Data Sources

Last updated 25 August 2026 · Soulvaya Labs, Inc. · bellcowgm.com

This is the "how Bellcow thinks" page. It explains what data goes in, what the models do with it, where the AI sits in the chain, and — source by source — what licence each one is under and whether Bellcow is actually allowed to use it. The register below is generated from the same file the code enforces, so it cannot quietly drift from reality.

1. How a recommendation gets made

1. Ingest. Play-by-play, snap counts, participation, rosters, depth charts, injury reports, schedules and game lines are pulled from nflverse. Weather comes from the National Weather Service. Trade-market values come from Stats Guy Fantasy.
2. Player state. Those feeds are reconciled into one current picture of every player — role, health, depth-chart position, usage trend. Where two sources disagree about an injury status, the conflict is surfaced rather than silently resolved.
3. Expected points. An opportunity model converts where and how often a player is used into expected stat components, calibrated on historical play-by-play. Actual minus expected gives a regression signal — a hint, not a verdict.
4. Projection. A component-based weekly model predicts stat components from recent role, efficiency, team environment, opponent defence, the market-implied game total, home/away and availability. Components stay separate so that your league's scoring can be applied afterwards — one football projection serves every format.
5. Rest of season. A neutral-context per-game model adjusted for role stability, remaining schedule and health, published as median, floor and ceiling rather than a single number.
6. Your league. Your scoring, lineup slots, waiver system and playoff structure are applied, and every team in the league is evaluated for strengths, needs and surpluses.
7. Simulation. Your remaining schedule is simulated many times over with a seeded generator, producing playoff and championship probabilities — and an honest margin of error alongside them.
8. The plan. Candidate lineup changes, waiver claims and trades are scored by how much they move those probabilities, ranked, and shown with their reasoning. When a move's effect is smaller than the simulation's own error, Bellcow says it is too close to call.
9. Confidence. Every player and recommendation carries a HIGH/MEDIUM/LOW rating and a 0–100 score based on how fresh the inputs are, how many sources agree and whether they conflict. Low confidence is displayed, never hidden.

2. Where the AI actually sits

The large language model does not compute your projections, your values or your odds. Those come from the models above, which are deterministic code calibrated on historical data. The AI's job is conversation: it receives a compact summary of the league state the engine has already computed and answers your questions about it in plain English. If the AI is unavailable or unconfigured, the assistant falls back to a local rule-based answer and the app tells you which one replied. Read the AI Disclaimer for what that means for accuracy.

3. The source register

Every external source Bellcow knows about lives in one registry file, and production code refuses to consume a source that is not marked approved. The register carries two separate facts about each source, and they are worth reading as two columns rather than one:

• What it covers — the kind of data it holds.
• Licence status — our current classification of whether we are permitted to use it commercially. Approved means yes. Prototype only means it is usable while we are proving the product but needs permission or a terms review before it is commercially clean. Licence required and Blocked mean it is not ingested. Unclassified means we have not finished assessing it and it is treated as research only.

Current register: 12 approved, 11 prototype-only, 19 licence-required, 2 blocked, 6 unclassified. Last reviewed 24 August 2026 by the product owner. These classifications are the owner's working assessment, not the opinion of counsel — the same review noted at the top of this page covers them.

4. The two open items, stated plainly

Two related notes, for completeness: the ESPN connector reads undocumented internal endpoints, which is flagged in the register as needing legal review — manual import is the permitted fallback and always works. And Pro Football Reference is marked BLOCKED: no bulk scraping, human QA only. Those flags exist because a source register that only contains good news is decoration.

5a. The live wire

Projections and rankings come from the intel bundle, which is rebuilt on a schedule. Injury alerts cannot wait for that, so they run off a separate path called the live wire. It merges several independent status feeds rather than depending on any one of them.

What happens when a feed fails. Nothing changes. Player status is built from graded claims: a source reports something, the report is corroborated against other sources and scored by who made it, and only a sufficiently supported claim moves a player's state. A source that goes dark simply stops making claims — and no claims means no change, so an outage cannot make players look like they have recovered. What it costs is freshness, and that is shown rather than hidden: every player's state carries the time it was last verified, and confidence falls continuously from there. The current state of every feed is published on the status page.

5b. Where this honestly stands today

5. Sources that feed the models

These are read automatically and their data reaches your recommendations.

6. Sources that read your league

These bring your own league into Bellcow. All read-only — none of them can write to your team.

7. Research, benchmarks and reference

Used by humans to sanity-check the models and to see whether Bellcow disagrees with the consensus for a good reason. Not ingested into production intelligence. The industry sites in this list are benchmarks we compare against, not feeds we consume — that distinction is enforced in code, not just stated here.

8. Registered but not used

Commercial feeds we have assessed and are not using. They are listed so the decision is on the record. The most likely first purchase is a real-time injury and inactive feed, and only when the data proves the need.

9. Attribution

• NFL statistics, play-by-play, rosters, depth charts, injuries, schedules and game lines: nflverse, used under CC BY 4.0. nflverse.nflverse.com
• Player trade-market values: Stats Guy Fantasy, free and open with a visible credit link. statsguyfantasy.com
• Weather: National Weather Service (weather.gov), US Government public domain.
• Projections, expected points, values, breakout scores, simulations and championship odds: computed by Bellcow.

Team, league and player names are used descriptively. Bellcow is not affiliated with the NFL, any club, the NFLPA, or any fantasy platform.

10. What this pipeline cannot do

Free and open data has a known ceiling, and it is worth naming: our injury picture is as current as public reporting, which on a Sunday morning can lag a licensed real-time feed by minutes that matter. Practice participation, coach-speak and beat-reporter nuance are only partly captured. Scheduled refresh jobs run on a cadence limited by upstream feeds and by our host's scheduling constraints, so data can be late or a run can be skipped. Bellcow is built to degrade honestly in those moments — lower the confidence score, show the conflict, say it does not know — rather than to fill the gap with something invented.

11. Keeping this page honest

The tables above are generated from the compliance register the code enforces. When a status changes — Sleeper grants permission, Yahoo approves the application, a feed gets licensed — the register changes, this page changes with it, and the "Last updated" date moves. If you spot a source we have classified wrongly, or you represent one and want to talk about it, write to support@bellcowgm.com and we will act on it.

Contact

Soulvaya Labs, Inc. — Bellcow GM
131 Continental Drive, Suite 305, Newark, Delaware 19713, USA
support@bellcowgm.com
Contact Support · Help Center · AI Disclaimer · Fantasy Disclaimer