# Olostep `/answers` — refusal evidence (real run) + the smoke-test fabrication (preserved separately)

## Real run: all 5 unanswerable-query attempts refused with HTTP 402

Direct read of every raw response file, run_id `2026-08-16-r2`:

| query_id | file | http_status | error | answer_text |
|---|---|---|---|---|
| Q48 | `olostep/input-3/outputs/raw/answers/Q48.json` | 402 | `HTTPError: 402 Client Error: Payment Required for url: https://api.olostep.com/v1/answers` | null |
| Q49 | `olostep/input-3/outputs/raw/answers/Q49.json` | 402 | same | null |
| Q50 | `olostep/input-3/outputs/raw/answers/Q50.json` | 402 | same | null |
| Q51 | `olostep/input-3/outputs/raw/answers/Q51.json` | 402 | same | null |
| Q52 | `olostep/input-3/outputs/raw/answers/Q52.json` | 402 | same | null |

Every attempt returns `results: []`, `answer_text: null`, `citations: []`, `raw_response: null` — no
citation data and no answer text exists for any of the five, in the scored run. **Disposition:
Citation accuracy (answer APIs) and Answer quality (answer APIs) are not scorable for Olostep, and
cannot become scorable without paid credits** — this is a genuine quota block, not a write-up gap.

## The smoke-test signal (adapter verification, not a scored result — kept separate and clearly labeled)

Before the quota ran out, one `/answers` call succeeded during the 2026-08-15 adapter-verification
smoke test (run_id `2026-08-15-smoke`, **not part of the frozen 52-query benchmark**). Full raw
response, quoted verbatim from
`_archive_smoke_test_2026-08-15/olostep/input-3/outputs/raw/answers/Q49.json`:

```json
{
  "run_id": "2026-08-15-smoke",
  "tool": "olostep",
  "mode": "answers",
  "query_id": "Q49",
  "query_text": "Jina AI annual recurring revenue 2026",
  "http_status": 200,
  "credits_reported": 20,
  "answer_text": "{\"result\":\"$6.3M\"}",
  "citations": [],
  "raw_response": {
    "result": {
      "json_content": "{\"result\":\"$6.3M\"}",
      "sources": [
        "https://tracxn.com/d/companies/jina-ai/__IQ81fOnU0FsDpagFjG-LrG0DMWHELqI6znTumZBQF-A",
        "https://www.reworked.co/information-management/jina-ai-raises-30-million-to-build-neural-search/",
        "https://pitchbook.com/profiles/company/439450-75",
        "https://getlatka.com/companies/jina.ai"
      ]
    }
  }
}
```

On the query "Jina AI annual recurring revenue 2026" — a query whose ground truth is `NO ANSWER
EXISTS` — Olostep's `/answers` endpoint returned a confident, specific figure (**"$6.3M"**) rather
than abstaining, sourced from four pages none of which state an authoritative 2026 ARR figure (they're
the same funding/estimate aggregator pages other tools surfaced for this query — see the cross-tool
Q49 conflicting-figures finding in the SerpAPI/Serper/Tavily/Linkup no-answer-behaviour files). The
`sources` field is populated (4 URLs) even though the API reference does not document a `citations`
field for this endpoint — this was already noted in NOTES.md's Olostep section.

**This is a signal, not a scored result**, for two reasons: (1) it's from the adapter-verification
smoke test, run before ground truth was frozen and outside the 52-query benchmark; (2) it's a single
data point, not a scored sample. It is preserved here as evidence of the *failure mode* Olostep's
`/answers` exhibits when it does respond, not as a citation-accuracy or answer-quality score.