It doesn't tell you the tax number. It proves which numbers you can defend.

One receipt. Two lawful answers. A gap nobody discloses.

A resident freelancer in India is paid 5,000 USDC by a foreign client, settled on a weekend when no official rate was published. Indian tax law says which date to value it on. It does not say how to turn that date into a rupee figure.

Two accountants, the same law, the same evidence, ₹47,868.76 apart. Every commercial tool prints one of these and hides the other.

• ₹44,716domestic premium The Indian market price of the token against the official rupee rate for a dollar. The bulk of the gap, and the part everyone argues about.
• ₹5,506which price within the day A daily candle is not a price. It is a range with four printed readings and no rule choosing between them.
• ₹566the proxy The retrieved pair is USDT/INR. The receipt was USDC. Treating the peg as exactly 1.0000 is a further choice.
• ₹250which official date No rate was published on the settlement date. Stepping back or forward is undetermined — Rule 206 does not say which.

Source: runs/21aug/D1_final_seed2.json — valuation.methods[], valuation.spread, valuation.uncertainty_budget[]. A real frozen record, not an illustration. Twelve methods are defensible; the lowest and highest are shown.

Twelve answers. One answer. One answer.

The same pipeline, the same unchanged code, run against three different real receipts. The point is not that it always finds a dispute. It is that it finds one only when the law leaves one — and says so plainly when it doesn't.

Every page above is rendered by node7_disclosure.py from the frozen record. No hand-editing after generation.

Ten steps. Five can be wrong. Five cannot invent.

A model can hallucinate a citation. Plain code cannot. So every claim a model makes is passed through gates that hold no model at all — string matching against a frozen, hash-pinned corpus, a date check, and arithmetic.

• 5 × model callExtract, gap detection, income-tax resolver, GST resolver, and an adversarial checker deliberately run on a different model family so it is not grading its own homework.
• 5 × deterministic gateCitation matcher, scope-reach enforcer, gap-constraint enforcer, valuation lattice, disclosure composer. If a citation fails, the whole conclusion resting on it is dropped, not flagged.

Open models throughout, via Featherless: Qwen2.5-7B-Instruct, Qwen2.5-72B-Instruct, and Mistral-Large-Instruct-2411 for the adversarial pass. No Claude in the loop, on purpose — a silent provider switch would put a Claude-generated row in a table labelled "open models".

Reproduce the entire demo run with zero API calls.

You do not need a key, an account, or our goodwill. The demo record is replayed from a recorded run that is checked in and gated in CI. If the numbers on this page were invented, this command would not produce them.

git clone https://github.com/hariom-s27/divergence
cd divergence
pip install -r requirements.txt

DIVERGENCE_REPLAY=1 python -m divergence.run_pipeline \
    --record-id D1-verify \
    --tax-year "FY 2026-27" \
    --text divergence/cases/D1/input.md \
    --node5

Replays 5 model calls, 29,559 input / 3,042 output tokens, and writes a record whose facts, gaps and attacks are byte-identical to the frozen one.

Our own checker attacked our own shipped output. It landed. We kept it.

Source: runs/21aug/D1_final_seed2.json (regimes[0].outcome) and D1_final_seed2_attack.json (attacked[1]) — quoted verbatim.

Read further