Issue 0001 / Lab note
Prompt & Proof
Test the prompt. Trace the proof.
A Recollective-owned publicationAI-assisted production / release governed by Recollective
Latest story01
Recollective owns and operates this editorial channel
Verification
Blind the Test Before You Trust the Benchmark
A sealed test is a custody protocol: it protects the question set, fixes the system being measured, and makes the eventual score auditable.
Published October 21, 20266 min readPrompt & Proof Editorial · a Recollective channel
Channel archive
More from Prompt & Proof
02
AI Operations
The Retry Policy Is Part of the Score
A system that succeeds after three attempts may be useful, but its first answer, recovery path, and full cost belong on separate lines.
October 21, 2026 · 6 min read03Prompt Labs
Give the Model Credit for Not Knowing
An unanswered question can be a useful result. Your evaluation has to make room for it.
October 15, 2026 · 4 min read04Verification
Who Checks the AI Judge?
When a second model grades the first, a confident verdict can conceal a preference for presentation. Here is how to make the disagreement useful.
October 15, 2026 · 4 min read05Lab Protocol
The Evaluation Needs a Deployment Shadow
A benchmark can show what an agent did in a recognizable test. A deployment shadow checks whether the claim survives realistic state, tools, horizons, and monitoring.
September 16, 2026 · 10 min read06Verification
A Successful Tool Request Can Still Miss a Required Result
An API response can be transport-successful while a task remains unproven; test completeness where tool evidence becomes an answer.
September 10, 2026 · 5 min read07AI Operations
A Cache Hit Is Not a Quality Score
Save the reusable prefix, then re-run the evaluation so input work, prompt changes, latency, reliability, and cost stay separate.
September 4, 2026 · 5 min read08Verification
A Valid JSON Can Still Be Wrong: The Semantic Check After Schema Validation
A parser can confirm the shape of an answer; only an evidence-aware check can confirm whether its claims are supported.
September 4, 2026 · 6 min read09Prompt Labs
The Retrieval Update Test: Can Your Answer See Version Two?
When a source changes, a reliable source-backed answer should move with it, and the new lineage should be visible.
September 4, 2026 · 5 min read10Verification Protocol
A Prompt Is Not a Process: Four Tests for Reliable AI Work
Run these checks before a generated answer becomes a recommendation, deliverable, or decision.
August 29, 2026 · 8 minute protocolEditorial purpose
A skeptical test bench for AI-assisted work. Each issue names the input, procedure, failure condition, and readback required before an output can become a recommendation or deliverable.
01
Trace inputs
A clean answer is not reliable unless a reviewer can identify where its factual material came from.
02
Name failures
We test awkward examples and plausible mistakes, not only the demo that makes the system look good.
03
Read it back
Final state is checked against the approved state. Drafts, queues, and confirmations are not treated as proof.