Reconciliation engine.

The same engine, graded on a set it was tuned on vs one it never saw.

A flatters the engine by construction. B is adversarial - and its failures are shown in red, not hidden. B is the number a diligence team should read. Full breakdowns below.

The honest benchmark - failures included.

A second generator whose noise is not aligned to the engine’s tiers: vendor re-coded series, digit transpositions, Rule 46(b) 16-character truncation, wrong-branch GSTINs under the same PAN - plus collision traps (recurring same-amount billing, doc-number reuse) that break the “every invoice number is unique” assumption which pinned the old false-match rate at zero. The engine was not tuned on this set; the score below is what it actually earns.

Synthetic labelled data, frozen seed (deterministic across page loads), scored in-process on every request against the same engine the table below uses. This measures engine behaviour under adversarial-but-realistic noise - it is not a claim about accuracy on any customer’s production books.

The original seed scores near-perfect - by construction.

The demo seed mutates the register with exactly the transformations each tier was built to reverse - a one-character GSTIN flip for the fuzzy tier, ±₹0.50 inside the ₹1 tolerance band, a one-month slide inside the 3-month window, punctuation the normalizer strips - and every invoice number is globally unique, so false matches are structurally near-impossible. Grading the engine on undoing its own mutations is a tautology. It stays here as a regression harness for the cascade (if any number below moves, a tier broke); Benchmark B above is the one that measures.

Why FN and FP are rupee lines, not ML metrics.

Worked example (assumptions stated, all illustrative): an enterprise with ₹100 Cr of GST-bearing annual spend at a blended 18% rate carries an ITC pool of ≈₹18 Cr a year. Against that pool, the two failure modes Benchmark B counts price out very differently:

How this would price: platform fee per GSTIN per month, plus a success fee only on incremental ITC recovered above the customer’s pre-deployment baseline match rate - measured with the same FN/FP definitions this page uses, so the vendor is paid on the same scorecard the buyer audits. Success-fee pricing is only honest if the benchmark is; that is what Benchmark B is for.

Illustrative model - the ₹100 Cr spend, 18% blended rate and derived figures are stated assumptions, not customer data or market claims. Statutory references (Sec 16(4), Sec 50(3), ASMT-10) are real.