Skip to content
ARCHYTASAI

ENV-FIN-101 · Investment Research & Valuation

Research & Valuation. Produce reliable institutional-quality analysis, not just extracted figures or a convincing investment narrative.

A financial research workspace where agents investigate companies across filings, financial statements, earnings materials and spreadsheets, reconcile conflicting information, build and repair models, and reach defensible valuations.

Custom buildFundamental Research · Financial Modeling · Valuation

Overview. What the environment trains, and the work it's built around.

Core competency

Produce reliable institutional-quality analysis, not just extracted figures or a convincing investment narrative.

  • Retrieving figures with source attribution, using only what was available at the research cutoff
  • Normalizing statements and assessing quality of earnings
  • Building and repairing linked three-statement models and DCF valuations
  • Testing an investment thesis against counter-evidence

Built around the work of

Equity Research Analyst · Investment Banking Analyst · Private Equity Associate

Simulated systems

  • Financial data terminal
  • Document repository
  • Modeling workspace
  • Research registry

Sample task

You have been given a financial model for a proposed acquisition of Aster Manufacturing. The model indicates that the target is attractively valued, but several assumptions are inconsistent with the source financial statements. Review the supporting documents, identify material inconsistencies, correct the model, and produce a revised valuation with a concise explanation of the changes. Use only documents available as of the specified research cutoff date.

Specification. The environment, field by field.

spec · env/research-valuation.yamlENV-FIN-101
product_type:
Executable RL training and evaluation environment
category:
Fundamental Research · Financial Modeling · Valuation
commercial_model:
Environment development and validated task packages
environment_id:
ENV-FIN-101
target_roles:
Equity Research Analyst · Investment Banking Analyst · Private Equity Associate
core_competencies:
  • Retrieving figures with source attribution, using only what was available at the research cutoff
  • Normalizing statements and assessing quality of earnings
  • Building and repairing linked three-statement models and DCF valuations
  • Testing an investment thesis against counter-evidence
data:
Synthetic issuers with filings, earnings materials, transcripts and market data on a point-in-time clock. Licensed historical data can be added where an engagement allows it.
simulated_systems:
Financial data terminal · Document repository · Modeling workspace · Research registry
interfaces:
Structured research tool API · Spreadsheet workspace · Python for calculations · Document search
interaction:
Stateful, multistep tool use across documents, data services and a spreadsheet model, with an artifact submitted at the end.
state:
A versioned document store and modeling workspace on a point-in-time clock. Each episode starts from an isolated, seeded snapshot.
disturbances:
Documents published after the cutoff, restatements, inconsistent figures across sources, broken formulas and assumptions that contradict the filings.
task_families:
  • TASK-FIN-101-01 · Point-in-time company research
  • TASK-FIN-101-02 · Statement normalization & quality of earnings
  • TASK-FIN-101-03 · Financial modeling & valuation repair
  • TASK-FIN-101-04 · Investment thesis & counter-evidence audit
evaluation:
Source figures, adjustments, model linkages and valuation math recomputed by the verifier, plus expert rubrics for the written conclusion.
reward:
Deterministic checks on figures, citations and model math, combined with rubric scores where professional judgment applies.
training_use:
RL training, held-out evaluation and SFT on reference research artifacts.
deployment:
Containerized, isolated execution, with a customer-controlled deployment option.
licensing:
Commercial use rights. Exclusivity, ownership and any third-party data licenses are agreed for each engagement.

Workflow coverage. What agents do, and what the verifier checks.

  • Workflow

    Company research

    Example tasks

    Compute trailing figures from filings and releases available at a cutoff date.

    What is verified

    Figures match permitted sources, each is cited, and nothing post-dates the cutoff.

  • Workflow

    Earnings quality

    Example tasks

    Normalize reported EBITDA for non-recurring and owner-related items.

    What is verified

    Each adjustment is supported by a document and computed correctly.

  • Workflow

    Model repair

    Example tasks

    Find and fix errors in a DCF and its statement linkages.

    What is verified

    Linkages hold, the valuation recomputes, and every fix is explained.

  • Workflow

    Thesis review

    Example tasks

    Test a stated thesis against the evidence and report what weakens it.

    What is verified

    Counter-evidence is found and quantified, and the conclusion follows from it.

Task families. Each family has its own datasheet: prompt, verification criteria and failure modes.

Agent tools. What the agent can call.

The agent searches filings and transcripts, queries financial data on a point-in-time clock, works in a spreadsheet model and submits a research artifact for grading.

tools · env/research-valuationJSON-schema tools
  • search_filings
  • get_financial_statements
  • get_earnings_transcript
  • query_financial_metrics
  • open_spreadsheet
  • run_financial_calculation
  • save_model
  • submit_research_artifact

Verification methodology. Checked against state the verifier computes itself.

Source figures against permitted records, evidence for every adjustment, formula and statement linkages, reproducible valuation math, citations and the research cutoff, and whether the recommendation reflects the model's assumptions and uncertainty.

The verifier checks

  1. 01Every source figure matches a permitted record available at the research cutoff.
  2. 02Every adjustment is supported by a cited document.
  3. 03Model formulas and three-statement linkages hold.
  4. 04The valuation recomputes from the submitted model and the stated assumptions.
  5. 05Citations resolve to the documents and dates claimed.
  6. 06The recommendation is consistent with the model's assumptions and states its uncertainty.
  7. 07Reference solutions and adversarial cases run against the verifier before training use.

Reward

Valuation allows professional judgment. A different discount rate or growth forecast isn't an error if it's defensible and consistent with the task. Invalid calculations are.

Scores zero on

  • Using documents published after the research cutoff
  • Source figures that don't match the permitted records
  • A valuation that can't be reproduced from the model
  • A recommendation the model's assumptions don't support

Included deliverables. What you receive, and what we agree before we build.

Delivered with the environment

  • Containerized environment with point-in-time document and data stores
  • Validated task instances with reference research artifacts
  • Verifier code and expert rubrics
  • Agent traces and failure-mode analysis across repeated runs
  • Technical datasheet and integration documentation

Agreed before development

  • Company universe and document types
  • Research cutoff conventions
  • Valuation methods accepted and tolerance thresholds
  • Rubrics for judgment-based conclusions
  • Any third-party data licenses

Every environment is built to order. Tell us which workflow you need to train or evaluate, and where you are in the process.

Request sample tasksDiscuss custom developmentRequest technical integration detailsSee ways to buy