| """Clarify two language actions in every reference-feasible r1 cell, remotely.""" |
|
|
| from __future__ import annotations |
|
|
| import json |
| from pathlib import Path |
|
|
| from build_cliff_pilot import dispersed_blocks, fact_blocks |
| from run_evaluate import oracle, read_jsonl, write_jsonl |
|
|
|
|
| def language_block(profile, presentation): |
| a = {action["id"]: action for action in profile["actions"]} |
| paid, application, receipt = (a[key]["allowed_days"][0] for key in ("L_paid", "W", "L_waiver")) |
| paid_cost = f"，当天支付{a['L_paid']['cost']}元不退" if presentation == "C" else "" |
| receipt_cost = f"，当天支付{a['L_waiver']['cost']}元不退" if presentation == "C" else "" |
| return ( |
| "语言有两种等效办法。付费考试办法：" |
| f"第{paid}天参加英语测试，记作动作L_paid{paid_cost}。" |
| "校内等效证明办法由两个独立动作构成：" |
| f"第{application}天申请校内英语证明减免费，记作动作W，免费。" |
| f"第{receipt}天领取等效英文证明，记作动作L_waiver{receipt_cost}，须先完成W至少14天。" |
| "选择校内等效证明办法时，W和L_waiver必须各办一次并各列一项；领取不能替代申请，" |
| "申请也不会自动完成领取。L_paid与L_waiver互斥。") |
|
|
|
|
| def main(): |
| source = Path("data/cliff_dev_r1") |
| destination = Path("data/cliff_dev_r1_lang") |
| profiles = read_jsonl(source / "profiles.jsonl") |
| by_id = {profile["id"]: profile for profile in profiles} |
| references = [{"profile_id": profile["id"], **oracle(profile)} for profile in profiles] |
| thresholds = {row["profile_id"]: row["minimum_initial_cash"] for row in references} |
| requests = [] |
| for original in read_jsonl(source / "requests.jsonl"): |
| if original["B"] < thresholds[original["profile_id"]]: |
| continue |
| profile = by_id[original["profile_id"]] |
| presentation = original["presentation"] |
| old = (fact_blocks(profile) if presentation == "C" else dispersed_blocks(profile))[1][1] |
| new = language_block(profile, presentation) |
| if old not in original["prompt"]: |
| raise ValueError("Source language block absent: " + original["id"]) |
| requests.append({**original, "id": original["id"] + "__lang", |
| "source_request_id": original["id"], |
| "language_action_clarification": True, |
| "prompt": original["prompt"].replace(old, new, 1)}) |
| write_jsonl(destination / "profiles.jsonl", profiles) |
| write_jsonl(destination / "oracle.jsonl", references) |
| write_jsonl(destination / "requests.jsonl", requests) |
| design = json.loads((source / "design.json").read_text(encoding="utf-8")) |
| design.update(version="cliff_dev_r1_lang", request_count=len(requests), |
| source="all reference-feasible cells from cliff_dev_r1, without failure selection", |
| inference_buffer_selection="two original reference-feasible balances per profile; original physical controls are retained in r1", |
| prompt_changes_from_r1="only language notice: independent W and L_waiver sentences and explicit each-once/each-listed semantics", |
| paired_facts="all facts, calendars, graph and cost locations unchanged; C costs remain beside actions, D costs remain in finance", |
| scientific_scope="adaptive language-action granularity diagnostic; not a financial-improvement claim") |
| (destination / "design.json").write_text(json.dumps(design, ensure_ascii=False, indent=2), encoding="utf-8") |
| (destination / "change_note.txt").write_text( |
| "After all 12 r1 finals and the three positive-claim structural failures were read, the next bounded diagnostic " |
| "clarifies that W is a free application and L_waiver is a later, paid receipt of the equivalent proof. They are " |
| "independent actions, each performed and listed once when choosing this language method; neither substitutes " |
| "for the other or completes automatically. The old facts already named both actions, their dates and dependency, " |
| "and required free prerequisites; this clarification does not retroactively treat the old facts as absent. " |
| "Every original reference-feasible C/D cell is retained (not just failed cells), with the same two profiles, " |
| "two feasible balances, seed1729, native4B settings, budget and batch4. Old below-R* controls are not rerun. " |
| "No fee, date, graph, finance-notice location, boolean claim contract or cash rule changes. All r1 outputs remain. " |
| "Any improvement is a post-result action-granularity diagnostic, not evidence of financial reasoning improvement. " |
| "Counterfactual replays retain all three original balances without new model calls.\n", encoding="utf-8") |
| print(json.dumps(design, ensure_ascii=False), flush=True) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
