---
kind: "stage-reading-key"
stage: "D2"
title: "Adaptation: labelers and verdicts"
purpose: "Ask the original community leads what the model explains and misses."
status: "finished"
closed_legs: 1
publication: "c070d990be88"
status_note: "This describes source publication c070d990be88, not live state."
authority: "This is a reading key. The linked contract and card hold the rules."
---

# D2: Adaptation: labelers and verdicts

Ask the original community leads what the model explains and misses.

This is a reading key. The linked contract and card hold the rules.

## 1. Reading order

1. [Contract](../reading/contract.md), [method](../reading/method.md), [standing instructions](../reading/standing.md).
2. Your recorded platform: [Claude](../reading/platform-claude.md) or [Codex](../reading/platform-codex.md).
3. [Full D2 card](../reading/stage-D2.md), then the run's head, diary and handler deliveries as that card prescribes.

## 2. Inputs and permitted reading

Use the full card's startup packet and reading boundaries. Do not use files from later stages merely because this repository contains them. Historical interviews use the files in their selected bundle.

## 3. Outputs, quoted from the card

- `adaptations.jsonl` (ingested, validated)
- `verdicts-D.jsonl`
- `ingest-D.md`
- `_views/rooms-x-elements.tsv` (generated)
- `rooms/<sub>/model-D.md` and `verdict-D` copied
from the drops
- `notes-D2.md`
- the D2 counts in `RUN.json`
- a `run_manifest.json` entry for
the labelers' model and the readers' model
- `handback-D2.md`.

## 4. Legs

| Leg | Role | Repeats | Interview point |
|---|---|---|---|
| `D2a` | check | no | - |

## 5. Checks and handback

The stage card governs its checks. The [handler](../reading/handler.md) and [leg controller](../reading/legs.md) govern opening, close, pause and recovery. Do not infer permission to start from this page. Keep notes and the head current; the conversation and filesystem must match at any resumable close.

## 6. Published record

Status: **finished**, 1 recorded closed legs. This describes source publication `c070d990be88`, not live state.

## 7. 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 retain resumed_with=5.6; that field is not 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.

[Historical context and evidence](../guides/history.md).


[Checkpoint interviews](../interviews.html?stage=D2) · [Source hashes](../sources/manifest.json)