Ground agent answers in cited document blocks

01 · Role in the system

What Jev does here

Parsing and structural indexing happen in the browser, keeping block IDs with section, location and page. No lexical, vector or keyword retriever filters blocks before Jev: documents of 24 blocks or fewer are judged in full, larger ones are split into contiguous structural groups of at most 12, Jev sees a preview from every group, and full blocks arrive only from groups whose signals pass a permissive discovery gate. Each block gets four Nouls - relevant, evidence-bearing, contradicts-premise and prompt-injection - catching injected instructions alongside evidence selection. The result is a cited evidence set with per-block traces carrying section, paragraph, page and scores; the LLM writes the answer from it and can call verify_claim to check individual claims against the cited blocks. When evidence is absent or coverage incomplete, the tool says so rather than fabricating.

02 · Control boundary

Where Jev sits

Jev as the evidence judge in front of the writer: typed Nouls per block select and guard the evidence set, code keeps the trace, the LLM only writes from cited blocks, and per-claim verification closes the loop.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• No hosted demo; runs as a local UI plus a Cloudflare Worker.
• On large documents the discovery gate trades full-coverage judging for cost; group previews cap at 225 characters per block.
• Evidence quality depends on document parsing (PDF, DOCX, Markdown, JSON, source code, ZIP).

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• Milothik ↗Community · github · public · checked 2026-09-28