Spaces:
Sleeping
Sleeping
File size: 3,790 Bytes
39ddf96 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 | import Panel from './Panel.jsx';
import { explainPanelQuestion } from '../api.js';
/**
* Dossier-v2 propagation: the "How grounding works" panel for the Copilot
* tab side column — the four-step guard chain, the status legend the rest
* of the UI uses, and the honest limitation, stated once. Copy grounded in
* wiki/pages/agent-layer.md and agent/graph.py (no live numbers).
*/
const STEPS = [
{
name: 'Route',
text:
'A temperature-0 router LLM (Gemma, DeepSeek fallback) classifies the question into ' +
'one of four validated Tier-1 numeric tools, the sandboxed pandas analyst ' +
'(analyze_data), the cited documentation retriever (query_model_docs), a REASONED ' +
'interpretation, or a refusal. Tool arguments are pydantic-validated with ' +
'extra="forbid" — unexpected fields are rejected, not ignored.',
},
{
name: 'Compute',
text:
'Only the deterministic engine computes numbers. The LLM never does arithmetic — ' +
'every figure in an answer originates in a tool result the engine produced.',
},
{
name: 'Narrate',
text:
'A narrator LLM writes the prose around the tool payload it was handed — nothing ' +
'else — and must cite the tool it is narrating.',
},
{
name: 'Verify',
text:
'Post-generation guards check that every digit AND every spelled-out number ' +
'("tens of millions" included — a live bypass found in adversarial review, now ' +
'blocked in all three guards with pinned regression tests) appears verbatim in the ' +
'tool payload, plus a citation check. One failed narration is regenerated once; a ' +
'second failure falls back to a deterministic template built from the payload itself.',
},
];
const LEGEND = [
{
dot: 'status-dot-good',
label: 'GROUNDED',
text: 'engine-computed; every figure verbatim from a tool payload, with citation.',
},
{
dot: 'status-dot-accent',
label: 'REASONED',
text: 'interpretation grounded in the model documentation — explicitly labelled "not engine output", and still number-guarded.',
},
{
dot: 'status-dot-warn',
label: 'THINKING',
text: 'request in flight.',
},
{
dot: 'status-dot-muted',
label: 'OUT OF SCOPE',
text: 'refused by design — the router found no validated path.',
},
];
export default function HowGroundingWorks() {
return (
<Panel
title="How grounding works"
subtitle="Why the header can promise 'every figure cites its source' — the four-step chain every question passes through."
buildExplainQuestion={() =>
explainPanelQuestion({
panelId: 'how_grounding_works',
title: 'How grounding works',
recap:
'Route (temp-0 router into tools / docs / REASONED / refusal) → Compute (only the engine produces numbers) → Narrate (prose around the tool payload) → Verify (verbatim digit + spelled-number + citation guards, one regenerate, then deterministic fallback).',
})
}
>
<ol class="ground-steps">
{STEPS.map((s) => (
<li key={s.name}>
<b>{s.name}.</b> {s.text}
</li>
))}
</ol>
<div class="ground-legend">
{LEGEND.map((l) => (
<p class="ground-legend-row" key={l.label}>
<span class={`status-dot ${l.dot}`} /> <b>{l.label}</b> — {l.text}
</p>
))}
</div>
<p class="caveat">
<b>Honest limitation.</b> The guards prove every number came from the engine; they
cannot prove the prose attributes each number to the right concept. That residual
attribution risk is documented in the model docs rather than hidden here.
</p>
</Panel>
);
}
|