菜单
Loss-2 judge (codex role). Over one topic's 6 shuffled research-design samples, pairwise-rank by quality using the D1–D5 standard. Emit the pairwise log; the harness computes the order and the ladder verdicts. Judge quality difference, never against academic standards.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | ladder-quality-order |
| description | Loss-2 judge (codex role). Over one topic's 6 shuffled research-design samples, pairwise-rank by quality using the D1–D5 standard. Emit the pairwise log; the harness computes the order and the ladder verdicts. Judge quality difference, never against academic standards. |
You rank ONE topic's 6 research-design samples (each a research_graph + research_result pair) by quality. The samples arrive SHUFFLED and anonymous — you see 6 positions (0–5), never their true rung id or config. You judge only on the D1–D5 standard:
Judge only on the D1–D5 standard above; never on academic-publication criteria of any kind. You never see any quality-check list.
You will be asked to compare two positions at a time. For each pair (i, j)
decide the winner (the higher-quality position) and give a one-line reason
grounded in D1–D5. Do not assign absolute scores — only pick a winner per pair.
The graph is structure-aware context; read it holistically, do not run any
checklist over it.
The harness enumerates all 15 pairs (i<j over 6 positions), Copeland-aggregates
your winners into an induced order, un-shuffles to true ids, and computes
Kendall τ against the intended order id0 > id1 > … > id5 (id0 = highest
quality). You only emit {winner, reason} per pair.
You will also be asked, K independent times, to compare the two extreme samples
(the harness picks them and presents them as just two options, A and B).
Return {"winner": "A" | "B"} — exactly the label of the higher-quality one.
Judge each call independently and honestly; do not try to be consistent with a
previous call you don't remember. (This is a two-way A/B label, distinct from
the position integers used in the pairwise rank above.)
If the topic carries a same-substance / different-framing triplet, rank it first. The order must NOT change with framing alone (buzzword vs neutral wording is not a quality difference under D1–D5). If your order tracks framing, say so in the reason — the harness will treat this topic's ladder as untrustworthy.
The harness assembles loss2.json: tau, monotonicity_pass
(τ ≥ the τ line AND no adjacent endpoint inversion), endpoint_separation_pass
(endpoint majority), rigor_floor_flag (endpoints a near-tie — a possible
genuine quality floor, NOT a tuning bug), and pairwise_log (your winners +
reasons, un-shuffled to true ids). Your only job: honest per-pair winners and
D1–D5 reasons.
Append research process and results to the current Phase's context file. Each append MUST contain >=500 lines of markdown covering both process and results. Use this skill at plan-designated checkpoint points — typically after each strategy completes or at key decision nodes within a research Phase.
Create a new context file for a research Phase. Called once at Phase start to initialize the file that subsequent context-checkpoint calls will append to. Use this skill whenever a new research Phase begins and a fresh context file is needed.
The optimizer brain for the ladder-foundry pretraining loop. Runs the two-level nested batch loop, delegates gating to gate_eval, attributes a failing batch to one weight (attribute-first), and recovers from disk after compaction. Control flow is fully scripted; only the backprop attribution is a judgment call.
Loss-1 judge (codex role). Given one sample's de-identified dialogue and its PolicyCard, decide axis-by-axis whether the user-simulator enacted the card's per-axis pressure. Judge enactment of the card, never whether the research is good.
Closing skill for the research-executor, loaded as the last step of formated-specs. Summarize the design just produced into one research-result JSON fenced block in your reply. Do not execute the research.
Spec-slot skill for the research-executor. Emit the 4-layer DARE orchestration of the assigned topic as one research-graph JSON fenced block in your reply. Replaces the generic spec-writing step.
