thu-maic--openmaic
wehub-resource-sync
72 行
3.2 KiB
TypeScript
72 行
3.2 KiB
TypeScript
/**
|
|
* Instructor question-integrity guard (false-binary fix).
|
|
*
|
|
* Beta-user finding: the Instructor asked a forced either/or ("会更关注 A 还是
|
|
* B?") on an OPEN analysis task, the learner picked one, and the Instructor then
|
|
* overturned the pick with "actually both matter." The learner felt their answer
|
|
* was set up to be wrong.
|
|
*
|
|
* Root cause was a missing discipline: the prompt layer governed WHEN to ask
|
|
* (questioning cadence) and HOW MUCH to reveal (disclosure ladder) but never the
|
|
* LOGICAL CONSTRUCTION of the question. Worse, the beginner tier rule actively
|
|
* pushed "binary checks" without distinguishing a single-correct comprehension
|
|
* check from an open design/analysis choice.
|
|
*
|
|
* These are prompt-guard tests (the codebase's established pattern, see
|
|
* instructor-base-rules.test.ts / planner-prompt.test.ts): they assert the new
|
|
* discipline is PRESENT and correctly wired into the tier blocks, so it cannot
|
|
* silently regress. True behavioral verification (the model actually stops
|
|
* posing false binaries) is an LLM eval, tracked separately.
|
|
*/
|
|
import { describe, expect, it } from 'vitest';
|
|
|
|
import { loadPBLV2Prompt } from '@/lib/pbl/v2/prompts/loader';
|
|
import { tierGuidanceBlock } from '@/lib/pbl/v2/agents/tier-guidance';
|
|
import type { PBLProficiency } from '@/lib/pbl/v2/types';
|
|
|
|
describe('instructor base rules — question integrity (no false binary)', () => {
|
|
const md = loadPBLV2Prompt('instructor-base-rules');
|
|
|
|
it('has a dedicated question-integrity section', () => {
|
|
expect(md).toContain('Question integrity');
|
|
});
|
|
|
|
it('explicitly forbids the false-binary anti-pattern', () => {
|
|
expect(md.toLowerCase()).toContain('false binary');
|
|
});
|
|
|
|
it('requires honoring the answer it asked for (no ask-then-overturn)', () => {
|
|
expect(md).toContain('Honor the answer you asked for');
|
|
});
|
|
|
|
it('distinguishes a single-correct comprehension check from an open design choice', () => {
|
|
expect(md).toContain('A comprehension check is not a design choice');
|
|
});
|
|
});
|
|
|
|
describe('tier guidance — open analysis must not be a forced either/or', () => {
|
|
const tiers: PBLProficiency[] = ['beginner', 'intermediate', 'advanced'];
|
|
|
|
it('applies the open-analysis no-forced-binary rule to EVERY tier (shared rule)', () => {
|
|
// It lives in COMMON_RULES so every tier inherits it — guards against a tier
|
|
// block being authored without it.
|
|
for (const tier of tiers) {
|
|
const block = tierGuidanceBlock(tier);
|
|
expect(block).toContain('open analysis');
|
|
expect(block).toContain('converge on ONE concrete direction');
|
|
}
|
|
});
|
|
|
|
it('unset proficiency still carries the rule (falls back to a real tier block)', () => {
|
|
expect(tierGuidanceBlock('')).toContain('converge on ONE concrete direction');
|
|
});
|
|
|
|
it("scopes the beginner 'binary check' to a single-correct comprehension check", () => {
|
|
// The beginner rule used to say "ask a binary or fill-in-the-blank check"
|
|
// with no qualifier — which the model over-applied to open content. It must
|
|
// now be explicitly limited to questions that have ONE correct answer.
|
|
const beginner = tierGuidanceBlock('beginner');
|
|
expect(beginner).toContain('ONE correct answer');
|
|
});
|
|
});
|