PRISM reads tax statutes clause by clause, ties every number to the words it came from, runs the law on the population it governs, and checks the reading against the law itself.
It obliges, penalises, grants a right, sets a threshold, names an actor or a beneficiary. PRISM separates them the way a prism separates light. In the Income-tax Bill, 2025:
Each stack here is one of them, as tall as its sections; point at one to open it. Parliament amends them every year.
To know what a budget does to a salary, someone has to turn sentences back into numbers: slabs, rebates, thresholds, reliefs. By hand, for every rule, every year. And amendments rarely restate a rule. They swap words inside the old one.
Language models will happily fill in the numbers. PRISM asks the question that matters: do the numbers they read lead to the same conclusions as the law? When they don't, which clause is to blame?
The PDF becomes text with exact character offsets and its place on the page, so anything PRISM extracts can be traced back to the words that say it.
A statute is cut the way lawyers cite it: chapter, section, sub-section, clause, proviso. The tree here is the consolidated Income-tax Act 2025; the bright branch is the rebate.
Each clause is tagged for who must act, what they must do, what they may claim, what it costs them and where the lines are drawn. The prism passes over the clause and the tags light up.
Conditions are linked to their consequences and rated by risk: a missed condition behind a penalty matters more than one behind a definition. Behind the clause, every pattern in the Act.
Document, chapter, clause, pattern, entity: one graph, so any figure PRISM produces can be walked back to its words. Point at a clause in the web to open it.
Each point is a group of taxpayers, built from the income-tax department's count of returns by income range. The lines mark tenths of the population; teal points pay less under the AY 2026-27 law than under the year before.
The same engine turns the coded law into a schedule: what every salary pays under the new regime for AY 2026-27. The dim line is the year before.
Questions are answered from the corpus, with the sections that answer them. One benchmark question, “Penalty for under-reporting and misreporting of income”: the right section, s.439, comes back at rank 1.
Section 20 of the Finance Act 2025 raises the rebate by swapping words inside section 87A. Qwen 3.8 27B read it: income up to ₹12,00,000, a rebate of at most ₹60,000.
Clause (b) is the marginal relief. The amendment only changes its numbers, so the words never appear on the page.
Across the population, 1.7 million returns would be taxed differently, and the reform's revenue cost would be off by ₹4,077 crore.
10 income-tax reforms, from the Finance Act 2020 to the Income-tax Act 2025, read by 6 readers, from a 120-billion-parameter model to hand-written rules. A reading counts only if it assembles into a complete schedule.
The torn curve did not change the policy story: the same deciles gain, the cut stays progressive. The small local models assembled none of the 20 readings they attempted. Point at a tile to see its run.
The kinds of slip a reader makes: a dropped slab, last year's table, a lakh read as a crore.
Rebate dropped flipped up to 65% of conclusions. Bound off by one rupee flipped 0%. Each column is one kind of error; each spark, one injection. Point at a column for its numbers.
In an agent-based model of compliance, the share of tax evaded settles near 0.33. Reading the 200% penalty as 50% raises it by 0.071 on average; the penalty rate is the second most important input. 2 readers found the penalty; 4 found none.
The words a model cites as evidence are no more faithful than random words: comprehensiveness 0.72 for its spans, 0.77 for random ones.
Recall at five of 0.72 over 339 questions; unanswerable ones separate with an AUROC of 0.85.
In sample the population model lands within 9.4% of reported tax, but its forecast to 2023-24 scored worse than carrying the last year forward (skill -4.26).
The app's agent-based policy simulator is illustrative, labelled as a sandbox, and none of the results above use it.
20. In section 87A of the Income-tax Act, with effect from the 1st April, 2026,–– (a) in the proviso,— (i) in clause (a),–– (I) for the words “seven hundred thousand rupees”, the words “twelve hundred thousand rupees” shall be substituted; (II) for the words “twenty-five thousand rupees”, the words “sixty thousand rupees” shall be substituted; (ii) in clause (b), for the words “seven hundred thousand rupees” at both the places where they occur, the words “twelve hundred thousand rupees” shall be substituted; (b) after the proviso, the following proviso shall be inserted, namely:–– “Provided further that the deduction under the first proviso, shall not exceed the amount of income-tax payable as per the rates provided in sub-section (1A) of section 115BAC.”.
| Field | Model read | The law |
|---|---|---|
| Income limit | ₹12,00,000 ✓ | ₹12,00,000 |
| Maximum rebate | ₹60,000 ✓ | ₹60,000 |
| Marginal relief | false ✕ | true |
| Decile | Previous year's law | AY 2026-27 law | Qwen 3.8 27B's reading |
|---|---|---|---|
| 1 | 0.00% | 0.00% | 0.00% |
| 2 | 0.00% | 0.00% | 0.00% |
| 3 | 0.00% | 0.00% | 0.00% |
| 4 | 0.00% | 0.00% | 0.00% |
| 5 | 0.00% | 0.00% | 0.00% |
| 6 | 1.74% | 0.00% | 0.00% |
| 7 | 4.03% | 0.00% | 0.00% |
| 8 | 5.32% | 0.00% | 0.00% |
| 9 | 8.98% | 5.63% | 6.00% |
| 10 | 26.07% | 23.62% | 23.62% |
| Reform | gpt-oss-120b | Qwen 3.8 27B | gpt-oss-20b | Phi-3.5 mini | Gemma 2 2B | Rule-based |
|---|---|---|---|---|---|---|
| FA 2020 · new regime | 0 flips | 0 flips | no conclusion | no conclusion | no conclusion | 0 flips |
| FA 2020 · old schedule | 0 flips | 0 flips | 0 flips | no conclusion | no conclusion | 0 flips |
| FA 2023 · new default | 0 flips | no conclusion | 0 flips | no conclusion | no conclusion | no conclusion |
| FA 2024 (No. 2) · budget | no conclusion | 0 flips | 0 flips | no conclusion | no conclusion | no conclusion |
| FA 2025 · rebate ₹12L | 0 flips | 0 flips | 0 flips | no conclusion | no conclusion | no conclusion |
| FA 2023 · old schedule | 0 flips | 0 flips | 0 flips | no conclusion | no conclusion | 0 flips |
| FA 2024 (No. 2) · old schedule | 0 flips | 0 flips | no conclusion | no conclusion | no conclusion | 0 flips |
| FA 2025 · old schedule | no conclusion | 0 flips | no conclusion | no conclusion | no conclusion | no conclusion |
| FA 2026 · old schedule under ITA 2025 | no conclusion | no conclusion | no conclusion | no conclusion | no conclusion | no conclusion |
| ITA 2025 · new act | no conclusion | no conclusion | no conclusion | no conclusion | no conclusion | no conclusion |
Pipeline counts are from the demo document: 624 pages, 2,757 clauses, 9,063 entities, 755 cause–effect patterns. Paper numbers: 10 reforms.
Reading the law, one clause at a time.