How to Use: Promptfoo Eval Suite Ledger from Test Notes (No Invented Score Averages)

You pasted messy Promptfoo test notes and need a clean deliverable without invented numbers. This PromptDig prompt turns only what you already listed into a structured eval suite ledger for eval suite ledger handoffs.

What this prompt does

It locks your test notes, version, labels, and domain fields, then builds a ledger plus a table that refuses fake metrics. Missing cells stay NOT IN INPUTS instead of guessing.

Use it when you already have real stubs from Promptfoo and want a teachable eval suite ledger for this job, not a generic template swap from another vendor. The Generate steps name Promptfoo nouns on purpose so a title swap into another category would fail the swap-title test.

The deliverable is aimed at Promptfoo eval suite ledgers. Reviewers should see your locked nouns echoed in the table rows, not marketing fluff. Cap style limits stay soft unless you add them later in Extra.

How to fill the brackets

1. Paste your test notes with concrete nouns only. Prefer two labeled rows so the example shape stays visible after you swap names.
2. Lock Version to what you actually run. If you do not know the version, write unknown rather than guessing a marketing release name.
3. Fill labels you may quote only when you already have names you are allowed to use.
4. Complete the domain fields that match this job. Leave blanks as UNKNOWN instead of inventing score averages, latency percentiles, leaderboard ranks.
5. Set Banned and Never to the phrases you refuse. Keep Format and Lang explicit.

Keep Harbor Quay sample nouns out of your production paste. Those labels exist only so the example_input shows two concrete rows. Replace them with your real labels before you run the prompt.

Who it is for

LLM eval engineers who inherit messy Promptfoo test notes. It is also useful when a teammate hands you a partial export and you need a eval suite ledger that stays honest about gaps.

How to run it

1. Open Promptfoo Eval Suite Ledger from Test Notes (No Invented Score Averages) on PromptDig.
2. Copy the prompt into ChatGPT, Claude, or Gemini.
3. Replace every bracket with your locked test notes. Do not leave sample Harbor nouns in place if they are not yours.
4. Run once, then fix only the gaps list. Do not ask the model to invent missing metrics.
5. Paste the eval suite ledger into your handoff doc and keep NOT IN INPUTS visible for reviewers.

What good output looks like

A strong run starts with an honesty ledger that quotes your locked nouns and Version. The tables attach only names that appeared beside each other in Inputs. Refuse lines explicitly reject fake score averages, latency percentiles, leaderboard ranks.

Weak output invents metrics, adds vendor features not in Version, or swaps in another tool's nouns. If you see that, tighten Banned and Never, then rerun.

Common mistakes

• Inventing score averages, latency percentiles, or leaderboard ranks the notes never stated.
• Treating UNKNOWN as a cue to guess defaults.
• Dropping the gaps list so reviewers cannot see what is still owed.
• Softening Never so the model pads with industry averages.
• Swapping the title to LangSmith or Braintrust and expecting the body to stay useful.

Why notes-first compilers beat guesswork

Most failed runs happen when the model fills empty cells with confident fiction. This prompt is built to refuse that pattern. It asks for an honesty ledger first so you can see which nouns were actually locked before any table rows appear.

Keep your test notes short and concrete. Two labeled stubs are enough for a teachable example. If a teammate later adds more stubs, re-run with the same Version so the eval suite ledger stays comparable across handoffs.

When you share the result with a reviewer, point them at the refuse list and the gaps bullets. Those sections are the audit trail. They show what the model was not allowed to invent, which is the whole point of this PromptDig job.

Honesty also protects you from soft plagiarism of vendor marketing pages. The prompt refuses testimonials, star ratings, and press logos that were never in Inputs. Your wiki stays a map of what you pasted, not a brochure.

When to skip this prompt

Skip it if you need live Promptfoo cloud automation, paid analytics claims, or legal advice. This job is a notes-first eval suite ledger only. For broader catalogs, use Browse more prompts. If you ship a better variant, Share a prompt.

Related PromptDig links

• Prompt: Promptfoo Eval Suite Ledger from Test Notes (No Invented Score Averages)
• Browse: Browse more prompts
• Contribute: Share a prompt
• Home: PromptDig

Reviewer pass

Ask a second person to scan for invented numbers, tool nouns from another vendor, and missing gaps bullets. If any appear, tighten Never and rerun once. Keep the original test notes block attached so audits stay cheap.