# Reading a Report

A report opens when a test finishes, and from any row under **Checks**. It starts with the verdict and its confidence, then shows the evidence behind it.

## Verdicts

| Verdict | Meaning | What to do |
|---|---|---|
| **Genuine** | The answers match the claimed model, and the competence and refusal checks, if run, found nothing wrong. | Nothing. A competence check confirms the quality as well. |
| **Degraded** | The claimed model, but measurably weaker than the official deployment: fewer items right, or much longer reasoning. | Confirm at the Thorough tier. Typical causes are quantization below the official precision or a lowered reasoning budget. |
| **Abliterated** | The endpoint answers nearly all the benign requests the official model declines, with no sign of an injected prompt. | The weights were modified. See [Refusal Check](https://llm-id.sh/docs/refusals). |
| **Substituted** | The answers match a different model in the baseline, which the report names. | Confirm at the Thorough tier before raising it with the provider. |
| **Unknown Model** | No model in the baseline fits: the answers fit the claimed model worse than its genuine runs do. | Confirm at the Thorough tier. It can be a model outside the baseline, a modified copy, or an added system prompt. |
| **Mixed Routing** | Requests go to several models or backends. The report names them and their shares where it can. | See [Mixed Routing](https://llm-id.sh/docs/mixed-routing). |
| **Inconclusive** | Not enough evidence either way. | Run again at the Thorough tier, or later if rate limits cut the test short. |

## Confidence

The rating next to the verdict bounds the chance of seeing evidence this strong if the verdict were wrong:

| Rating | Confidence |
|---|---|
| **Very High** | 99.9% or more |
| **High** | 99% or more |
| **Moderate** | 95% or more |
| **Low** | below 95% |

After the rating, the report names what the confidence rests on: identity, competence, refusals, the blend test or the split test. The bound assumes the baseline describes each model correctly. It doesn't cover models outside the baseline or changes since the baseline was collected, and the report notes either where it matters.

## The Report's Sections

**The header** gives the verdict in a sentence, then the endpoint, the model name used, the claimed model and what the test sent. Notes below it explain anything unusual: a sibling model that answers almost alike, rate limits, an injected prompt.

**Identity** shows how well the answers fit each model in the baseline:

- **Fit Among Genuine Runs** places this test's fit to the claimed model between the floor that genuine tests stay above and their median. A marker below the floor means the answers don't look like the claimed model.
- **Fit per Answer** lists the closest models. The longest bar to the right is the best fit; bars to the left fit worse than a model that knows nothing.
- The line above them gives the likelihood ratio against the runner-up: 10^13 means the answers are ten trillion times likelier under the best model.

**Competence** appears when the competence check ran: items right against what the official model would get, the accuracy statistic against its line, reasoning length against official, then each task family and the cells with the largest change in reasoning. See [Competence Check](https://llm-id.sh/docs/competence).

**Refusal Probes** appears when the refusal check ran: how many requests the endpoint complied with, what the official model would comply with, the line for "more than official" and the mark for a complete removal. Each probe shows one square per answer, colored when the endpoint complied, and its first answer. See [Refusal Check](https://llm-id.sh/docs/refusals).

**Weights or Prompt** shows how many prompt tokens the endpoint adds to each request compared with the official route. Fewer than 256 extra tokens is a different chat template; 256 or more is a system prompt the endpoint injects.

**Mixed routing** reports add the requests grouped by prompt-token count, the split test, the blend test on the answers alone, and earlier tests of the same endpoint.

**Run** lists latency, time to first token, tokens per second, errors, the provider that served the requests where the router names it, hidden reasoning, and reported cost.

## Buttons

- **Rerun** opens the test form filled in with this test's settings.
- **Show Answers** lists every identity probe with this endpoint's answers and what the claimed and closest models usually say. A dashed answer was never seen in the baseline.
- **Export JSON** saves the full report.
- **Certificate**, below the buttons on a Genuine or Abliterated test, issues a public, signed certificate with badges for your site. See [Certificates and Badges](https://llm-id.sh/docs/certificates).

## The Checks List

**Checks** lists every test your account has run, newest first, with counts per verdict above it. The filters show one verdict at a time. A row shows the endpoint, the model name, the claimed model, the competence or refusal result and the confidence; **Open** opens the report. A test marked **Interrupted** stopped because the service restarted while it ran; run it again.

---

From the llm-id.sh guides: https://llm-id.sh/docs/reading-a-report