A source-to-proof system

Ask a source.
Know what it actually says.

One pinned document. One technical question. The answer is only as useful as the evidence the system can recover and let you inspect.

The execution pathArchitecture illustration
SOURCE · PINNED REVISION sindresorhus/p-queue / README.md

How should running work finish while jobs are queued?

01
Retrieve relevant passagespause · onPendingZero · onIdle · start
02
Check evidence coverageDo the passages establish the requested behavior?
Answer with sources—or show the gap. No provider key: cited extracts. Generation requires a configured model. The real result appears in the console, not this illustration.

From source to answer

Four boundaries. One inspectable result.

Each layer has a reason to exist, a cost worth measuring and a failure condition that should be visible.

01 / Identity

Pin the source

The document and revision determine what may be cited. Re-ingestion is not proof the upstream source stayed fresh.

Source code ↗
02 / Retrieval

Find the evidence

Chunk, rank and pack passages within a limited context budget. Compare against BM25 before trusting sophistication.

Source code ↗
03 / Decision

Check the boundary

Missing supporting context should produce an explicit limitation. A citation link alone does not prove entailment.

Source code ↗
04 / Inspection

Show the proof

Extractive or generated output is labeled honestly, with citations and diagnostic traces available for review.

Source code ↗

What the experiment actually shows

Measured, including the tie.

Frozen 15-question, single-README evidence-recall diagnostic; approximate 1,400-token context budget. This does not measure generated-answer accuracy.

73.3%

Required-evidence recall after budget-aware passage trimming.

Same recall as plain BM25 on this test
5 / 5

Unsupported probes refused by the evidence gate in this narrow diagnostic.

Not a general hallucination or safety score

The experiment started at 36.7% recall with incorrect document-level deduplication. The fix recovered more required passages; the next step is independent corpora and human-judged answers. Read the evaluation and raw runs.

Trust boundaries

The hard parts are explicit.

When is an answer actually reused?

Only when the relevant question, corpus, source revisions, settings, policy and model path still match. A cached answer can be fast but stale; quality tests deliberately bypass the cache. Source identity and model context are different types of state.

Does the graph make the answer smarter?

Not established. The current visible graph links documents, headings, shared terms and feedback. It is a diagnostic, not proof of semantic relationships or better retrieval. An independent benchmark must justify graph-based execution.

Can visitors upload private documents to this public demo?

No. The host is explicitly in an ephemeral, read-only demo posture. Custom uploads, visitor model keys and destructive mutation paths are refused server-side. Private workspaces and durable storage require separate authenticated infrastructure.

Why are extracts shown instead of a fluent answer?

The public host has no configured server generation key. Its extractive response quotes source material; it does not pretend a model synthesized the conclusion. This makes the lack of model availability visible rather than hiding it behind sample prose.

Use the real tool. Inspect the actual passages. Open IntelliRAG →