Cited research
Agents read 150,045 BSE filings in full text, 43 news feeds, prices and the desk's own documents. Every number in the writeup names the document and the panel it came from.
No number from a model
v0.1-beta
Scroll for more
[Cursor]
for equity analysts.
Design partner YES Securities
877K events measured
Sitrep is an AI-native operating system for investment teams. Agents research at 100x the speed of an analyst. Every number cites the document it came from. The market scores every call after the fact. The team decides.
BSE filings
NSE prices
GDELT
Google News
IIM-A factors
India VIX
[N.01/11]> What it does
Agents do the work.
The team decides.
Request access
Precedent evidence
Paste a news item. Sitrep returns the similar past events and what each did to the stock, sector and market at +1, +3, +5 and +10 days. The cases that disagree stay in.
Shows the cases that disagreeGraded theses
Each thesis carries a version, a date and a window. Sitrep measures drift as the evidence moves. When the window closes, the market scores the call. The record stands as the rationale.
Graded against the market[N.02/11]> The corpus
877,067 events.
Every one measured.
Request access
877,067
Measured events
150,045
Filings with full text
+1/+3/+5/+10
Day windows, abnormal terms
43 + GDELT
News feeds, titles to 2015
[Macro base rates] Nifty 50 median return after 460 release days, by kind
Day +1Day +3Day +10
[N.03/11]> Core layers
Three layers.
One desk, one record.
Tata Motors JLR wholesale volumes fall 11% in Q2
TMPVAutomobile and Auto ComponentsPress Release / Media Release5 years
After the 20 similar past items (10 of TMPV's own, 10 from Automobile peers), the stock's abnormal return was a median -1.0% by day 3 (30% were up) and -1.8% by day 10. Over the same 3 days the sector did +0.2% and the market -0.1%.
Similar items20all 20 measured
Stock, day 3-1.0%p25 -2.4% · p75 +0.8%
Stock, day 10-1.8%p25 -4.9% · p75 +0.9%
vs expected, day 10.8×|move| over implied
| Date | Headline | D1 | D3 | D5 | D10 |
|---|---|---|---|---|---|
| 2026-01-06 | Shares fall 4% after muted Q3 JLR update | -1.5% | -0.9% | -1.8% | +0.1% |
| 2025-10-07 | JLR posts quarterly volume drop after cyber incident | -2.6% | -5.7% | -47.3% | -46.1% |
| 2025-07-08 | Group global wholesales decline 9% in Q1 FY26 | +1.7% | +0.7% | +0.8% | +3.2% |
| 2025-04-08 | Global sales slip 3% in Q4 FY25 | -0.4% | +1.5% | -1.0% | +4.3% |
| 2024-10-08 | Slides as Q2 global wholesales drop 11% YoY | +0.8% | -0.1% | -1.8% | -2.9% |
Abnormal return over what the market explainsas of 2026-09-11
Works with your data
Add the desk's coverage, positions and internal documents. Connect the market data you already run on. Sitrep reads. It never places an order.
Built to disagree
Review mode reads your thesis and returns three to six challenges, each anchored to a panel row. It takes no side.
Graded after the fact
The market scores every call once its window closes. Event time and ingestion time sit in separate columns, so nobody can backdate a record.
[N.04/11]> How it works
News in. Precedents out.
The desk decides.
Request access
// 001
A filing, a wire story, a broker note, a line pasted from the desk chat. Any source. The resolver names the company and the event class.
A news item comes in
// 002
Same company first, then the same event class across the sector, matched on the text itself, inside the lookback you choose. Each precedent carries its measured reaction, cap band and regime.
Agents find precedents
// 003
Median, IQR, hit rate stated both ways, n. The cases that went the other way stay on the list. The analyst reads it, draws the conclusion and seals it with the evidence attached.
The analyst concludes
// 004
The window closes. The market answers. Sitrep scores the thesis per axis against dated thresholds and files the result next to the sealed evidence. The record is the rationale.
The market grades it
[N.05/11]> Integration
// Add Sitrep to any MCP client: Claude, Cursor, your own agent. { "mcpServers": { "sitrep": { "url": "https://mcp.sitrep.so", "headers": { "Authorization": "Bearer $SITREP_API_KEY" } } } } // 14 tools: precedents, base_rates, thesis, watch, memo, vault ... // Every number the model quotes carries the panel and the document it came from.
from sitrep import Desk desk = Desk(api_key=os.environ["SITREP_API_KEY"]) # One news item in, the precedents out result = await desk.precedents( text="Tata Motors: JLR wholesale volumes down 11% YoY", lookback_years=5, n=20, ) for case in result.cases: print(case.date, case.headline, case.abnormal_5d, case.vs_implied) # Summary across the list: median, IQR, hit rate both ways, n print(result.summary.median_5d, result.summary.hit_rate, result.summary.n)
# Every panel as JSON, with provenance on every number curl -s https://api.sitrep.so/v1/precedents \ -H "Authorization: Bearer $SITREP_API_KEY" \ -d '{ "text": "RBI holds repo at 6.5%, stance unchanged", "subject": "NIFTY", "lookback_years": 8 }' # → { "cases": [...], "summary": { "median_5d": ..., "n": 31 }, # "provenance": { "measured_at": "...", "as_of": "..." } }
14 tools over MCP.
