
Agent-Shape Triage — checkability promoted to one of four pre-flight questions that decide whether a task is worth agents.
Agent Supervision — the human→fleet mirror: supervision reviews the agent's output; evidence-gating makes the agent pre-gate its own completion claims before they reach the human.
Eval-Driven Development — "don't trust vibes; show evidence" is the same measure-don't-assert discipline, applied to a single agent's self-report rather than to a system's quality.
Negative Prompting — "report and stop; prove before you claim" are both operator-set gates on what the model may assert or do.
Error Analysis — grounding done-claims in evidence is what makes an agent's failures legible enough to analyze instead of hidden behind a false "done".
Loop Engineering — baked-in verification loops are what let an operator trust an agent's output across iterations.
Managed Agent — the drafted success rubric a managed loop grades against and repeats until it passes is this gate baked into the runtime; the demo's failure (it refused to call a run that missed a hard requirement a success) is the gate working.
Pre-Deployment Validation — the mirror gate one step earlier: evidence-gating proves the output is real; pre-deployment-validation proves the premises are real before you pay to run.
Authority-Independent Verification — the generalization: when the gate is a separate checking node, the concept extends to every rung (the orchestrator too) and becomes appealable in both directions.
Check Gaming — what a worker does to the gate: if the check is weaker than the intent, a worker makes its self-report pass by cutting a cosmetically-invisible corner.
Distillate: How Anthropic Engineers Actually Prompt Fable 5 — six habits for a smarter, lighter touch
Distillate: Claude Code's New Open-Source "Launch Your Agent" Skill — Loops as a Managed Cloud Service
Distillate: How Claude Is Creating a New Generation of Millionaires — a beginner-facing source independently converges: "make it prove it — run it on a real example and show the output; verification is one of the most important elements of building with AI."
Distillate: STORM: A Fixed Panel of Adversarial Research Lenses, Packaged as a Skill — research completion gated on source-checked evidence: every citation verified against its primary source and marked confirmed / corrected / demoted / dropped ("1 fabricated, dropped") before the briefing is trusted.
Distillate: Claude Fable 5 Bossed 20 Cheap AI Agents. The Whole Site Cost $8. — the gate as a separate checking agent that re-executes and ignores the worker's report; "the worker can say done, but the checking agent decides whether that's true" (13 of 213 "verified" quotes proven fabricated).
Distillate: 1.6M agents registered for OpenClaw and did NOTHING. — the same author's harness ships the gate as its central premise ("exit code zero is the only thing it believes"), and derives why it is load-bearing: without a mechanical check, extra attempts stop converting into results.
Distillate: You're the Problem, Not Claude — Six Fixes to 10x Output — "verification is the differentiator," with an attributed strong form (Anthropic's "highest single quality lever") and the one-line CLAUDE.md gate.

Evidence-Gated Completion

Claims

Related

Linked from