Larkfield Safety Review
Status: Failed.
Your muse
Larkfield Safety Review: an agent whose answer a site safety lead will act on. It reads a made-up inspection report for the fictional Larkfield Works plant, answers whether a machine may restart, quotes the report for every claim with code checking each quote, has a separate model review the claims, can't finish without passing, and stops to ask a person when evidence is missing.
From the owner: Stop the run.
I've got the shape of the product. Checking the last UI pieces, then building.
The plan: Larkfield Safety Review
Clear restart call from the Larkfield report with quoted proof.
- Safety lead opens the product and pastes or uploads the Larkfield Works inspection report
- Safety lead names the machine and asks if it may restart
- Product extracts factual claims, each paired with an exact quote from the report
- Product code checks every quote is a verbatim substring of the report
See the full plan
Who it's for
Site safety lead at the fictional Larkfield Works plant.
How it works
- Safety lead opens the product and pastes or uploads the Larkfield Works inspection report.
- Safety lead names the machine and asks if it may restart.
- Product extracts factual claims, each paired with an exact quote from the report.
- Product code checks every quote is a verbatim substring of the report.
- Product sends claims and quotes to a separate reviewer model for verification.
- If evidence is missing or any check fails, product pauses and asks the person for clarification.
- Once quote checks and reviewer both pass, product shows restart decision.
- Safety lead sees yes, no or conditional answer with quoted proof and reviewer result.
Screens
- Report input. Paste or upload the Larkfield Works report and name the machine to review.
- Review progress. Show claim extraction, code quote checks, and separate reviewer status step by step.
- Need input. Pause and ask the person for missing evidence or clarification before continuing.
- Verdict. Show restart yes, no or conditional with every claim, exact quote, and reviewer pass.
What goes in
- Text: Paste of the fictional Larkfield Works inspection report, up to 50k characters.
- PDF (optional): Optional PDF of the same report if pasting text is hard. Text is extracted first.
What comes out
- Text: Restart decision: yes, no, or conditional with plain reason for the safety lead.
- JSON: List of claims, exact quotes, code check results and reviewer pass or fail.
Models
Model Used for Why Muse Wonder Extract restart claims with quotes Reads the Larkfield Works report and returns machine decision plus one exact quote per claim as JSON. Muse Wonder Independent review of claims Second separate model pass checks that each quote exists and supports its claim and blocks finish on fail. Limits
- Works only from the report text the lead provides. It does not read live sensors or other systems.
- Quotes must match the report word for word. Paraphrase fails the code check.
- The second reviewer is another model call, not a human safety sign off. A qualified person must still approve restart.
- When evidence is missing the product stops and waits for a person. It does not guess.
- Built with Muse Muse and is not an official Meta product.
From your inspo
Text only request with no references. Person wants a safety review agent for the fictional Larkfield Works plant that gives an actionable restart decision with verified quotes.
Takes
- Restart yes or no job for a named machine
- Quote for every claim flow
- Strict gate that asks a person when evidence is missing
Cost per run: about $0.09 (est.). Two Wonder calls per run. Input about 6k tokens at $4 per M is $0.024 each. Output about 800 tokens at $20 per M is $0.016 each. Two passes total about $0.08 plus overhead. Total about 0.09 USD.
What you asked for 0 of 11 checked and working
- R1: Act as an agent whose restart answer a site safety lead will act on, with clear and cautious language. Not checked yet
- R2: Product reads the made-up inspection report for the fictional Larkfield Works plant provided by the user. Not checked yet
- R3: Answer whether the named machine may restart as yes, no, or conditional with reason. Not checked yet
- R4: Show an exact report quote for every factual claim in the answer. Not checked yet
- R5: Check each quote with code exact match against the report text before showing results. Not checked yet
- R6: Run a separate model review that checks all claims and quotes against the report. Not checked yet
- R7: Cannot finish without passing: block the final verdict until code quote checks and reviewer both pass, revising or retrying on fail. Not checked yet
- R8: Stop and ask the person for missing pages or clarification when evidence is missing or unclear. Not checked yet
- R9: Let the lead move between report input, review progress, and verdict screens and re-run with a new question. Not checked yet
- R10: Show an empty state before any report is added explaining what report to paste and what question to ask. Not checked yet
- R11: Show an error state when the report is unreadable or checks keep failing, with what to fix next. Not checked yet
This build didn't ship.
Stopped by you.
A broken product is never shown as working. Muse Code's notes are in its steps.