A leverage ratio of 7.3x says more than every Jane Street news article combined. The methodology that surfaced it, and an open repo to reproduce the work.
In 2024, Bill Hwang lost $20 billion, and the industry learned about it from the news. Not from a 13F. Not from a 13D. Not from Form SHO. Archegos never filed a 13F - its positions in ViacomCBS, Discovery, Tencent Music were structured through total return swaps, which the regulatory framework doesn't treat as "holdings." That's the ceiling of what can be hidden inside the law.
Before decoding Jane Street - where SEBI put out 105 pages of material Jane Street would much prefer to keep private - we need to agree on what's visible and invisible in public data at all. Otherwise, any LLM pipeline will build you a beautifully written hallucination.
This article is about the Jane Street vs SEBI case and the methodology that decoded it in an evening instead of a senior analyst-week. Methodology implemented as an open repo with six techniques, a monitoring layer, and pattern propagation. Link at the end.
Part 1. What 13Fs and public documents don't show
If your analyst still thinks a 13F is a portfolio snapshot, fire the analyst. A 13F is a filtered, lagged projection of a portfolio:
- Long positions only, in Section 13(f) securities. No shorts. No swaps. No fixed income. Options reported as notional only - no strikes, no direction
- 45-day lag. By the time you see Q3 holdings, the fund may already be in Q4
- Confidential treatment requests. Berkshire routinely hides building positions for 3-12 months; Agarwal et al. in the \Journal of Finance\ showed empirically that confidential holdings generate systematically higher alpha than disclosed ones
- Window dressing. Rotating into "respectable" positions toward quarter-end - well-documented academically, invisible by construction in 13F
- Form SHO (which would close part of the short loophole) pushed to February 2028
Any methodology starts with an explicit list of what's structurally impossible to extract. Otherwise, you're building a model of the fund's public mask, not the fund.
The good news: the mask has to be consistent. To hide a strategy, a fund needs costly actions every quarter - confidential treatment requests, swap restructuring, cross-jurisdictional allocations. These leave second-order traces: language shifts in Form ADV between years, hiring patterns on LinkedIn, patent filings in USPTO, academic publications by employees, regulatory documents from other jurisdictions.
LLMs don't find new sources. The sources are the same. What changes is the cost of cross-reference. Connecting a USPTO patent to an academic paper by the same author, who a year ago spoke at a conference also attended by an ex-employee of the fund - that used to be a senior analyst's full workday. With Claude plus prompt caching: $2 in API credits and 15 minutes.
On Opus 4.7, cache reads cost $0.50 per million tokens instead of $5. Load the 105-page PDF into cache once, then run hundreds of prompts at pennies each. The workflow shifts from "one expensive analysis" to "a thousand cheap questions."
Part 2. Jane Street vs SEBI
Timeline:
- April 2024: Citadel sues Jane Street in NY for trade-secret misappropriation by two ex-employees. The complaint references India options strategies - which tipped SEBI off
- February 2025: NSE issues an explicit advisory to Jane Street
- July 3, 2025: SEBI publishes its 105-page ex-parte interim order, freezes ₹4,843 crore ($565M)
- July 21, 2025: ban lifted after escrow deposit
- September 2025: Securities Appellate Tribunal pauses personal hearings, demands SEBI disclose more documents
- January 2026: investigation expands to Sensex and other strategies. No final order yet
Profit SEBI attributes to Jane Street over January 2023 - March 2025 in Indian options: ₹43,289 crore (~$5B). The frozen $565M is only "unlawful gains" from specific trading days.
Jane Street's defense: "basic index arbitrage, retail demand facilitation." The line between market making and manipulation in derivatives is genuinely thin. But 18 expiry days with a nearly identical pattern, persisting after NSE's explicit advisory - a harder defense.
What follows: how to reach the same conclusions yourself from public documents.
Part 3. Six techniques
Most LLM-for-finance tutorials say "load the PDF, ask for a summary." Wasted context.
3.1. Hostile hypothesis testing
The basic mistake - asking the LLM to "explain the strategy." You get a polished narrative in which Claude used your leading questions as rails.
The correct pattern is inverted. You formulate the hypothesis; Claude gets the role of hostile reviewer:
LLMs are good as proof-readers, bad as proof-generators. Use them in "find holes in my thinking" mode, not "think for me" mode.
3.2. Ensemble of competing hypotheses
What if the right hypothesis isn't the one you formulated? Run multiple Claude instances in parallel with different interpretive frames: "this is manipulation," "this is legitimate market making," "this is something else." A meta-prompt compares the three outputs and identifies points of divergence. Where all three frames converge on the same facts but reach different conclusions - that's where the data is insufficient. Diagnosing your own unknowns, which in research is worth more than any single answer.
3.3. Temporal differencing
Same document type (Form ADV is the most productive target) from the same firm in different years. Ask Claude to find language changes, especially in boilerplate. Most funds copy ADV verbatim year to year; when something changes in risk management language, it's signal that compliance decided the old wording no longer protects them.
