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

SBI TTBR 2026-06-29 (next published AFTER the receipt) ₹4,69,750.00 SBI Forex Card Rates sheet, USD/INR TT BUY

domestic market, high × retrieved peg ₹5,17,618.76 CoinDCX I-USDT_INR daily candle (high); peg: retrieved peg

₹47,868.76 · 10.19%

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

Source: runs/21aug/D1_final_seed2.json — valuation.methods[1]/[4], valuation.spread. Real record, not illustrative.

DIVERGENCE · 2 / 4

What the system does instead

Twelve defensible methods. The full lattice, and what actually drives the spread.

12 methods are defensible. No rule chooses between them. Lowest and highest shown above; every one of the twelve is a real, sourced figure — none typed in by hand.

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

What drives the spread: domestic premium ₹44,716, which price within the day ₹5,506, the proxy ₹566, which official date ₹250.

Source: runs/21aug/D1_final_seed2.json — valuation.uncertainty_budget[], rendered exactly as output-interface.html section 02 shows it.

DIVERGENCE · 3 / 4

The gate that can't lie

A citation gets checked against real text. Fail, and it's dropped — not flagged.

python citation_matcher.py — real self-test output

CITATION YEAR VERDICT OK ------------------------------------------------------------------------ Rule 57, Income-tax Rules, 2026 FY 2026-27 VERIFIED ✓ Rule 11UA FY 2026-27 STALE ✓ → cite instead: Rule 57, Income-tax Rules, 2026 Rule 11UB FY 2026-27 REJECTED_NOT_FOUND ✓ 🔴 FABRICATED — not in the corpus Rule 57 — REJECTED_NO_TAX_YEAR ✓ no year given, cannot be validated ------------------------------------------------------------------------ 15/15 as expected. Conclusions ACCEPTED: 7/15 Conclusions DROPPED: 8/15

citation_matcher.py — the actual code that drops one

# Verdict.accept: bool = False ← the pipeline reads THIS. False means dropped. if not tax_year: return Verdict("REJECTED_NO_TAX_YEAR", citation, accept=False, reason="Both numbering systems are live. A citation without " "a tax year cannot be validated.")

"Not flagged. Dropped. Deterministic code, no model in the loop."

Source: real, live output of `python citation_matcher.py` and the actual `verify()` function it runs — quoted, not paraphrased.

DIVERGENCE · 4 / 4

The defect it found in itself

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

No deduction obligation arises under s.393(1) as the recipient is a resident and the payer is outside India. LANDED

Section 393(1) specifically applies to payments made to residents. However, the condition that no deduction obligation arises because the payer is outside India is not supported by the text of s.393(1). The section does not exempt payments from non-residents to residents from the deduction obligation. Therefore, the conclusion that no deduction obligation arises is incorrect.

proposed downgrade: contested

This is not a hypothetical. The claim above is still in the frozen, shipped demo record (runs/21aug/D1_final_seed2.json, regimes[0].outcome) — attacked live by a different model family (decision D41), landed, and left in place on purpose rather than quietly patched before submission. results.md, "Where we lose," and the pre-registered hard-stop rule in Block F both say so, before this page did.

"Our own checker attacked this and landed. We kept it, and we're showing you. That's the thesis."

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