f89e074dd1
Governed substrate for autonomous agents: scoped identity (passports), audited actions, MCP workspace. Infra IPs and secrets redacted for public release.
98 lines
4.4 KiB
JavaScript
98 lines
4.4 KiB
JavaScript
// @vitest-environment node
|
|
/**
|
|
* DCD fast-guardrail contract (ADR-157), shadow-mode rule-of-thumb.
|
|
*
|
|
* Pins:
|
|
* - retrieved context that shares the query's terms → grounded
|
|
* - retrieved context disjoint from the query → not grounded (the hallucination
|
|
* case the guardrail exists to flag)
|
|
* - empty recall → not grounded ("no retrieved context")
|
|
* - a contentless query is not flagged (cannot assess)
|
|
* - the verdict is pure data — never mutates inputs (shadow mode)
|
|
*
|
|
* Adaptive head (ADR-157 amendment — adaptive-head replaces fixed 150-window):
|
|
* - a near-front hit settles cheap (inspectedTokens stays small)
|
|
* - support living PAST the old fixed window is still found (window expands)
|
|
* - "not grounded" is only declared after the whole context is inspected
|
|
* (inspectedTokens === totalTokens) — the honest-exhaustion guarantee
|
|
*/
|
|
import { describe, it, expect } from 'vitest';
|
|
import { assessGrounding, GROUNDING_THRESHOLD } from '../fastGuardrail.js';
|
|
|
|
describe('assessGrounding', () => {
|
|
it('marks context grounded when it shares the query terms', () => {
|
|
const v = assessGrounding('how is the deploy pipeline configured', [
|
|
{ text: 'the deploy pipeline runs on pm2 and is configured per release' },
|
|
]);
|
|
expect(v.grounded).toBe(true);
|
|
expect(v.score).toBeGreaterThanOrEqual(GROUNDING_THRESHOLD);
|
|
});
|
|
|
|
it('flags ungrounded context disjoint from the query', () => {
|
|
const v = assessGrounding('what are the vacation payroll rules', [
|
|
{ text: 'nginx restarts when the docker container exits unexpectedly' },
|
|
]);
|
|
expect(v.grounded).toBe(false);
|
|
expect(v.reason).toMatch(/no grounding/i);
|
|
});
|
|
|
|
it('returns not-grounded with no retrieved context', () => {
|
|
const v = assessGrounding('anything', []);
|
|
expect(v.grounded).toBe(false);
|
|
expect(v.reason).toMatch(/no retrieved context/i);
|
|
});
|
|
|
|
it('does not flag a query with no content terms', () => {
|
|
const v = assessGrounding('what is it', [{ text: 'some unrelated memory' }]);
|
|
expect(v.grounded).toBe(true);
|
|
expect(v.reason).toMatch(/not assessable/i);
|
|
});
|
|
|
|
it('is pure — does not mutate the inputs (shadow mode)', () => {
|
|
const memories = [{ text: 'deploy pipeline release' }];
|
|
const snapshot = JSON.stringify(memories);
|
|
assessGrounding('deploy pipeline', memories);
|
|
expect(JSON.stringify(memories)).toBe(snapshot);
|
|
});
|
|
|
|
it('settles a front-loaded hit cheaply — inspects far fewer than total tokens', () => {
|
|
// Query terms appear in the first sentence; a long tail follows. The
|
|
// adaptive head must accept early without reading the whole context.
|
|
const tail = Array.from({ length: 300 }, (_, i) => `filler${i}`).join(' ');
|
|
const v = assessGrounding('deploy pipeline pm2', [
|
|
{ text: `the deploy pipeline runs on pm2 per release. ${tail}` },
|
|
]);
|
|
expect(v.grounded).toBe(true);
|
|
expect(v.inspectedTokens).toBeLessThan(v.totalTokens);
|
|
expect(v.inspectedTokens).toBeLessThanOrEqual(40); // first probe window
|
|
});
|
|
|
|
it('expands to find support living PAST the old fixed 150-token window', () => {
|
|
// Support sits at ~token 200 — beyond the retired HEAD_TOKENS=150 cutoff.
|
|
// The fixed window would have falsely reported "no grounding"; the adaptive
|
|
// head doubles past it and finds the hit.
|
|
const filler = Array.from({ length: 200 }, (_, i) => `noise${i}`).join(' ');
|
|
const v = assessGrounding('vacation payroll', [
|
|
{ text: `${filler} the vacation payroll rules are documented here` },
|
|
]);
|
|
expect(v.grounded).toBe(true);
|
|
expect(v.inspectedTokens).toBeGreaterThan(150);
|
|
});
|
|
|
|
it('rejects only after exhausting the whole context (honest exhaustion)', () => {
|
|
const v = assessGrounding('quantum cryptography lattice', [
|
|
{ text: 'the deploy pipeline runs on pm2 and nginx restarts on crash' },
|
|
]);
|
|
expect(v.grounded).toBe(false);
|
|
expect(v.reason).toMatch(/no grounding/i);
|
|
// Honest reject: we read everything available before saying "not grounded".
|
|
expect(v.inspectedTokens).toBe(v.totalTokens);
|
|
});
|
|
|
|
it('caps inspection at HEAD_MAX on very large ungrounded context', () => {
|
|
const huge = Array.from({ length: 5000 }, (_, i) => `token${i}`).join(' ');
|
|
const v = assessGrounding('unrelatedterm anotherterm', [{ text: huge }]);
|
|
expect(v.grounded).toBe(false);
|
|
expect(v.inspectedTokens).toBeLessThanOrEqual(600); // HEAD_MAX safety cap
|
|
});
|
|
});
|