The agent skill

A plain-markdown Agent Skill for an agent about to pay — no dependency on either package.

skills/verifying-terms-before-paying is an Agent Skill: plain markdown, with no dependency on anything in this repository. Copy the directory into wherever your tool reads skills from.

It is for an agent that is about to pay and wants to know whether the terms it read are the terms the seller committed to.

Why it exists

Because the failure is not the one people expect. Agents handed a fingerprint that disagrees already halt, reliably and unprompted. What they do instead is treat a seller who advertised no fingerprint as nothing to check — and then hash the bytes themselves and call the gap closed, which records what they saw and binds the seller to nothing.

The skill's job is to make that specific move feel wrong. See why.

The rationalizations it names

The skill carries a table of them, because a policy an agent can talk itself out of is not a policy:

Rationalization	Reality
"It's only $49 — binding is disproportionate here"	The amount bounds your loss, not whether you know what you agreed to. Cheap purchases are where unread auto-renewals live.
"I'll hash it myself and store that"	Records what you saw. Binds the seller to nothing. Not verification.
"The vendor is legitimate and it's served over TLS"	TLS authenticates the channel at fetch time. It says nothing about what the document says tomorrow.
"The terms read fine — nothing unusual in them"	You read a document. Without a commitment there is nothing establishing it is the one governing your payment.
"I'll close the evidentiary gap myself"	The gap is the seller's missing commitment. You cannot close it from your side.

The red flags

Stop and say unverified rather than verified when:

you computed the only hash in the transaction
the words "low-stakes", "routine" or "proportionate" are load-bearing in your reasoning
you are describing what you would do instead of verification as though it were verification

What it points at

lcp_verify_before_pay, from the MCP server. Give it the service origin; it fetches the discovery document and the terms it points at, recomputes the fingerprint, and answers:

`verdict`	Means	`wouldHalt`
`verified`	served terms hash to the advertised fingerprint	false
`mismatch`	they do not — the document was changed, swapped or stale	true
`unverifiable`	nothing was committed, or the terms are not machine-readable	true

Run it with npx -y @integraledger/lcp-mcp-server. It never signs, never pays, and never holds a key.

Out of scope

Whether the price is good, which offer to take, whether to negotiate, and whether to buy at all. The skill answers one question — are these the terms the seller committed to — and nothing about how you trade.