7.9 KiB
7.9 KiB
VYNDR — Clean-Room Design Log
This log documents every external codebase studied during VYNDR's development. It proves independent creation — no code was copied from any external source. When we examined a project under a copyleft license (GPL/AGPL/LGPL) we restricted the study to concept and behavior, not implementation.
| Date | Source | License | What Was Studied | Insight (Our Own Words) | Code Copied? |
|---|---|---|---|---|---|
| 2026-05-24 | nba_api (github.com/swar/nba_api) | MIT | API wrapper patterns for stats.nba.com endpoints | NBA.com exposes hidden JSON endpoints with rate limiting; our adapter calls these directly with our own session + retry layer. | No |
| 2026-05-24 | pybaseball (github.com/jldbc/pybaseball) | MIT | Statcast data access patterns | Baseball Savant provides pitch-level data via public CSV endpoints; we shape our own query layer on top of those endpoints. | No |
| 2026-05-28 | OpenBB (github.com/OpenBB-finance/OpenBB) | AGPL-3.0 ⚠️ | UI layout patterns for dense financial data display ONLY — no code studied | Terminal-style UI with panel-based layout suits intelligence products; applied as a design pattern only. | No — concept observation, no code read |
| 2026-06-05 | River (github.com/online-ml/river) | BSD-3-Clause | Incremental ML API design, ADWIN drift detection algorithm | learn_one/predict_one interface enables single-sample model updates; ADWIN splits an error window for drift detection. | No |
| 2026-06-05 | kyleskom/NBA-ML-Sports-Betting | Unlicensed ⚠️ | XGBoost + NN dual-model approach, Kelly Criterion sizing | Two independent model architectures voting together increase robustness; sigmoid calibration converts raw scores to probabilities. | No |
| 2026-06-05 | Lisandro79/BeatTheBookie | GPL-3.0 ⚠️ | Market consensus deviation concept ONLY | When one book's line diverges from consensus, it signals either sharp action or a trap. | No — concept only |
| 2026-06-05 | charlesmalafosse/sports-betting-customloss | MIT | Custom loss function concept | Training loss should weight correct predictions by implied probability — a correct call on +150 is more valuable than one on -200. | No |
| 2026-06-05 | AagamanVarma/DriftGuard | Unknown | Champion-challenger deployment pattern | Challenger model earns promotion by outperforming champion on holdout set; a model pointer file enables atomic swap. | No |
| 2026-06-06 | Serwist (serwist.pages.dev) | MIT | Service worker integration with Next.js | @serwist/next wraps next.config and injects a service worker entry; defaultCache provides reasonable runtime caching presets out of the box. |
No — documentation read; code is our own SW wiring |
| 2026-06-06 | web-push (github.com/web-push-libs/web-push) | MIT | VAPID-based push delivery semantics | A 410 Gone response on a subscription is the signal to delete it permanently; other statuses are transient. | No |
| 2026-06-06 | SharpAPI (sharpapi.com) | Commercial | API surface only (docs read, no source available) | Player-prop endpoints expose over/under odds per book; consensus is computed client-side by aggregating across books. | No — commercial API, no code |
| 2026-06-06 | OddsPapi (oddspapi.io) | Commercial | API surface only (docs read, no source available) | Pinnacle closing-line endpoint requires a bookmaker filter; the docs make no guarantee about latency between game start and line availability. | No — commercial API, no code |
| 2026-06-06 | ParlayAPI (parlayapi.io) | Commercial | API surface only (docs read, no source available) | Free tier exposes 3.7M historical prop closing records via a credit-based bucket — careful pacing is mandatory because credits don't refresh mid-month. | No — commercial API, no code |
| 2026-06-06 | PropOdds (prop-odds.com) | Commercial | API surface only (docs read, no source available) | Specializes in player props specifically; useful as a second consensus point so line-divergence trap signal isn't a two-book artifact. | No — commercial API, no code |
| 2026-06-06 | College Football Data (collegefootballdata.com) | Free / personal API key | API surface only (docs read, no source available) | Talent composites + PPA fill the gap nba_api / ESPN don't cover for college. Bearer-token auth keeps the key off URL logs. | No — commercial-style API, no code |
| 2026-06-06 | Sports-Reference (basketball-reference.com) | Site TOS (HTML scraping with credit) | Public HTML tables only — referee + coach index pages | Data-stat attributes on table cells are the cleanest extraction key. We rate-limit at 1 req per 5s and identify with a User-Agent so the site can contact us if anything's off. | No — read public HTML, parsed table data only |
| 2026-06-06 | OpenRouter (openrouter.ai) | Commercial | API surface only (docs read, no source available) | OpenAI-compatible /chat/completions endpoint with a model-selector header. Bearer auth keeps the key off URLs. Free-tier per-model rate limits are not documented per-account; conservative 20/min budget is safer than relying on 429 retry. | No — commercial API, no code |
Template for future entries
| YYYY-MM-DD | <source URL> | <license> | <what we studied> | <our own-words insight> | No |
Rules
- Always log GPL/AGPL/LGPL projects with the ⚠️ marker and limit study to concept observation. Do not read source files line-by-line.
- Unlicensed code is studied as ideas only. Without a license the work is default copyrighted; we do not duplicate structure.
- Document the insight in our own words. If an entry reads like a paraphrase of the source, rewrite it.