D3 ยท Adaptation: adjudication
Decide what the return changes
D3 asks which answers from communities require a change to the model, which describe local variations, and which remain unresolved. It reviews the substance of the answers collected in D2.
Where this stage starts
D2 has supplied local accounts, adaptation lines, independent verdicts and a record of missing returns. The aggregator has the fixed model those communities answered. It reads the answers by model element, so corrections to the same claim can be compared across communities.
What the aggregator does
For each proposed adaptation, the aggregator returns to its evidence and records a decision: accept it, modify it with a reason, reject it, or keep it as a supported local particular. Unsupported or unresolved claims stay pending. The original answers stay in the record.
Similar corrections from more than one community trigger a closer comparison. Repetition does not show a shared mechanism; the aggregator shows that the evidence supports the same change, condition or distinction. A claim supported in one community only can stay in that local account without becoming a new element of the population model.
The aggregator also examines the communities the model could not explain. Some organized in a way the model misses. Others fall outside its subject or have too little evidence to judge. These are different limits, and each is reported on its own. Disputed verdicts and tentative claims from the Field account also receive explicit rulings.
If the model changes, the affected communities see the revised version. The aggregator records how it identified them, and an uncertain scope means a return to all communities. The handler resumes their community coders, and the revised accounts and verdicts are checked again. A small pilot cannot stand in for every affected account.
What the next stage receives
D4 receives one frozen model version and the ruled record of adaptations tied to it. First fit, current local adequacy and shared corrections stay separate. The frozen version is the one served to the affected communities, or an earlier one.
What the output files contain
Markdown files hold prose; JSONL files hold one structured record per line.
Decisions on each answeradjudications.jsonl, verdicts-D.jsonl, findings.jsonl
adjudications.jsonl records rulings and reasons and
keeps the original lines. verdicts-D.jsonl keeps the reader
verdicts beside the later rulings. findings.jsonl records
claims that the return changed or bounded.
The revised model and coding trailmodel/, model/history.jsonl, tier_history.jsonl, scheme_v2.jsonl, assignments_final.jsonl
model/ holds the model; model/history.jsonl
records revisions. tier_history.jsonl records category
changes. Where categories change, scheme_v2.jsonl and
assignments_final.jsonl are rewritten so the levels of
evidence stay consistent.
Who was re-checked, and against whatrecheck-scope.json, handoff-D/recheck/, adaptations-recheck.jsonl, verdicts-recheck-D.jsonl
recheck-scope.json sorts communities into affected and
unaffected, with reasons. handoff-D/recheck/ contains the
revised tasks and model. adaptations-recheck.jsonl and
verdicts-recheck-D.jsonl keep the new answers beside the
earlier ones.
Exceptions and limitsuncomposed-D.md, exception-report-D.md
uncomposed-D.md explains the cases the model could not
account for. exception-report-D.md sorts adaptations into
shared changes, local particulars and unresolved claims.
The frozen result and handoffnotes-D3.md, handback-D3.md, out/_closes/model-D3.sha256
notes-D3.md explains what the re-check established and
what remains uncertain. handback-D3.md summarizes the
handoff. out/_closes/model-D3.sha256 names the exact model
version carried into the taxonomy.
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 3 legs
D3a: element group. Repeats: yes. Interview point: -.D3b: recheck. Repeats: no. Interview point: -.D3c: freeze and re-read. Repeats: no. Interview point: end.
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. D3 closed at D3c on 24 October 2026 in source commit 0990f903d412. The baseline catalog c070d990 predates that close. See the historical context for decision rows and per-community fork evidence.
Source versions and hashes identify the exact technical documents reproduced here.