"""Check that this release reproduces the MACE study without model calls. cd code && mkdir -p .cache && tar -xzf ../cache/llm.tar.gz -C .cache AGMEM_DATA=$PWD/../benchmark python verify_release.py 1. Rebuilds every evaluation job (all packages, including the 2026-11-05 amendment P8 and P9) from benchmark/ and study/, and compares each job, with its exact memory list, to study/evaluation_jobs.jsonl.gz. 2. Recomputes the request hash of every caption and every text diagnosis and checks that the cached reply exists, which means the prompt is reproduced byte for byte and a rerun replays the recorded answer at no cost. 3. Checks that each cached diagnosis reply gives the prediction stored in runs//predictions.jsonl. Agentic (P6) and image-reference (P7) calls are not hashed here; their records are in cache/agentic.tar.gz and the traces. """ import gzip import json import sys from collections import Counter from pathlib import Path from agmem import config as C, data, kb as KB, llm, packet, prompts from mace10 import amend, conditions, jobs def diag_key(j, caps, kbd, classes): """Request hash of a text diagnosis, built exactly as mace10/run.py builds the request.""" if j["condition"] == "sage": text, _ = packet.build(j["crop"], kbd, caps[j["image_id"]], None) else: memory = [{"id": "m_" + m["image_id"], "label_shown": m["label"], "label_true": m["label"], "caption": caps[m["image_id"]]} for m in j["memory"]] text, _, refs = conditions.build(j["variant"], j["crop"], kbd, caps[j["image_id"]], memory, j["image_id"], classes, j.get("retrieved_ids")) assert not refs content = [llm.image_block(data.path(j["image_id"])), {"type": "text", "text": prompts.diagnosis_user(j["crop"], classes, text)}] return llm.sha({"m": C.MODEL, "s": prompts.DIAG_SYS, "c": content}) REL = Path(__file__).resolve().parent.parent saved = {j["job_id"]: j for j in (json.loads(l) for l in gzip.open(REL / "study/evaluation_jobs.jsonl.gz", "rt"))} built = {j["job_id"]: j for j in jobs.build() + amend.build()} diff = [k for k in saved if json.dumps(built.get(k), sort_keys=True) != json.dumps(saved[k], sort_keys=True)] print(f"1. jobs rebuilt {len(built)}, saved {len(saved)}, differing {len(diff)}") caps = json.loads((REL / "captions.json").read_text()) miss_cap = [i for i in caps if not jobs.cached(jobs.caption_key(i, data.images().loc[i, "crop"]))] print(f"2. captions {len(caps)}, cache misses {len(miss_cap)}") pred = {} for p in (REL / "runs").glob("P*/predictions.jsonl"): for l in p.open(): r = json.loads(l); pred[r["job_id"]] = r kbs = {} hit, miss, agree = Counter(), Counter(), Counter() for j in saved.values(): if j["condition"] in ("agentic", "imageref", "imageref_dev"): continue c = j["crop"] if c not in kbs: kbs[c] = (KB.build(c, data.classes(c)), data.classes(c)) k = diag_key(j, caps, *kbs[c]) tag = (j["pkg"], j["condition"]) if k and jobs.cached(k): hit[tag] += 1 rep = json.loads((C.CACHE / "llm" / f"{k}.json").read_text()) p = prompts.clean_prediction(llm.parse_json(rep["text"]), kbs[c][1]) if j["job_id"] in pred and p.get("prediction") == pred[j["job_id"]]["prediction"]: agree[tag] += 1 else: miss[tag] += 1 for t in sorted(set(hit) | set(miss)): print(f" {t[0]:<4} {t[1]:<13} cache hits {hit[t]:>5} misses {miss[t]:>3} reply = saved prediction {agree[t]:>5}") ok = not diff and not miss_cap and not sum(miss.values()) and sum(agree.values()) == sum(hit.values()) print("3. all text diagnoses replayed:", sum(hit.values()), "| replies matching saved predictions:", sum(agree.values())) print("RELEASE REPRODUCES" if ok else "MISMATCH FOUND") sys.exit(0 if ok else 1)