why-agent — an agent that answers why, and cites every sentence

Nebius × NVIDIA Global AI Hackathon · Best Apps and Agents track · runs on Nebius Token Factory with NVIDIA Nemotron 3

Try it — the reconciler, running in this tab

Pick a case, untick a source, or type a different value into a claim. The verdict on the right is recomputed from what is on screen: this is the reconciler the Python runs, not a video of it. No key, no server, no network — the corpus and the reconciler are both in this page. The JavaScript is a port of evidence.reconcile(), and test_demo_app.py runs both over every item in every corpus and fails if a verdict, a conflict, a gap or an aspect differs.

What it does

Give it a work item. It reads what the card says, what the receipt says, what the commit says, what the check run said, and what two Nemotron reviewer models said — and then it tells you what happened, with the citation attached to each line. When the sources disagree it does not pick a winner: it shows every side, cites every side, and says which source is authoritative for what.

The answer it gives

This is the tool's real output, not a mock-up — the excerpt is taken verbatim from python3 demo.py when this page is generated, so it cannot drift into a nicer story than the tool tells.

CASE why-alpha — the sources disagree
------------------------------------------------------------------------------
# Why: Ship the alpha release notes

**Item:** `why-alpha`
**Verdict:** CONFLICTED
**Sources read:** 4

The sources disagree. That is the finding.

## What the sources say

### outcome

- **receipt** — the burn-down receipt says nothing was claimed → `not-claimed`  ·  `synthetic/notes/alpha.md — synthetic/notes/alpha.md:4`

### state

- **card** — the card was marked DONE → `DONE`  ·  `synthetic/cards/alpha.md — synthetic/cards/alpha.md:9`
- **commit** — a commit merged the fix → `merged`  ·  `git — 0c4f1a9 (synthetic)`
- **check** — the gate run is still red → `failing`  ·  `ci — synthetic://run/alpha/42`

## Conflicts

### state

- `DONE` — the card was marked DONE  ·  `synthetic/cards/alpha.md — synthetic/cards/alpha.md:9`
- `merged` — a commit merged the fix  ·  `git — 0c4f1a9 (synthetic)`
- `failing` — the gate run is still red  ·  `ci — synthetic://run/alpha/42`

**A single answer here would be a guess.** Each side is cited; the reader decides which source is authoritative for that aspect.

## Unknowns

- **cause** — no source says anything about this.

These are UNKNOWN, not passes. An aspect nobody reported is a gap in the evidence, and it is recorded as one.

## What to check next

- Resolve the state conflict between `synthetic/cards/alpha.md — synthetic/cards/alpha.md:9`, `git — 0c4f1a9 (synthetic)`, `ci — synthetic://run/alpha/42` — one of them is out of date, and which one is a decision, not a lookup.
- Find a source for cause: nobody reported it, so it is UNKNOWN rather than fine.

> A card says DONE, a commit says merged, a check says red, a receipt says nothing was claimed. Four sources, four answers. The agent does not pick one — it shows all four and says which one is authoritative for what.

------------------------------------------------------------------------------

What the run measured

An adversarial corpus of 14 planted-claim work items, with a truth file the lanes never see. The deterministic reconciler is the only thing that decides; the model lanes contribute claims and an advisory verdict.

The lanes err in both directions — the fast lane believed a card that claimed a failed deploy had shipped. That gap is the architecture, not a bug: the lanes advise, the reconciler decides. Reviewer-lane cost $0.033, 0 retries. Reproduce it: python3 benchmark/score.py --eval-results benchmark/eval-benchmark-2026-10-28.json

The lanes, and what each one is for

The live-web lane queries Tavily at run time, and a page becomes evidence only if the claimed value literally appears in it — 3 rows across 3 of the corpus items are Tavily's, and one of them records a miss rather than dropping the row. A page that does not match cannot quietly become a pass.

Every lane output enters as a claim, never as a verdict, because any integration can be poisoned. The model may write the prose; it may not add a fact — every value a narration states is checked against the evidence, and a value that is not there is reported, not removed.

Run it yourself

git clone https://github.com/Snedz/why-agent && cd why-agent
python3 demo.py                                   # the cases, as a reader sees them
python3 benchmark/score.py --eval-results benchmark/eval-benchmark-2026-10-28.json
python3 demo_lab/flip_check.py                    # proves the model lanes change the verdict

Earlier build

Synthetic data only — no real card, customer or identifier. The reconciliation is deterministic Python; the model writes the prose and is checked against the evidence. The model runs were real: 14 items, 51 claim rows, recorded 2026-10-28.