D2 ยท Adaptation: labelers and verdicts
Collect the communities' answers
D2 brings back two kinds of answer: each community coder's account of its community in the model's terms, and independent readers' judgments of how well the model fits. It checks that the answers arrived intact before anyone interprets them.
Where this stage starts
D1 has prepared and fixed the tasks. The adaptation handler resumes each community's own community coder, with its original conversation and files. Separate readers receive the verdict packets. The aggregator's check leg opens after these jobs have returned their files.
What the community coders, readers and aggregator do
Each community coder rewrites its local account in the population model's terms. It records where the model applies, where it needs detail or a condition, where the community contradicts it, and where a new idea may be needed. Each claim cites evidence from that community. A community that needs no adaptation still submits an explicit answer; silence counts as a missing return.
Two independent readers judge whether the model accounts for the community, accounts for it with something left over, fails to account for it, or cannot be judged from the evidence. Their reasons and disagreements stay visible. The community coder's own assessment sits beside these verdicts, and each is kept.
The ingestion tools check model versions, identifiers, citations, required fields and completeness. The handler can send a failed return back for repair. The aggregator then reads the ingestion report and a seeded sample of ten local accounts: do they describe how those communities organized, or only repeat the population model?
These checks show whether a return can be read and traced. Whether a proposed correction holds is decided in D3. A high disagreement rate or many exceptions is reported as found, and no answer is adjusted to improve the numbers.
What the next stage receives
D3 receives the accepted adaptation records, both verdicts, the local accounts and the explicit gaps. D3 judges the substance of the corrections; D2 only collects them.
What the output files contain
Markdown files hold prose; JSONL files hold one structured record per line.
The adaptation ledgeradaptations.jsonl
adaptations.jsonl collects the validated return lines
with their evidence and model version. "Validated" means the required
structural checks passed; the aggregator rules on each claim in D3.
The independent judgmentsverdicts-D.jsonl
verdicts-D.jsonl keeps the readers' fit judgments and
reasons beside the community coders' assessments. Each reading stays
visible, with no vote between them.
The local accounts and comparison tablerooms/<sub>/model-D.md, verdict-D, _views/rooms-x-elements.tsv
rooms/<sub>/model-D.md holds each returned local
account, with its verdict-D record beside it.
_views/rooms-x-elements.tsv sets communities against model
elements, so D3 can read all the evidence on one claim together.
What arrived, failed or remains missingingest-D.md, notes-D2.md, RUN.json, run_manifest.json, handback-D2.md
ingest-D.md reports accepted lines, failed rules and
missing communities. notes-D2.md records the sample checks
and cautions on substance. RUN.json records return counts;
run_manifest.json records the models used.
handback-D2.md carries defects and open operational
questions forward.
Full stage description
Read the complete technical card for exact tasks, reading boundaries, checks and closing rules. The explanation above is a guide; the card holds the instructions.
Show how the stage is divided into its 1 legs
D2a: check. Repeats: no. Interview point: -.
Show the changes recorded for this run
Historical instructions and settings
Study-specific settings: the 15 October decision specified a 5.6 resume followed by a 6.1 return. Native D2 turns checked for all 14 shared communities use gpt-6.1-sol, medium. Their run.json files keep resumed_with=5.6, a field that does not name the model of those answer turns. On 21 October the D3 re-check was assigned gpt-6.1-sol, medium. Its return records name 6.1 and the same D2 sessions, but omit effort. This website review did not inspect every D3 native turn or all 705 D2 histories. See the historical context for decision rows and per-community fork evidence.
Source versions and hashes identify the exact technical documents reproduced here.