No validation infrastructure
Chronological splitting, purging, embargoing and a trial ledger are real engineering work — a small team often has neither the time nor the mandate to build them before the research itself can start.
For fintech and prop teams
One methodology, shared by the whole desk.
A small team can standardise on validation infrastructure without building it first: projects isolate each question, workflows are versioned, and every run leaves a record — not because someone remembered to write it down, but because the workspace does.
01A quarter with Quant Compose
From onboarding a team to a quarterly evidence pack a downstream system can consume, across a single quarter.
Month 1
Standardise on projects and templates
Every research question gets its own isolated project — its own universe, dataset, models, workflows and audit trail — built from one of five workflow templates or a blank canvas, so a new hire starts from the team's own pattern rather than a blank page.
Workflow Composer Named research templates to start a pipeline from. Month 2
Assign roles and invite the team
Owner, admin, quant researcher, analyst and viewer cover a desk's natural division of access. Production registration requires an invitation from an owner; invite links are stored hashed and work once, and a workspace always keeps at least one owner.
Settings Workspace settings: profile, research defaults, appearance and data connections. Month 3
Read the audit log
An immutable, chronological log records every action, run and artifact the workspace produced, traceable back to its project and route — the reference when a number from three months ago needs explaining.
Audit & Governance Audit & Governance: the project experiment lineage, one row per recorded action with its actor, artifact and status, filterable and exportable as CSV. Month 4
Export an evidence pack for downstream use
The same evidence model that renders the PDF, HTML and DOCX pack is also available as a JSON payload, so a risk system or dashboard downstream reads the identical figures a human reads in the report.
Reports Report structure: typed evidence blocks arranged into sections, each block re-orderable and traceable to the module that computed it. Month 5
API keys, with an enforcement gate
Provider credentials resolve from environment variables or a server-side store; the API can report that a credential is configured and its source, but the credential store never returns the value itself, on request or by accident.
Settings Macro-economic data and enterprise adapters in Settings, each with its status and the setup it needs.
02Three problems it removes
Named specifically, not as a category.
Methodology varies per person
Without a shared workspace, one person's split logic differs from another's, and two results that look comparable were never actually produced under the same conditions.
Nothing auditable
When a number from last quarter needs defending, a team without a shared, immutable log is reconstructing the answer from memory and scattered notebooks.
03Auditability and team standardisation
Method that holds up when someone checks.
What a team gets by standardising on one workspace is not convenience alone — it is a record that survives staff turnover and a compliance question.
- Append-only trial ledgers and write-once verdicts
- There is no update or delete for a recorded trial, and a validation verdict already attached to an experiment cannot be replaced — a different verdict requires a new experiment, not an edit to the old one.
- An environment and content-hash fingerprint per record
- Each experiment record carries a SHA-256 hash of its inputs — strategy version, dataset id and hash, feature set, universe, period, parameters, split design and seed — so a supposedly identical re-run that produces a different answer is flagged, not missed.
- A per-step run record for every workflow
- Each step logs which steps ran, which were skipped and why, which failed and where, and what parameters were actually used after defaults were resolved.
- Tenant isolation and an egress allowlist
- Every store reads and writes only its own tenant's documents, and outbound data calls are restricted to an approved host list with SSRF defences on every resolved address.
- The credential store never returns a value
- The API reports that a credential is configured and its source, but never returns the secret itself, and every secret is redacted from logs before being written.
04Objections, taken seriously
The questions this audience actually asks.
Is our data safe?
The workspace is local-first: persistence is JSON files by default or PostgreSQL when configured, on a server your team controls or the hosted deployment. Outbound calls are restricted to an allowlisted set of hosts with SSRF defences on every resolved address, and secrets are redacted before they reach a log.
Can we self-host?
Today, yes, as Python on your own machine or via the bundled Docker image. A supported on-premise or VPC deployment for a team, with SSO and signed audit export, is stated as design intent for the Institution plan rather than a shipped product today.
Is there an API?
Yes — the same specification objects and evidence model the interface uses are reachable through the API, and API key limits scale with plan.
Is billing live?
No. The default payment gateway records instruments locally and settles against a local ledger; a live Stripe integration is scaffolded but dormant. Published plans and prices are stated as planned, not charged today.
05In beta, for you
What a small team gets access to during the private beta.
- Multi-seat access with the five defined roles and hashed, single-use invitations
- The full audit log, trial ledger and evidence export, including the JSON payload for downstream systems
- Account-level data export as a downloadable JSON document, itself audited
- On-premise and VPC deployment, SSO/SAML and SCIM provisioning remain design intent for the Institution plan, not shipped today
Other audiences, and further reading
Related capabilities
Standardise the desk on one auditable record.
Teams evaluating a shared deployment should get in touch directly; the security model is documented in full.