The most interesting findings: where old language disappears. A removed statement almost always means it's no longer true.
3.4. Cross-document triangulation
One document is a narrative. Three documents are triangulation. Claim from source A (SEBI order), confirmation in B (academic literature on expiry-day microstructure), contradiction in C (Jane Street internal communications). Claude runs the intersection against a hundred claims in one session with confidence scores. Claims passing all three checks - operating ground truth. Claims passing A and contradicted by B and C - where the regulator overreached.
3.5. Prompt caching
A 105-page SEBI order is ~120k tokens. First load with 1-hour cache: ~$1.20. Each subsequent query: ~$0.06. After 50 queries - $4 and 50 different analytical cross-sections.
What changes isn't cost, it's the structure of the work: from "three questions per document" to "three hundred."
3.6. Adversarial document interrogation
Document from party A (SEBI order). Feed it to Claude and ask it to play A's adversary - Jane Street's defense team - and construct targeted questions that would undermine specific factual claims. The resulting list is a forecast of how the defense will be structured. The final order may be a year away, appeals two more - but the map of potential vulnerabilities is in your hands today.
Part 4. How to defeat hallucinations
Without this, the six techniques become beautifully written bullshit. Three rules, enforced in code via Pydantic:
Provenance chain enforced. Every claim carries a full evidence trail: \source_url\, \source_document_hash\, \source_page\, \extraction_timestamp\, \extraction_model\, \verification_status\. Without a complete trail - Pydantic rejects.
Verbatim quote verification. Every quote from the LLM is programmatically string-matched against the source document. Not found - \verification_status="failed"\, claim discarded.
Ensemble voting. Every enrichment runs N times with different seeds (default 3). Claims with 2/3+ agreement get high confidence. Disagreement is visible in the final output.
The confidence formula in code:
Every weight has a required \weight_justification\ field. Default scheme: 0.4 for structural markers, 0.3 for numerical thresholds, 0.2 for regex patterns, 0.1 for LLM semantic extraction (the lowest, because most subjective).
On predictions - mandatory \adversarial_analysis\ and \falsification_criteria\. Pydantic rejects empty values.
LLM prompt engineering is ten percent of the work. Ninety percent is the code around the LLM that catches and discards what the LLM made up.
Part 5. Can you trust regulators
Natural objection: if the methodology is built on regulatory documents, and regulators are demonstrably biased - what holds the trust?
Validity of the critique. The revolving door between SEC and industry is documented (Mary Jo White, Jay Clayton). Selective enforcement is real - Madoff wasn't caught for 16 years despite Markopolos's whistleblower reports in 2000, 2005, 2007. After 2008, zero top executives of major banks received criminal prosecutions vs 1,000+ convictions after the S&L crisis of the 1980s. The SEC enforcement division is 1,300 people against an industry of millions.
Why this isn't relevant to the methodology.
Bias works in one direction. A regulator may decline to file a case. But publishing a 105-page order is a different category of decision. At that point, the regulator's risk is overstating, not understating: the opponent has the resources to dismantle weak arguments in court.
SEBI ≠ SEC. SEBI against a US firm - the politics is reversed: demonstrate sovereignty, defend domestic retail. Cross-jurisdictional enforcement is systematically harsher than domestic.
Numbers vs narrative. The SEBI order has narrative ("manipulation," "sinister scheme") and a factual section with specific numbers (Tables 7, 8, 16, 17). Narrative can be spun. Numbers come from exchange records and broker reports that Jane Street itself submitted into Indian regulatory systems. Falsifying numbers is criminal liability for regulator employees.
Adversarial validation. Jane Street is actively fighting. Top-tier lawyers. Filed an appeal. They contest interpretation ("this is not manipulation, it's arbitrage"), not facts ("we did not buy that much underlying"). If the numbers were fabricated, that would be their first argument at SAT. The adversarial system is doing its job.
What the methodology does with this. Separation by claim type: factual claims → trust (verifiable, adversarially validated); causal/intent claims → triangulate against other sources; legal characterizations → not our domain, wait for the final order. The adversarial document interrogation (Part 3.6) specifically buys the defense perspective from the document itself.
Regulators aren't neutral. But published enforcement orders go through an adversarial filter that makes their factual claims more robust than most alternative sources. Trust the numbers, not the narrative.
Part 6. Quant mechanics: the order book on January 17, 2024
"Jane Street held 20%+ of daily traded value" sounds like a smoking gun - on its own, it proves nothing.
BANKNIFTY is an index of the 12 largest Indian banks. Index options are cash-settled against the weighted average of the last 30 minutes of constituent trading. A market maker holding a large gamma position at that moment is required to hedge through cash and futures.
