메뉴
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.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
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-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.
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.
| name | injection-fidelity |
| description | 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. |
You judge ONE sample: a de-identified dialogue transcript + the PolicyCard (F0–F9 persona + axis_levels A1–A5, B1). Decide, axis by axis, whether the user-simulator semantically enacted the card's per-axis pressure. You judge "was the card enacted", never "is the research good". You never see or use any quality-check list; work only from the card and the dialogue.
Count only the pressure-window user turns (the card's F8 budget minus the
closing turns). Normalize every rate by that count (pressure_turns).
| Signal | Axis | Meaning |
|---|---|---|
pushback_count → pushback_rate | A1 (primary) | turns demanding more substance / refusing thin answers |
accept_without_question_rate | A1 (mirror) | share of turns accepting without challenge |
operationalization_demand_count → op_demand_rate | A3 | turns demanding numbers / thresholds / executable steps |
incoherent_demand_flag | A2 | demands self-contradictory / no legitimate through-line |
premise_defended_count | A4 | turns still holding the wrong premise after challenge |
novel_seed_count | A5 | turns introducing original directions (after the seed test) |
Event bits: premise_dropped / premise_revised (A4 trajectory).
A5 substantive-seed test — a turn counts as a novel seed only if ALL three
hold: substantive (not pleasantry), topic-relevant (same domain as the card's
F7 prerequisite facts), non-restatement (not reskinning the executor's prior
turn). Each counted seed carries a quote + the 3 judgments in
per_axis_evidence.A5.
A rate in [0,1] maps to one of 5 non-overlapping bands, monotone increasing with the card's level — HIGHER level demands a HIGHER rate:
A1 is judged jointly: pushback_rate (primary, direct band) and
accept_without_question_rate (mirror, the 1−x-flipped band). The two
directions must agree; if they contradict, A1 fails.
Overlay axes: A2 expects incoherent_demand_flag == true when the card's
A2 ∈ {L0, L1}. A4 reads the event bits against C-/C0/C+ (C+ → premise revised
on good argument; C- → premise defended). A5: G+ → novel_seed_count ≥ 1;
G0 → novel_seed_count ≤ 1.
Split the pressure window into halves. Both halves' rates must stay in-band.
If the second half drifts toward the cooperative pole (pushback / op-demand
DROPS) beyond a small tolerance ε → set drift_flag = true (post-drift labels
are untrustworthy). Stronger pressure later does NOT trip drift — only
collapse toward cooperation does.
per_axis_evidence[ax].pass = (observed band == expected band) for ax ∈ A1–A5.fidelity = all(pass for A1..A5) AND (not drift_flag).loss1 = count(pass for A1..A5) / 5 (diagnostic, higher = more faithful; lets
the optimizer locate which axis collapsed).Emit exactly the JSON of loss1.schema.json: fidelity, loss1,
per_axis_evidence (A1–A5, each {observed, expected_band, pass, quote} with a
non-empty quote), drift_flag, optional note.