HiveBase

Guides

Model Intelligence

How HiveBase publishes research-based model defaults, attributed evidence, and observational HiveBase-owned outcomes — without a leaderboard or rehosted scores.

Last updated 2026-09-06

HiveBase routes by work class, not a single leaderboard. Coding is ranked as (harness, model, role). The public record is hivebase.ai/models and the same sanitized snapshot at GET /api/models/ledger.

This page is durable selection guidance. Architecture stays in the engineering docs.

What the public page shows

  • Current research-based default for each work class — useful before HiveBase outcome data accumulates.
  • Applicable tasks and meaningful limitations.
  • Research basis plus freshness (when the snapshot was generated and which sources last refreshed).
  • Links to attributed sources. Artificial Analysis is named and linked; numbers are never rehosted.
  • Outcome statistics in a separate block — HiveBase-owned dogfood only, labeled observational, with population, dates, denominator, and uncertainty.

The preview snapshot used when the live feed is missing does not invent preferred model names.

How a default is allowed to exist

  1. Bind the work class (and harness, for code.*). Do not pick a global “best model.”
  2. Research can establish a default from attributed public sources and the HiveBase catalog. That default stays labeled research-based.
  3. First-party outcomes never override a measured comparison just because n ≥ 20. n=20 is a reporting threshold, not a causal gate.
  4. Community recency and catalog chatter do not rank quality.
  5. Saturated suites (SWE-bench Verified, HumanEval, retired open LLM leaderboards) are weight-zero.

Outcomes versus research

Existing eval consent does not authorize publishing customer traces. HiveBase initially publishes only HiveBase-owned / dogfood outcomes, clearly labeled. Missing instrumentation stays unknown. Seven-day survival stays pending until the window closes.

Changelog and notes

Routing changes and rollbacks publish automatically as a factual changelog. Narrative Model Notes are drafted when material findings warrant them, reviewed weekly, and published only through the existing human editorial workflow.

API

GET /api/models/ledger?work_class=&harness=&since= returns the same sanitized numbers as /models. It never returns private claims or raw customer traces. Immutable vintages (v-YYYY-MM-DD) stay off the marketing page; slug current is the live pointer.