← FIELD applications

FIELD · Agents and workflows · 4 min read

Financial research with inspectable evidence

One document corpus and defined issuer, currency, period and as-of semantics.

Arunank Sharan · Kuzushi Labs ·

In this article

This worked application uses explicit planning assumptions to demonstrate FIELD. Quantities below are inputs, calculated implications or proposed targets; they are not reported customer results. The engagement decision remains conditional on the evidence described.

Planning basis and engagement boundary

ItemBasis for this application
Demand and context500 research tasks/day; 20 source passages × 400 tokens per task
First delivery boundaryOne document corpus and defined issuer, currency, period and as-of semantics
Acceptance to agree and testAdjudicate a proposed 200-task set for supported claims, calculations and abstention
Exclusion or constraintThe service supplies research evidence; trading and payment actions are excluded.

F — Frame the outcome

Help analysts produce defensible financial comparisons faster, with source, period and unit correctness as part of the outcome.

Value measure: Time to an accepted sourced comparison and material numerical/citation error rate.

I — Investigate the workflow

Observe how analysts verify filings, amendments and calculations. Establish supported questions and distinguish research assistance from unsupported personalized decisions.

First evidence task: Compare an amended filing with an older period containing different units.

E — Engineer the solution

Keep provenance/versioned ingestion and deterministic calculations; connect numerical and citation requirements to separate evidence checks and parsing/index budgets.

The implementation assessment below makes the authority, workload, resource and failure assumptions concrete. Its numbers are calculations from the stated scenario, not a vendor performance guarantee.

Contract

Define supported questions, source authority, effective dates, numerical tolerance and what to do when sources conflict. Separate information extraction/calculation from personalized financial decisions. A correct answer needs the right period, units and source revision, not merely a relevant citation.

Implementation plan

Ingest immutable filings with publication/effective dates and supersession links. Parse sections, tables, headers and notes into linked elements. Hybrid retrieval finds candidate evidence; reranking and table expansion assemble coherent context. A deterministic calculator receives typed values with units and provenance. Generation explains only verified results and flags unresolved conflicts.

Data/state

Store document/version/page/element IDs, table cell coordinates, extracted value/unit, parser version and source references. Cache computations using the evidence-version hash. Keep access controls through retrieval and final output. Prefer latest authoritative amendments where the question requires current facts, while historical questions pin the appropriate as-of revision.

Sizing

Assume 100,000 reports×50 pages=5M pages. At a measured-or-to-be-measured parser rate p pages/s, initial parsing takes 5M/p seconds before source limits. Five million 768D float32 chunk vectors alone use 15.36GB. OCR, table reconstruction, reranking and model token load are separate resource pools; a raw vector estimate cannot size the full system.

Handling conflicting evidence

Normalize “in thousands” before arithmetic. For fiscal-year changes, distinguish comparable-period growth from raw reported-period growth and explain missing comparability. Do not combine an amended numerator with a superseded denominator unless that is an explicitly requested historical comparison.

Evaluation

Use held-out source families with amendment, footnote, multi-table, unsupported and conflict cases. Grade evidence recall, citation support, numerical/temporal correctness and task completion separately. Include oracle-context generation to isolate retrieval from reasoning. Compare against a simpler search-plus-calculator workflow.

L — Launch with evidence

Evaluate held-out source families and amendment/unit traps. Measure analyst verification time as well as factual correctness before expanding coverage.

Critical decision condition: Unsupported material numbers or mixed incompatible reporting periods block accepted output.

Attach raw artifacts, the tested configuration, population, denominator, date and limitations to the release record. Missing or stale evidence holds the dependent decision. A favorable value forecast cannot offset a failed critical condition.

D — Develop ownership and reuse

A research workflow owner maintains source authority rules and exception review; operations owns ingestion and amendment reconciliation.

Candidate reusable asset: Provenance-aware calculation contract.

Commercial decision and next commitment

Accountable decision owner: Research workflow owner with data steward.

The initial commitment is the bounded journey above, with the stated exclusions. Confirm access, customer staffing and acceptance ownership during discovery; price and schedule unresolved dependencies explicitly. A new source, tenant class, language, action or operating envelope requires an impact review rather than silent scope growth.

Approve only the next exposure level supported by the evidence. The next useful investment is: compare an amended filing with an older period containing different units.

Method and references

Apply the FIELD method, economic model and evidence rules. Methodological and technical influences are listed in the source register.