How to check every number in the paper, from the shipped evidence
This data layer is the request-level evidence behind AI Service QoS: An Independent Outside-In Measurement Methodology, with Case Study 1. It exists so that the paper's claims can be checked without asking anyone's permission — including mine.
One command does the checking
git clone https://huggingface.co/datasets/2makeitwork/paper_ai_Qos ds && cd ds
pip install pyyaml # the only dependency
python3 scripts/analyze.py; echo "exit code: $?"
It re-derives every published figure from evidence/ and asserts the prose against what it finds — a number that no longer follows from the data fails the run, with the claim, the expected value and the measured value printed. Exit code 0 means the paper and the evidence agree at the byte level it can reach. scripts/verify_dataset_card.py does the same for every quantity this card states.
What is here
294 requests across two collection rounds — round 1, 2026-09-01 to 09-09 (135 requests, 12 conversations sess-01…sess-12), and round 2, 2026-09-10 (159 requests, 32 conversations S1…S34 with two retired identifiers). One provider (Qwen3.8-Max through the Qoder CN client on Alibaba Cloud Model Studio), one subscription plan, client version 1.1.3. Per-request timing, outcome classification by the client's own code plus the provider's own code, context occupancy where the client reported it, and the configuration the provider pushed to the client.
Two things worth knowing before you grep
Count from evidence/logs/chat_finish_census.txt, not from a convenient extract. That file is a census of every streamed chat_finish outcome. The extracts are keyed on different log patterns: the timeout code 80408 appears zero times in token_size_lines.txt and quota_error_lines.txt, because the client writes banner events to a separate log file. Reading one extract and reporting its count is how the same incident got understated by more than half — 13 where the census says 33 — and that error is in this project's own history, recorded in the source repository's release notes.
The rounds use different pseudonym namespaces on purpose. Round 1's identifiers are ordinal and assigned in first-appearance order, so any new conversation would renumber the rest and invalidate the frozen bundle. Round 2 therefore uses a distinct scheme, and the two rounds share no identifier — deliberate, because it means neither can corroborate the other.
What the data will not tell you
No response content and no prompts — the client writes neither to the log, and a content-bearing copy of a shared coding conversation is not something this bundle can carry. No cache-hit counts, no provider-internal state. Three negative results are published with the same weight as the positive ones: the client-side panel shows every request regardless of origin (refuting a claimed display filter), server-side prompt-cache residency does not decline during the stall window, and a request designed to inflate an agent's context occupancy by making it read files did not produce a measurable rise.
Citing it
Cite the tag, not a branch; name the collection round you read; and cite the client version plus the service, not the model alone — the finding is partly that a client-side watchdog gets misattributed to a model. The paper is paper_ai_QoS.md in the source repository, the case report is report_qwenAliServiceQuality.md, and the released software is archived under concept identifier 10.5281/zenodo.22701430. Text, tables and data are CC-BY-4.0; the scripts here are MIT.
Not included, and why
The raw log trees stay out on purpose: they contain local filesystem paths and real conversation identifiers, neither of which any transformation of this data can recover or needs. A reproduction run against them is described in methodology.md, and the anonymisation policy is anonymization.md.
Corrections, gaps and questions are better raised on the source repository than answered here — the measurement code, the release notes and the working history of every figure are there.