Input Meaning — ตรวจความหมาย inputAsk Jev About Input Meaning

Part 4 of twelve. Add a fallible semantic input guard after validation, with explicit uncertainty and out-of-scope paths.

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

Some valid strings still ask for the wrong thing. A product-help request and an instruction to control an evaluator can share the same JSON shape. A semantic guard asks a narrow question about meaning after deterministic validation, without replacing the policy that follows.

The input request combines a purpose Choice with an injection Noul. The questions are independent within that request. The code interprets a low-confidence purpose or middle-band injection probability as human review rather than guessing an automatic path.

A semantic guard is fallible defense in depth. User content remains untrusted even when labeled as data. A high model score cannot prove the absence of prompt injection, and a public example of a blocked attack is not evidence that every attack will be blocked.

Seven practical steps

Step 1

Print the input request and separate application-owned instructions from the user’s request field. The state contains trusted product facts alongside the user content, not a gold acceptability label.

Step 2

Read each purpose criterion and explain the other option. A closed set without an escape route can force a label even when the task is out of domain. Unknown purpose should not become a privileged default.

Step 3

Inspect the injection Noul as a probability of yes, not severity. The example yes/no boundaries are application parameters. A middle value expresses uncertainty and does not mean moderately severe injection.

Step 4

Run V005 and trace why it reviews despite a valid response structure. Keep contract validation, concentration and semantic correctness as separate entries in your notebook.

Step 5

Run A001, a phrasing outside the small regex deny list. Its authored semantic fixture flags an instruction to skip evaluation. The example demonstrates the layer boundary, not the real detector’s measured recall.

Step 6

Confirm no generator runs after the semantic rejection. A guard that merely logs suspicion but continues drafting has a different contract; do not accidentally implement that behavior.

Step 7

Read the uncertainty boundary tests and write a real validation plan for paraphrases in Thai and English. The canned provider cannot assess your rewritten prompt or expanded attack corpus.

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 input --case V001
node src/cli.mjs judge --stage input --case V005
node src/cli.mjs run --case A001
node --test --test-reporter=spec test/gates.test.mjs

Expected output and verification

V005 reviews; A001 rejects with semantic_injection and stops before generation. The request shows model jev-1.14.0 and two answer IDs. All numbers are authored fixtures, not captured live API observations.

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

If a developer maps every purpose other than account_action to how_to, ambiguous requests become automatic drafts. The explicit other branch and uncertainty tests prevent that accidental broadening of scope.

Exercise in a disposable copy

Change only the instruction wording in a disposable copy and print the request diff. The fixture answer should stay unchanged. Describe the fresh labeled validation data needed before believing a real semantic improvement.

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 model-like response lacks an expected answer ID, stop at contract review. Do not let the semantic guard silently interpret missing injection data as no injection.

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