# The TAB Index

> The TAB Index is an independent, deterministic daily measurement of AI model output
> quality, published by TAB Platform LLC. Scoring is by fixed rules against known ground
> truth — no LLM-as-judge, so the judge cost is zero. Every run is a dated, version-stamped
> observation, and records are never overwritten. It is a data publisher and auditor: it
> makes no model recommendations, and there is no composite score.

## What it measures

- Six per-property categories: Security (12 cases), Trust (14), Agentic (15),
  Integrity & Provenance (15), Resilience (15), Instruction Authority (15).
  Instruction Authority joined the public board August 6, 2026; Claude Opus 5 joined
  August 8, 2026. These are historical milestones, not current roster counts.
- Weekday rotation (began Aug 7, 2026): Integrity & Provenance and Instruction Authority are
  measured every night; one other category is added per weekday — Security Monday, Trust
  Tuesday, Agentic Wednesday, Resilience Thursday — and weekends measure Integrity &
  Provenance and Instruction Authority only. Each model is measured every night on that
  night's scheduled categories, not on all six every night.
- Measurement is per-property only — no composite — because capability rank and trust rank
  differ and a single number hides that.
- When a category is scheduled it runs at least three times that night under identical
  conditions; the public scores are from the most recent run, not an average. A category not scheduled on a
  given night keeps its most recent measurement, shown with the date it was measured.

## Publication and qualification

- Configured models are the admitted, non-retired routes scheduled for measurement.
  Configuration alone does not mean a completed observation or score exists.
- The live board and `GET /api/index/latest` show published observations. Read the
  returned model rows and each category's `measured_at`, `publication_status`, `scored`,
  `unscored`, and `below_floor`; a row count is not a count of fully scored models.
- `measured` means a scored observation is available. `provisional` means an observed
  score has incomplete comparability admission; it is published with that limitation.
  Neither status certifies a model or establishes mature change-detection history.
- `unavailable` means there is no usable score for that cell, including wholly unscored
  attempts or categories with no observation yet. It is not zero performance.
  Partial coverage is disclosed separately; below-floor cells show coverage rather than
  a comparable headline score. Unscheduled categories retain their last observation,
  if any, with its own measurement date.
- Since October 26, 2026, valid completed results publish without a staging wait,
  manual promotion, complete-category rotation, or seven-day embargo. Early history
  develops alongside publication. Newly configured models may have no row until their
  first completed observations arrive. Normal API caching can briefly delay refresh.
- Authoritative diagnosis is separate: explicit diagnosis eligibility AND sufficient
  evidence for the particular series are required. Public visibility, a measured score,
  or a stored tripwire flag does not by itself qualify a model for authoritative diagnosis.
- Model counts change. Use the live board/API for current published rows, not a fixed
  count in this document. Dated reports retain their historical membership and results.

## How it scores

- Deterministic, rule-based scoring against case-supplied ground truth. No model judges
  another model, so the judge cost is $0.00.
- A model that declines a case is recorded as unscored, never as a false zero.
- A calibration suite (28 synthetic cases, 12 checks) periodically runs synthetic agents with
  known behaviour through the real scorers to detect scorer drift.

## Records and cost

Published suite-cost figures count scored cases only; they are not total cycle charges.

- One measurement cycle per night at 06:00 UTC (02:00 America/New_York during daylight saving time; 01:00 during standard time). Every run is dated and
  version-stamped; records are never overwritten, so change over time stays visible.
- Cost per suite run and tokens per suite run are published as measured figures, from observed
  token counts and published list rates for the route used.
- The public board and GET /api/index/latest are free during the current promotional period.

## Live API

Base: https://tabverified.ai

- `GET /api/index/latest` — latest published observation per model/category, not a
  single-cycle report. `cycle_id` is null by design; use each cell's measurement date.
  No history-selection parameters. Open during the configured promotional window;
  an API key is required outside it. Responses may be cached for up to five minutes.
- `GET /api/index/history/{model_name}` — per-cycle history for models in the current
  public-visibility list; API key required. Optional `?category=<slug>`. The lifecycle
  label `staging` does not exclude a publicly visible model. Unknown or currently
  non-public models return 404 after authentication; retired models' dated reports
  remain preserved even when this current-membership endpoint excludes them.
- `GET /api/index/tripwire` — stored flags for publicly visible models in the endpoint's
  selected publishable cycle; API key required. Flags are reported verbatim and are
  not authoritative diagnoses or a promise of mature history for every visible model.
- Supported key headers: `X-API-Key`, `X-TAB-API-Key`, or `Authorization: Bearer <key>`.
  Never put API keys in URLs or public documents.

## Start here

- [/the-tab-test](/the-tab-test) — live numerical results board.
- [/daily-results](/daily-results) — latest completed cycle in plain language.
- [/report/](/report/) — dated historical reports.
- [/methodology](/methodology) — how the Index scores.
- [/about](/about) — what the Index is and who runs it.
- [/five-findings.md](/five-findings.md) — a historical worked example; not the current roster or current results.
- [/llms.txt](/llms.txt) — root index for AI crawlers and agents.

## About

- The TAB Index is a data publisher and auditor. It does not build, sell, host, or recommend
  models, and has no financial relationship with any model provider.
- Legal entity: TAB Platform LLC.
- Contact: rod@tabverified.ai
- Website: https://tabverified.ai

