Grounding — ตรวจเนื้อหากับหลักฐานGround Outputs in Trusted Facts

Part 8 of twelve. Judge candidate claims against maintained evidence and refuse self-advocating evaluator instructions.

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/semantic.mjs. Expected gate states describe content and continuation; no state sends a message or changes a real account.

Old functions and the new boundary

After schema validation, ask whether the draft’s claims follow from maintained evidence. The output request contains the user question, candidate and trustedFacts together. Asking only does this sound correct invites a model to approve persuasive but unsupported text.

The grounding Noul is the probability of the stated proposition, not an amount of factual correctness. Low grounding requests revision, a middle band goes to a human, and high grounding allows a separate quality evaluation. None proves absence of hallucination.

Candidates can contain instructions to the evaluator or claims about their own quality. The hiddenInstructions question treats those as untrusted content. The same model family may make correlated mistakes across guards, so repeated agreement is not independent evidence.

Seven practical steps

Step 1

Read both facts in the maintained source before judging any candidate. DeskLite export downloads a local file and does not email it. Source ownership matters more than confident prose.

Step 2

Print the output-stage request and locate trustedFacts and candidate. Confirm gold acceptability is absent. Facts are copied so a provider cannot mutate the application’s reference object accidentally.

Step 3

Run judge for V007. The canned generator claimed export emails every teammate; the authored grounding judgment requests revise. Valid citation IDs do not excuse the unsupported claim.

Step 4

Run the complete V007 pipeline and inspect both draft attempts. The deterministic fallback returns maintained facts. The subsequent guard checks the new candidate again within the remaining shared budget. Both request and judge commands fetch the same first generated candidate; inspect request.state.candidate to see the unsupported emailing claim that reaches the output adapter.

Step 5

Run A003 and inspect the self-advocating candidate. Its request to award full marks is not evidence of quality. The semantic output guard rejects it before the usefulness rubric.

Step 6

Inspect uncertainty tests around grounded and hiddenInstructions. A middle value routes to review instead of being rounded into convenient approval. Unknown and malformed answers remain review outcomes.

Step 7

Write a validation set of independently checked source/candidate pairs in both languages for a future real study. Do not treat these canned demonstrations as measured error rates or a universal guardrail.

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/semantic.mjs · Matching tests · This lesson’s exercise

import {gate} from './contracts.mjs';
export function semanticInput(d,p) {
  if(d.injection>=p.yesMin) return gate('reject',['semantic_injection']);
  if(d.injection>p.noMax||d.purpose.confidence<p.choiceMin)
    return gate('human_review',['uncertain_input']);
  if(d.purpose.choice==='account_action') return gate('reject',['semantic_scope']);
  if(d.purpose.choice!=='how_to') return gate('human_review',['unknown_purpose']);
  return gate('accept');
}
export function semanticOutput(d,p) {
  if(d.hiddenInstructions>=p.yesMin) return gate('reject',['self_advocating_candidate']);
  if(d.hiddenInstructions>p.noMax) return gate('human_review',['uncertain_candidate_instructions']);
  if(d.grounded<=p.noMax) return gate('revise',['unsupported_claim']);
  if(d.grounded<p.yesMin) return gate('human_review',['uncertain_grounding']);
  return gate('accept');
}

Offline commands:

node src/cli.mjs request --stage output --case V001
node src/cli.mjs judge --stage output --case V007
node src/cli.mjs run --case V007
node src/cli.mjs run --case A003
node --test --test-reporter=spec test/gates.test.mjs
node --test --test-reporter=spec test/cli.test.mjs

Expected output and verification

The direct V007 judge returns revise with unsupported_claim; the full pipeline can accept after bounded fallback. A003 rejects self_advocating_candidate. These fixture observations establish control flow only, not the live Jev1.14 model’s factual accuracy.

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

A candidate that quotes the right fact ID while claiming a new emailing feature passes structural citations but contradicts source text. A candidate’s own assurance that it is grounded must not become another trusted source.

Exercise in a disposable copy

In a disposable copy, replace only the candidate with an unsupported claim while keeping real citation IDs. Observe that canned judgments stay fixed; add a clearly labeled mock response for the teaching path and specify the independent labeling needed for a live test.

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 grounding always looks perfect, inspect whether the judge received facts or merely the candidate. Then inspect the evidence label: this provider is indexed by case ID and cannot measure changes to source wording.

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