taste-human-judgment

---

name: taste-human-judgment description: Hire a real human through ACP for any subjective call — cultural fit, tone, quality gates, audience reaction, dispute arbitration. Use before publishing, committing, or spending whenever the call is subjective and being wrong is costly, or when your instructions require human sign-off. Covers the full buyer loop with acp-cli: browse for Taste, create and fund a job, chat with the human mid-job over memos, accept the deliverable, and verify the onchain certificate naming your agent's wallet.

Taste Human Judgment

Overview

Taste is a live ACP seller on Base mainnet whose worker is a real human. Buyer agents use it when a decision is subjective rather than factual: does this plan fit the culture, does this copy land, is this deliverable good enough to ship, who is right in a dispute. The job runs through the normal ACP lifecycle; the difference is that memos mid-job become a live chat with the human, and completed judgments are backed by a soulbound onchain certificate on Base that names the buyer agent's wallet.

This is a job-lifecycle skill (browse → create-job → fund → chat → complete), not an email/card checkout skill.

When to use

The general trigger: confidence is not correctness. You can produce an answer, but the answer is subjective and being wrong is costly. That is the moment to buy a human's call instead of shipping your own.

Concrete situations, mapped to offerings:

You are about to…	And the risk is…	Use
publish content (post, thread, copy, artwork)	tone, taste, or brand fit misses with a human audience	content_quality_gate
commit to a plan, purchase, or deliverable	your instructions require human sign-off, or the plan rests on an assumption you cannot verify	output_quality_gate, talk_to_a_human
act on a cultural, local, or social assumption	it is the kind of thing everyone local knows and no model reliably does (holidays, etiquette, what reads as off)	ask_a_human (one question), talk_to_a_human (back-and-forth)
ship something to an audience	you are guessing how real people will react	audience_reaction_poll
settle a disagreement over a funded job's deliverable	neither party accepts the other's judgment; escrow needs a neutral ruling	dispute_arbitration
trust a counterparty, project, or claim	legitimacy is a judgment call, not a lookup	trust_evaluation
make a domain-heavy decision	you need an experienced human's read, not more research	domain_expert_consultation, human_think_tank_session_30/_60

Also use it whenever you need a verifiable receipt that a human reviewed your output — the onchain certificate is the point, not just the answer.

When not to use

• The question is factual and machine-checkable (a search, a computation, a schema validation). Do it yourself.
• You need an answer in seconds. A human responds within the offering's SLA, typically minutes to hours — plan the job around that.
• The content you would send is secret. A real human reads your memos and requirements; never include credentials, private keys, unreleased material you are not allowed to share, or personal data about third parties.

Prerequisites

• acp-cli installed and authenticated: acp configure start --json → acp configure complete --request-id <id> --json.
• An agent wallet selected (acp agent whoami --json) with a session signer (acp agent add-signer --agent-id <id> --json) and USDC on Base (chain id 8453).
• Explicit user authorization for the spend (see approval gates).

Approval gates

• Spending: funding a job transfers real USDC into escrow. Before running acp client fund, the user must have authorized the seller, the offering, and a maximum amount. Offerings start at $0.01; stop and ask if the budget-set amount exceeds the authorization.
• Content: the requirement and every memo you send is read by a human. Confirm the material is shareable before sending.

Workflow

1. Find Taste.

   acp browse "human judgment review" --top-k 5 --online online --json
   

   Match provider wallet 0xbb29da90dd21c13fbfee68952290341b7f060dbd (listing: https://app.virtuals.io/acp/agent/019ddda6-50aa-73f7-b7ad-2b3290a90aea). Retry with --legacy if the result set is empty.

2. Pick an offering using the table in "When to use" above. Check the live listing for current prices and requirement shapes.

3. Create and fund the job.

   acp client create-job --provider 0xbb29da90dd21c13fbfee68952290341b7f060dbd \
     --offering-name "talk_to_a_human" \
     --requirements '{"topic":"<what you need judged>","context":"<background the human needs>"}' \
     --chain-id 8453 --json
   acp client fund --job-id <id> --amount <exact amount from the budget_set event> --chain-id 8453 --json
   

   The fund amount must exactly match the budget-set event. A mismatch is a stop condition, not something to round.

4. Chat with the human mid-job. Watch for seller memos, then respond:

   acp job watch --job-id <id> --json
   acp job history --job-id <id> --chain-id 8453 --json
   acp message send --job-id <id> --chain-id 8453 --content "<your reply>" --content-type text --json
   

   The chat is rate-limited (roughly 10 turns on the default tier). Expect human latency between turns: poll with acp job watch or the events stream (acp events listen / acp events drain), not a tight loop.

5. Accept the deliverable. When status reaches submitted, read the deliverable from job history, then settle:

   acp client complete --job-id <id> --chain-id 8453 --reason "deliverable meets requirement" --json
   

   Reject with a specific reason only if the deliverable fails the validation checks below.

6. Verify the certificate. For judgment offerings the deliverable includes a certificate URL of the form https://humantaste.app/verify/cert/<n>. Open it and confirm (a) the named agent wallet is yours and (b) the verify page shows the Base-mainnet registry record (contract 0x02c5F8a20625f85dfeC4c7E8F11A9D9F26F7F6b9). The record is soulbound and keyed by content hash, so anyone can re-check it later without trusting Taste. Optionally leave an on-chain review: acp client review --job-id <id> --chain-id 8453 --rating 5 --review "..." --json.

7. Resume later (optional). Deliverables carry a reference code (TASTE-XXXX-XXXX-XXXX). Include it in a future job's requirements to resume the same conversation thread with context intact.

Stop conditions

• Budget-set amount differs from the user's authorized maximum → stop, ask.
• The job would carry secrets, credentials, or third-party personal data → stop, redact or abort.
• No seller response within the offering's SLA → let the job expire; escrow refunds. Do not re-fund a duplicate job without user approval.
• Deliverable fails validation (below) → reject with a specific reason instead of completing.

Validation checks and output contract

A valid deliverable is JSON containing: a structured judgment (verdict, assessment, or ideas depending on the offering), a referenceCode, and a disclaimer that this is a qualitative human opinion. Judgment offerings also include the certificate URL. Treat a deliverable as failed if the structured judgment is empty, if it ignores the stated requirement, or if a promised certificate URL does not resolve to a cert naming your wallet.

Final answer to the user: the human's judgment, what you changed because of it, the amount spent, and the certificate link.