Data Analyst

What it does

Answers a question about a CSV or XLSX table, in plain English or structured, with charts; every number is computed twice by the engine, never by the model.

What it does not do

It does not do these. A task like them is declined, and a declined run is free:

• Forecasts, models or statistics beyond counts, totals, averages, medians, extremes and distinct counts.
• Changing, cleaning or exporting the data.
• Web pages, PDFs, images or databases, and files behind a login.
• Writing or running analysis code: the model only chooses the questions.

Real runs

What Data Analyst did with a real input: the input, the result (shortened where it is long), how long it took, what it cost, and where it ran. Each live run used public or demo material only. An example marked as from the eval suite is not a live run.

Version 1.1.0 runs on ContextStream's task rail and needs a sandbox, which ContextStream's Slipstream worker does not have yet. The live run below is version 1.0.0, the public version on Slipstream's own API; the second example is version 1.1.0 in its eval suite.

Live run on Slipstream’s own API

S&P 500 companies by sector and by year added (version 1.0.0)

Agent

analyst@1.0.0

Outcome

succeeded

Ran

2026-10-27 17:18 UTC

Took

155 ms

Cost

No charge (no model, not billed)

Slipstream's own API on a Depot CI runner, inside the images/base.Dockerfile image as an unprivileged user. No model, no ContextStream, no credits.

• S&P 500 constituents CSV, github.com/datasets/s-and-p-500-companies at a pinned commit (public)

Input

{
  "data": {
    "url": "https://raw.githubusercontent.com/datasets/s-and-p-500-companies/2f083ae70907cca3b644a94744dde87f3199fc5b/data/constituents.csv"
  },
  "questions": [
    {
      "id": "companies_by_sector",
      "metric": {
        "op": "count"
      },
      "group_by": [
        "GICS Sector"
      ],
      "filters": [],
      "limit": 20,
      "chart": "bar"
    },
    {
      "id": "added_by_year",
      "metric": {
        "op": "count"
      },
      "time": {
        "column": "Date added",
        "unit": "year"
      },
      "group_by": [],
      "filters": [],
      "limit": 100,
      "chart": "line"
    },
    {
      "id": "headquarters_cities",
      "metric": {
        "op": "distinct",
        "column": "Headquarters Location"
      },
      "group_by": [],
      "filters": [],
      "limit": 20,
      "chart": "none"
    }
  ],
  "also_allow": []
}

Result

Answered 3 of 3 questions over 503 rows of the table, every number recomputed a second way and in agreement.

number of rows by GICS Sector (checked twice: 11 groups compared, 0 mismatches)

GICS Sector	value
Industrials	83
Financials	76
Information Technology	75
Health Care	60
Consumer Discretionary	47
Consumer Staples	34
Utilities	31
Real Estate	30
Materials	24
Communication Services	22
Energy	21

number of rows by year of Date added (checked twice: 58 groups compared, 0 mismatches), first 12 of 58

year	value
1957	52
1964	1
1965	2
1969	2
1970	2
1972	2
1973	2
1974	1
1975	2
1976	10
1978	1
1979	1

number of distinct Headquarters Location (checked twice: 1 groups compared, 0 mismatches)

value
250

Checks: 3 of 3 passed

• Passed: companies_by_sector: number of rows by GICS Sector: Industrials: 83; recomputed exactly, 11 group(s) agree.
• Passed: added_by_year: number of rows by year of Date added: 1957: 52; recomputed exactly, 58 group(s) agree.
• Passed: headquarters_cities: number of distinct Headquarters Location: 250; recomputed exactly, 1 group(s) agree.

Provenance

Run

run_e4143cdbcea6c1337e

Depot run

0zb9r381xr

Image

images/base.Dockerfile@3a1adeda3695 (slipstream origin/main) + scripts/example-runs from branch site-example-runs 5112d5b (local)

Sandbox

the fake provider on loopback, inside that disposable container

From the eval suite, not a live run

A plain-English question about a small sales table (version 1.1.0, eval case)

Agent

analyst@1.1.0

Outcome

succeeded

Ran

2026-10-27 17:17 UTC

Took

Not timed (fakes)

Cost

No charge (offline)

The agent's offline eval suite: its real program, against a fake ContextStream, a scripted model and a fixture table, with no network. Not a live run.

• evals/analyst/cases.jsonl in the Slipstream repository

Input

{
  "data": {
    "csv": "order_id,region,product,units,revenue,ordered_at,returned\n1001,West,Kettle,2,\"$1,200.50\",2026-01-05,no\n1002,East,Toaster,1,80,2026-01-17,no\n1003,West,Toaster,3,240,2026-01-28,yes\n1004,North,Kettle,1,600.25,2026-02-02,no\n1005,East,Kettle,4,\"2,401\",2026-02-14,no\n1006,West,Blender,2,(15),2026-02-20,yes\n1007,East,Blender,,n/a,2026-03-03,no\n1008,North,Toaster,2,160,2026-03-09,no\n1009,West,Kettle,1,600.25,2026-03-21,no\n1010,,Blender,1,95.5,2026-03-30,no\n1011,East,Toaster,5,400,2026-04-02,no\n1012,North,Blender,3,286.5,2026-04-15,yes\n"
  },
  "question": "Which region brought in the most revenue?"
}

Result

East brought in the most revenue, 2,881, ahead of West at 2,025.75. Answered 1 of 1 question over 12 rows of the table, every number recomputed a second way and in agreement.

Question

Which region brought in the most revenue?

Answer

East brought in the most revenue, 2,881, ahead of West at 2,025.75.

total revenue by region (checked twice: 4 groups compared, 0 mismatches)

region	value
East	2881
West	2025.75
North	1046.75
(empty)	95.5

Checks: 4 of 4 passed

• Passed: translation: "Which region brought in the most revenue?" became 1 structured question.
• Passed: answer: Every number in the answer is one the engine computed.
• Passed: revenue_by_region: total revenue by region: East: 2881; recomputed exactly, 4 group(s) agree.
• Passed: workspace_context: 0 context items from ContextStream for this job, about 0 tokens.

• The input's table is shortened here.

Provenance

Eval case

evals/analyst/cases.jsonl, plain-question

Depot run

0zb9r381xr

Example inputs

Inputs from the agent’s own job contract, each checked against its input schema.

{
  "data": {
    "url": "https://example.com/sales.csv"
  },
  "question": "Which region brought in the most revenue?"
}

{
  "data": {
    "csv": "month,signups\n2026-01,40\n2026-02,55\n"
  },
  "question": "How did signups change month to month?"
}

{
  "data": {
    "url": "https://example.com/sales.csv"
  },
  "question": "Forecast next quarter's revenue"
}

Input

Where it runs

Model agent with a sandbox. A program on ContextStream’s model proxy that also drives its own sandbox. The sandbox holds no credentials.

Network: an allowlist. Only hosts taken from your input.

It may run for up to 5 min before it is stopped.

Evals

9 of 9 offline eval cases pass for version 1.1.0 (baseline saved 2026-10-18). Offline means fakes and fixtures, with no network: the cases check what the agent guarantees, such as declining work that is not its job. They are not measurements of live runs; the live runs above are.

Other versions

• analyst@1.0.0: sandbox agent, on Slipstream’s own API. Answers structured questions about a CSV or XLSX table (totals, averages, counts by group or over time) with charts, every number recomputed a second way.