+++
title = "A reproducible model needs reproducible data rights"
description = "Saved code does not guarantee a repeatable model run. Data retention, access and licensing rights can determine what can be reproduced."
date = 2026-09-11
draft = false
[taxonomies]
topics = ["research-environments-reproducibility", "real-time-market-data", "historical-data-corporate-actions", "licensing-vendors-outsourcing-exit"]
kinds = ["second-order"]
[extra]
tier = "public"
schema_type = "Article"
article_class = "second-order"
related_concepts = ["data license", "retention right", "artifact reproducibility", "dataset version"]
sources = ["https://mlflow.org/docs/latest/ml/tracking/", "https://iceberg.apache.org/docs/latest/", "https://professional.bloomberg.com/products/data/data-connectivity/server-api/", "https://www.nyse.com/data-products/catalog/daily-taq", "https://www.federalreserve.gov/frrs/guidance/interagency-guidance-on-third-party-relationships.htm"]
source_details = [{ title = "MLflow Tracking", publisher = "MLflow Project", url = "https://mlflow.org/docs/latest/ml/tracking/", checked = "2026-09-09" }, { title = "Introduction", publisher = "Apache Software Foundation", url = "https://iceberg.apache.org/docs/latest/", checked = "2026-09-09" }, { title = "Server API", publisher = "Bloomberg", url = "https://professional.bloomberg.com/products/data/data-connectivity/server-api/", checked = "2026-09-09" }, { title = "Daily TAQ", publisher = "NYSE / ICE", url = "https://www.nyse.com/data-products/catalog/daily-taq", checked = "2026-09-09" }, { title = "Interagency Guidance on Third-Party Relationships", publisher = "Federal Reserve, FDIC and OCC", url = "https://www.federalreserve.gov/frrs/guidance/interagency-guidance-on-third-party-relationships.htm", checked = "2026-09-09" }]
faq = [{ question = "Does keeping the code guarantee a future rerun?", answer = "No. The specified data and execution dependencies must remain available and permitted." }, { question = "Does a data hash recreate a missing dataset?", answer = "No. It can support an integrity comparison, but it does not contain the dataset." }, { question = "Does every expiring data subscription prevent reproducibility?", answer = "No. Retained rights, a permitted archive or continued access can preserve the required inputs." }]
evidence_as_of = "2026-09-09"
related_articles = ["point-in-time-backtest-data", "notebook-to-model-service", "market-data-api-access"]
category_slug = "data-access"
category_name = "Data & access"
+++

Operational reproducibility is the ability to rerun a specified calculation using available, permitted inputs and a controlled execution environment. Retaining code and model artifacts does not by itself preserve future access to the required data.

## Computational identity and input availability

A reproducible calculation needs identifiable code, configuration, input versions and relevant execution dependencies. Experiment tracking can connect those objects through a run record. The record still needs references that remain resolvable when the replay occurs.

A dataset identifier tells a system which input was used. A stored hash can help compare candidate bytes with a recorded value. Neither supplies an unavailable dataset. Likewise, a snapshot reference identifies a table state only while the required contents and metadata remain accessible.

Data permissions add a separate condition. A technically retained copy and permission to use that copy are different facts. The actual agreement determines which retention and use arrangements exist.

## A model outliving its input agreement

Take a model retained for five years. Assume its input agreement ends after one year and permits neither retained copies nor further access after termination. Also assume the specified inputs cannot be obtained through another permitted route.

During the remaining four years, retaining the model code and environment does not provide an authorized rerun of that original calculation. The missing condition is access to the permitted input version. The example’s restriction is a stipulated contract term, not a claim about every market-data license or any named supplier’s retention policy.

Replacing the unavailable inputs with another dataset can produce a new calculation. It does not reproduce the old calculation unless the replacement satisfies the original specification and agreement criterion. Similar field names do not establish identical contents, timing or adjustment history.

## Rights and access are separate dependencies

A permitted archive can remain technically inaccessible after keys, credentials or storage are lost. Conversely, a technically reachable dataset can be outside the permitted use scope. Future replay needs both usable access and the relevant permission.

Market-data interfaces already illustrate distinct access boundaries. Bloomberg Server API delivery, in documentation checked in September 2026, requires an active Bloomberg Professional session for the user. That is an access condition for the documented service. It does not establish a universal restriction on retained historical data.

A replay inventory therefore connects the calculation’s required horizon to dataset versions, storage, keys, access paths and the permissions actually retained. Code openness does not resolve any missing condition in a separately licensed input.

## The lifecycle consequence

If a calculation must remain reproducible for a defined period, its indispensable input dependencies must remain available and permitted for that period under some supported arrangement. Retaining only the executable artifact leaves that requirement incomplete.

The mechanism changes how reproducibility is evaluated over time. A successful rerun today establishes current reconstruction under today’s access conditions. It does not establish that a replay after a contract or service change will have the same inputs available.

## Arrangements that preserve future replay

Perpetual retention rights, a permitted archive or equivalent continued access can preserve replay after a subscription ends. The result depends on the actual rights and technical arrangements, not on whether the original data arrived through a commercial API.

The conditional conclusion is limited: when required historical inputs become unavailable or impermissible to use, fixed code and environment alone cannot provide an authorized replay. A data-rights dependency is part of operational reproducibility without becoming a universal claim about data contracts.

## Questions about data license

### Does keeping the code guarantee a future rerun?

No. The specified data and execution dependencies must remain available and permitted.

### Does a data hash recreate a missing dataset?

No. It can support an integrity comparison, but it does not contain the dataset.

### Does every expiring data subscription prevent reproducibility?

No. Retained rights, a permitted archive or continued access can preserve the required inputs.