Python and REST too.
Request access
[N.06/11]> Documentation
Read the method.
Check the numbers.
Read the whitepaper
The whitepaper lives at research.sitrep.so.
Method, data and measured results. The cases that went the other way stay in. Every claim names the file it came from.
The method
//001Base rate44,514 events
ClassPress Release / Media Release
Median, day 1-0.10%
Middle half, day 1-2.2% to +2.1%
Up by day 348%
Beyond one expected move45% of 16,611
View
The research
//002Whitepaperresearch.sitrep.so
01Method: measurement, precedents, grading
02Corpus: 150,045 filings, 43 feeds, 460 macro days
03Findings: boilerplate, macro, cashflow
04The seven invariants
05What we got wrong, dated
View
[N.07/11]> Invariants
Seven invariants.
None of them move.
Two clocks
@event_time · @ingestion_time
Sitrep stores when it happened and when we learned it in separate columns. A precedent is only what the desk could have known that morning.
Invariant 01 · enforced in code
Read more
Point-in-time membership
Index and sector membership as they stood that day. Snapshotted by us. No survivorship.
Verified at the door
@ingestion
Ingestion sets the verification status, not the reader. An unverified item stays marked unverified on every panel it appears on.
Invariant 03 · enforced in code
Read more
Numbers from the data
@measurement_layer
Every number on a panel comes from the measurement layer and never from a model. CI fails the build if a model output reaches a number field.
Invariant 06 · CI-enforced
Read more
Grades stored per axis
One grade per axis, against a dated threshold. Re-read it years later and it means the same thing.
No order capability
@anywhere
Nothing in the system can place, route or size an order. There is no target price and no drafted view. The precedent list is the record.
Invariant 05 · by design
Read more
The analyst concludes
@the_desk
Sitrep retrieves and measures. The reading, the call and the responsibility stay with the person whose name is on the note.
Invariant 07 · the whole product
Read more
Invariant 06 · live
Try to put a model's number on a panel.
Every number on a panel comes from the measurement layer. A model output that reaches a number field fails the build. Flip the switch and watch the check run.
[N.08/11]> Pricing
Three plans.
Priced per seat.
- Includes:
- Precedent lookups, every NSE-listed company
- Cited research writeups
- Versioned theses, graded
- Email support
- Everything in Analyst, plus:
- Shared coverage, vault and watches
- Dated memos and desk notes
- MCP, Python and REST access
- Adversarial review mode
- Priority support
- Everything in Desk, plus:
- Your positions, mandate and documents
- Your market data connections
- Single tenant, read-only access
- Frozen as-of mode for audit
- A forward-deployed engineer
[N.09/11]> Changelog
Every change, dated.
Nothing backdated.
View all
Every change is a dated commit.
What we measured, what moved and what did not, in the order it happened.
Sep 14, 2026
Whitepaper live
The method, the corpus and the measured findings, as one static site. Every claim names the file it came from.
Sep 11, 2026
Boilerplate stripping, measured
Every BSE filing opens with the same covering letter. We stripped it before embedding and re-judged the company list: 59.3% of rows relevant before, 64.0% after. Not shipped yet.
Sep 09, 2026
Macro mode, similarity floor
460 dated macro events, CPI, FOMC, RBI, Budget, GST Council, OPEC+ and elections, with the Nifty's path after each. Cross-company lists now stop at a similarity floor instead of padding.
[N.10/11]> Research
Research notes.
Measured, not argued.
Read all
[N.11/11]> FAQs
Five questions.
Plain answers.
No. Sitrep returns the precedents and the measurements: what similar past events did to the stock, its sector and the market, and the summary across them. The analyst draws the conclusion. Sitrep drafts no view, sets no target price and holds no order capability anywhere. The precedent list is the rationale record.
From the measurement layer, never from a model. Sitrep measures each reaction in abnormal terms against the sector and the market at +1, +3, +5 and +10 days, and as a multiple of what options implied that day. A CI check fails the build if a model output reaches a number field. When the model quotes a number, it cites the panel and the document it read it from.
India first. Every NSE-listed company: 150,045 BSE filings with full text, adjusted prices, IIM-A factors, F&O implied moves, 43 news feeds plus GDELT and Google News with titles back to 2015, and 460 dated macro events for the index itself. Any market with an exchange feed comes next.
Yes. Add the desk's coverage, positions, mandate and internal documents, and connect the market data you already run on. Sitrep reads; it holds no order capability, so it cannot act on your book. The Enterprise plan runs single tenant with a frozen as-of mode for audit.
You seal a thesis with its evidence and a window. When the window closes, the market answers and Sitrep scores the thesis per axis against dated thresholds, so a grade re-read years later means the same thing. Event time and ingestion time sit in separate columns, so nobody can rehearse or backdate a record.