A market maker's hedge technically moves the market. If your options position requires selling $200M of underlying in 30 minutes, your own flow creates downward pressure. That's physics, not intent.
Numbers from the order
SEBI documents January 17, 2024 as the most illustrative day. Numbers from pages 25-40 (Tables 7, 8, 16, 17):
Part 7. What these techniques give you today
We decoded a 2025 order about 2023-2024 events - what does that give you now? If the only output were a reconstruction of an obsolete strategy, this would be methodology for journalists, not traders.
Each technique is a pattern detector that runs on today's data. Jane Street is a teaching case. Five applications to live data:
1. Leverage 7.3x as a detector signature. Any market maker observable through public data (combined 13F + Options Clearing Corp data + venue volume reports) running options notional / underlying position ratios above 5x is a Jane Street-pre-2025 setup. Active monitoring across the 20-30 largest HFT firms provides early warning for: (a) short volatility on ETFs with exposure if a regulator starts moving, (b) capacity opportunities if a firm exits after enforcement.
2. LTP impact analysis as due diligence. SEBI's method is now part of the regulatory toolkit. Any allocator after July 2025 should be asking for LTP impact metrics in the DDQ. If your fund runs market making - build this monitoring internally now.
3. Asymmetric P&L pattern for real-time screening. For every multi-leg strategy in your portfolio attribution, check for systematic loss-leader patterns between legs. If they exist - compliance risk regardless of intent.
4. Indian options as an actionable structural insight. 61% of global equity options by volume trade in India (April 2025 data). Jane Street withdrew - the liquidity vacuum is being distributed across Tower, Citadel Securities, Optiver, IMC. An open window for 6-18 months, after which SEBI is likely to introduce participation caps. Live trades: capacity build in India, short positions on firms with outsized India revenue without diversification.
5. JSI/JSI2/Singapore/Hong Kong as a regulatory arbitrage template. This structure was used by at least a dozen foreign HFT firms. SEBI is establishing precedent. For any foreign HFT you have exposure to (direct or via ETF), check public filings for DTAA routing. If present - adjust for regulatory risk premium upward.
Each is a historical pattern converted into a forward-looking screening criterion. The methodology for identifying (a) leveraged options structures, (b) aggressive LTP-impact behavior, (c) asymmetric PnL fingerprints, (d) opportunities in structurally distorted markets, (e) regulatory arbitrage architectures - live tools, applicable to today's 8-Ks, 13Ds, ADVs.
Part 8. What's in the repo
If each technique works in one-shot mode, nothing prevents converting it to continuous monitoring via regulator RSS feeds, USPTO patent assignments, court filing APIs. And further - extracting recurring structural patterns from analyzed cases and matching them to other funds. Architecturally, the same codebase. Three commits:
Commit 1: Platform. Six techniques in \src/reverse_quant/techniques/\. CachedCorpus. Anthropic client wrapper with retry and cost tracking. EDGAR fetcher for 13F and ADV. Notebook 01: end-to-end on the SEBI document - the analysis from this article, runnable for $4-8.
Commit 2: Monitoring layer. \monitors/\ - EDGAR RSS poller (working), USPTO/PACER/regulator press stubs. \pipelines/\ - triage/enrichment/dedup/orchestrator. \alerts/\ - stdout/file working, Slack/email stubs. SQLite storage with full provenance schema. Notebook 04 shows how verification catches hallucinated claims.
Commit 3: Pattern propagation. \patterns/\ - extraction/library/matching/prediction. Three seed patterns from the Jane Street case, hand-curated, marked \reviewed_by_human=True\. Transparent confidence formula in code. Mandatory adversarial analysis and falsification criteria on every prediction. Notebooks 05-06.
Stack: Python 3.11+, type hints, ruff/mypy clean, tests with mocked LLM clients, SQLAlchemy, Pydantic v2. MIT license.
Download the SEBI order via the link in \data/README.md\; the run costs ~$4-8.
What you can't get with even the best LLM
- Tick data for direct verification of SEBI claims isn't publicly available
- Internal compliance logs of Jane Street are absent - you don't know what the firm saw in real time
- Intent - an LLM can't distinguish manipulation from market making, requires discovery
- Numerical verification - Claude confused crore with millions of dollars at least once per ten runs. Every number gets verified by hand
- What happens next - the final order could exonerate, condemn, or partially confirm
Final
A regulator's document is the most information-dense artifact a fund leaves in public. Not because it reveals the strategy (it reveals less than a 13F), but because it's the only document not written by the fund and not by its marketing. SEC, SEBI, FCA - these people have subpoena power, and they write for courts, not for the public.
In 2020, decoding the structure of an average hedge fund cost $50-100k and 2-3 months of senior analyst time. In 2026 - $50-100 and an evening. A new equilibrium, in which funds aware their public footprint is being analyzed this way will start controlling it more carefully - creating the next round of the game.
If you work at a fund or allocator interested in internal use - closed consulting is available.
