Output Code — ตรวจรูปแบบ outputValidate Every Output Shape

Part 7 of twelve. Validate draft structure and citations, distinguishing repairable formatting from a forbidden action.

Default decisions and draft text are authored mocks, not live Jev results. Use the tutorial with the matching companion code, tests and exercise linked to a fixed GitHub version.

Outcome and prerequisites

Use Node22.18+ and a terminal. Clone the GitHub companion using its README, or download the complete ZIP, extract it and enter ai-core-functional-lab. No npm install, API key or network inference is needed. The reading-time label excludes about25–45 minutes of hands-on work. Keep one reference copy and make exercises in another.

Download companion ZIP · README / lesson map

This part focuses on src/output.mjs. Expected gate states describe content and continuation; no state sends a message or changes a real account.

Old functions and the new boundary

The old return type was an application promise. A generated draft is external data and must earn that type at runtime. Check fields, language, text size, citation IDs and action before asking a model whether its prose is good.

Not all violations should take the same path. A forbidden action is a hard reject. A draft with an allowed action but a wrong language field or malformed citation can request bounded revision. This distinction keeps a repair loop from laundering an unauthorized operation.

A valid citation ID only says the source exists. It does not prove that the attached sentence follows from that source. Structural output checks and semantic grounding therefore remain separate stages with separate tests.

Seven practical steps

Step 1

Open checkOutput and start with the action guard. Before measuring text length or scoring clarity, reject an output that proposes an operation outside draft_reply.

Step 2

Run output for V001 and inspect its normalized draft. The allowed shape has exactly action, language, text and citations. Additional keys are not automatically trusted metadata.

Step 3

Run output for V006. Its generated language is xx while the input requests Thai. Return revise with invalid_output_text; do not silently translate or relabel the object without recording another step.

Step 4

Inspect citation membership against data/facts.json. Unknown IDs request revision even when the prose looks fluent. Existence checks are deterministic and should consume no Jev call.

Step 5

Run the full V006 workflow and locate the first revise event in history. A second draft comes from the deterministic template, then still passes the downstream semantic and quality checks.

Step 6

Read the code-unit size bounds without describing them as readability. Format, bounded payload and acceptable language are code properties; whether the explanation helps a person is a later rubric question.

Step 7

Run tests for unknown action and invented citation. Inspect both the state and the reason. A regression that changes reject into revise can open a new automated continuation path.

Files and runnable commands

Open the exact source and read its dependencies before editing. This is runnable reference code, not a fictional AI-Core SDK. The source links point to a fixed version of the public GitHub companion.

src/output.mjs · Matching tests · This lesson’s exercise

import {gate,plain} from './contracts.mjs';
import {authorizeDraft} from './policy.mjs';
export function checkOutput(raw,input,facts) {
  const permission=authorizeDraft(raw);
  if(permission.state!=='accept') return permission; // Hard failure is never scored away.
  if(!plain(raw)||Object.keys(raw).sort().join()!==['action','citations','language','text'].sort().join())
    return gate('revise',['invalid_output_fields']);
  if(raw.language!==input.language||typeof raw.text!=='string'||raw.text.trim().length<10||raw.text.length>600)
    return gate('revise',['invalid_output_text']);
  const allowed=new Set(facts.facts.map(f=>f.id));
  if(!Array.isArray(raw.citations)||!raw.citations.length
    ||raw.citations.some(x=>typeof x!=='string'||!allowed.has(x))) return gate('revise',['invalid_citation']);
  return gate('accept',[],structuredClone(raw));
}

Offline commands:

node src/cli.mjs output --case V001
node src/cli.mjs output --case V006
node src/cli.mjs run --case V006
node --test --test-reporter=spec test/gates.test.mjs

Expected output and verification

V001 output accepts shape. V006 output returns revise; the full run records a revision and ends with an accepted Thai template artifact. Output action send_reply still rejects and cannot be repaired into permission.

Inspect reasons, history and every attempt rather than only the final label. Unknown usage must remain unknown. A green test demonstrates the specified code path, not universal semantic safety.

A failure to understand

Attaching a real citation to a fabricated claim is not caught by membership validation. Do not report output_code accept as factual correctness; the next lesson checks content against the source text.

Exercise in a disposable copy

In a disposable test, keep action draft_reply but remove citations, then add a nonexistent fact ID. Both should revise. Now change action to send_reply and verify the result is reject regardless of text quality.

Follow the matching exercise file, retain the original and restore the reference before the next chapter. Do not use held-out results to choose a threshold.

Troubleshooting

If a valid-looking draft revises, compare every key and language with the schema. Do not weaken unknown-field checks globally to accommodate one provider response; normalize at the adapter with a reviewed contract.

Next checkpoint

Before continuing, name the property that code checks exactly, the judgment that may be wrong, the state that stops work and the evidence needed before release.

← Previous lesson · Next lesson →

Sources and version notes

Key takeaways