# Headroom C2 "Same answers": rerun with large hosted models (2026-10-20 12:05-12:13 UTC+8)
Plan and pass mark committed first (20-c2-hosted-rerun-plan.md, commit 137fa02). Method unchanged from 09-c2b-local-llm.py:
same Headroom 0.41.0 venv, same compress() call, same 2 slices, same 10 questions, same prompt, same scoring. Compressed slices
reproduced exactly (11,477 -> 5,161 and 12,788 -> 7,953 Headroom tokens; 0 and 15 <<ccr:...>> markers).

## Models (OpenRouter free tier, $0 account, no credits bought; temperature 0, seed 1)
- nvidia/nemotron-3-ultra-550b-a55b:free (NVIDIA Nemotron 3 Ultra, 550B MoE / 55B active), served by Nvidia.
- dots-studio/dots-3-note-preview:free (Dots Studio Dots3-Note Preview, 280B MoE / 16B active), served by AtlasCloud.
- Planned second model thinkingmachines/inkling:free returned HTTP 403 ("only available on agentic harnesses") on the first request; not used, no answers recorded. No Llama-70B free model was listed today.

## Verdict method: slices, one question per request (20 requests per model)
| Model | Logs slice original | Logs slice compressed | Search slice original | Search slice compressed | Total original -> compressed |
|---|---|---|---|---|---|
| Nemotron 3 Ultra | 5/5 | 5/5 | 5/5 | 5/5 | 10/10 -> 10/10 (10/10 identical answers) |
| Dots3-Note | 5/5 | 5/5 | 5/5 | 5/5 | 10/10 -> 10/10 (10/10 identical answers) |
| (earlier) Qwen2.5-1.5B, local CPU | 4/5 | 2/5 | 5/5 | 2/5 | 9/10 -> 4/10 |
| (earlier) Qwen2.5-3B, local CPU | 5/5 | 4/5 | 5/5 | 1/5 | 10/10 -> 5/10 |

## Extra, NOT the verdict method: full payloads (now fit: 137K / 83K and 41K / 29K prompt tokens)
5 C2a questions per payload, asked in ONE request per input (budget), same scoring.
| Model | Logs (2,000 rows) original | compressed | Search (300 rows, 44 ccr markers) original | compressed |
|---|---|---|---|---|
| Nemotron 3 Ultra | 5/5 | 5/5 | 4/5 | 4/5 |
| Dots3-Note | 5/5 | 5/5 | 5/5 | 5/5 |
Nemotron's one miss is the same aggregate on both inputs ("how many results have wordcount > 5000": truth 27, answered 26 original / 23 compressed).

## Do misses trace to <<ccr:...>> markers? (27/28-c2-ccr-marker-trace)
No. Markers only ever replace the `snippet` field (15 rows in the search slice, 44 in the full search payload; 0 in logs). No question in the
slices needs a field behind a marker, and the large models had no compressed-input misses at all. The earlier Qwen misses also do not trace
to markers: their wrong values come from other rows of the compact table (e.g. pageid 23862 = row 0, 'Michael Palin' = row 42, size 17906 = row 40)
or from another column ('ledger' is a word in the FATAL msg, not the service). So the small-model failure is about reading Headroom's
header + CSV table format, not about hidden text. The full-payload snippet question (row 3) was not behind a marker either; we did not
test a question whose answer is only behind a marker (plain compress() gives the model no headroom_retrieve tool).

## Grading (pass mark from 00-test-plan.md, written down again before running)
V2 if compressed accuracy >= original - 1 and >= 95% retained or retrievable; V1 if retention 70-95% or drop 2-4; F otherwise.
Retention 20/20 (100%, C2a). Large hosted models: drop 0 and 0 -> V2. C2 = V2 for large models; the small-local-model drop (5/10) stays
on the page as a caveat. Overall = worst of C1b (F), C2 (V2), C3 (V1), C4 (V2) = F, unchanged.

## Requests
48 of the 50 free daily requests used (40 slice + 8 full-payload; the single Inkling 403 was not counted by OpenRouter). 0 retries, 0 HTTP 429. Usage $0.