#!/usr/bin/env python3 """analyze.py — recompute every number the report quotes, from evidence/ only. Reads the anonymized evidence tree, derives the statistics cited in report_qwenAliServiceQuality.md, writes analysis/summary_tables.md, and asserts each expected value. Exit code 1 = a report number no longer matches the evidence. Usage: python3 analyze.py (run from scripts/) """ from __future__ import annotations import hashlib import json import re import subprocess import sys from collections import Counter, defaultdict from datetime import datetime from pathlib import Path from statistics import median as statistics_median EV = Path(__file__).resolve().parents[1] / "evidence" OUT = Path(__file__).resolve().parents[1] / "analysis" / "summary_tables.md" TS = "%Y-%m-%d %H:%M:%S" fails: list[str] = [] checks_done = 0 # how many checks this run performed, i.e. the number prose may quote def check(name: str, got, want) -> None: global checks_done checks_done += 1 ok = got == want print(f" {'PASS' if ok else 'FAIL'} {name}: {got}" + ("" if ok else f" (report says {want})")) if not ok: fails.append(name) def ts_of(line: str) -> datetime: return datetime.strptime(line[:19], TS) # ---- timeout events --------------------------------------------------------- rows = (EV / "logs/timeout_resume_events.csv").read_text().splitlines()[1:] codes = Counter(r.split(",")[2] for r in rows) per_day = Counter(r[:10] for r in rows) times = sorted(ts_of(r) for r in rows) # densest 75-minute block best = max(range(len(times)), key=lambda i: sum(1 for t in times if 0 <= (t - times[i]).total_seconds() <= 75 * 60)) block_n = sum(1 for t in times if 0 <= (t - times[best]).total_seconds() <= 75 * 60) # ---- quota events (cluster lines within 2 s) -------------------------------- qtimes = [ts_of(ln) for ln in (EV / "logs/quota_error_lines.txt").read_text().splitlines() if re.match(r"^\d{4}-\d\d-\d\d \d\d:\d\d:\d\d", ln)] events: list[list[datetime]] = [] for t in sorted(qtimes): if events and (t - events[-1][-1]).total_seconds() <= 2: events[-1].append(t) else: events.append([t]) payloads = sorted(set(re.findall(r'"message":"([^"]{20,140})', (EV / "logs/quota_error_lines.txt").read_text()))) codes_seen = sorted(set(re.findall(r'"code":"?(-?\d+)"?', (EV / "logs/quota_error_lines.txt").read_text()))) # ---- chat_finish census ------------------------------------------------------ census = (EV / "logs/chat_finish_census.txt").read_text().splitlines() c_total = c_success = 0 for ln in census: m = re.match(r"\s*(\d+) (.*)", ln) if not m or '"command"' in ln: continue c_total += int(m.group(1)) if "success:200" in m.group(2): c_success += int(m.group(1)) # ---- sessions ---------------------------------------------------------------- sess = [ln.split(",") for ln in (EV / "stats/session_size_stats.csv").read_text().splitlines()[1:]] big = sorted(sess, key=lambda r: -int(r[3]))[:3] # ---- request lifecycle (frozen from a log snapshot via scripts/latency_from_logs.py) -- lc = [ln.split(",") for ln in (EV / "stats/request_lifecycle.csv").read_text().splitlines()[1:]] _fl = lambda x: float(x) if x else None _p50 = lambda v: (round(sorted(v)[min(len(v) - 1, int(0.5 * (len(v) - 1)))], 1) if v else None) _succ = [r for r in lc if r[2] == "S"] _stall = [r for r in lc if r[2] == "T"] _tto = [r for r in lc if r[2] == "T" and len(r) > 7 and r[7] == "resume_tool_call"] _tdi = [r for r in lc if r[2] == "T" and len(r) > 7 and r[7] == "permission_request"] _ttft = [x for x in map(lambda r: _fl(r[3]), lc) if x is not None] _tot = [_fl(r[4]) for r in _succ if _fl(r[4]) is not None and int(r[6] or 0) == 0] _stv = [_fl(r[5]) for r in _stall if _fl(r[5]) is not None] _fau = sum(1 for r in _succ if int(r[6] or 0) == 0) # ---- task lifecycle (quest.log; one master-mode task per conversation) -------- # The client's task object carries the conversation's own id (the log line names # it twice: "taskId" and "executionSessionId"). Every open IDE window writes the # same event into its own quest.log, so (session, status) repeats are clustered at # ≤2 s — the same rule used above for the quota markers. TLINE = re.compile(r"^\[(window\d+)\] (\d{4}-\d\d-\d\d \d\d:\d\d:\d\d\.\d{3}) \[\w+\] " r"task\.status\.update(\.skipNonQuestTask)? " r'\{"taskId":"(sess-\d+)","status":"(\w+)"') tl = [] for ln in (EV / "logs/task_status_lines.txt").read_text().splitlines(): m = TLINE.match(ln) if m: tl.append(m.groups()) tl_dupes = sum(1 for g in tl if g[2]) # .skipNonQuestTask twins tl_main = sorted((g for g in tl if not g[2]), key=lambda g: g[1]) task_ev: list[list] = [] # de-duplicated events for win, ts, _s, sid, st in tl_main: t = datetime.strptime(ts, TS + ".%f") if (task_ev and task_ev[-1][1] == sid and task_ev[-1][2] == st and (t - task_ev[-1][0]).total_seconds() <= 2): task_ev[-1][3].add(win) continue task_ev.append([t, sid, st, {win}]) task_status = Counter(e[2] for e in task_ev) task_windows = sum(1 for e in task_ev if len(e[3]) > 1) run_state: dict[str, datetime] = {} run_iv: list[tuple] = [] for t, sid, st, _w in task_ev: if st == "Running": if sid not in run_state: run_state[sid] = t else: if sid in run_state: run_iv.append((run_state.pop(sid), t, sid)) run_dur = sorted((b - a).total_seconds() for a, b, _ in run_iv) conc = sorted([(a, 1) for a, _, _ in run_iv] + [(b, -1) for a, b, _ in run_iv], key=lambda x: (x[0], x[1])) conc_now = conc_max = 0 for _t, d in conc: conc_now += d conc_max = max(conc_max, conc_now) # unordered pairs of conversations whose Running spans overlap in time conc_pairs = sum(1 for i, (s1, e1, _n1) in enumerate(run_iv) for s2, e2, _n2 in run_iv[i + 1:] if s2 < e1 and e2 > s1) # transition spans: Running -> ActionRequired is "sent, then went silent"; the # duration distribution of successful runs says whether the watchdog caps total # duration or only the silence between chunks. wd_span, ok_span = [], [] for (t1, s1, st1, _w1), (t2, s2, st2, _w2) in zip(task_ev, task_ev[1:]): if s1 != s2: continue gap = (t2 - t1).total_seconds() if st1 == 'Running' and st2 == 'ActionRequired': wd_span.append(gap) elif st1 == 'Running' and st2 == 'Completed': ok_span.append(gap) wd_mode, wd_mode_n = Counter(round(g) for g in wd_span).most_common(1)[0] ok_over = sum(1 for g in ok_span if g > wd_mode) # ActionRequired = "the banner is up, nothing is generating": pair it with the # 80408 resume-button events independently captured from the agent logs. Those # are read at full millisecond precision (`times` above is second-truncated). ar = [(t, sid) for t, sid, st, _w in task_ev if st == "ActionRequired"] resume_ms = [datetime.strptime(r.split(",")[0], TS + ".%f") for r in rows] ar_offsets = [min((abs((t - a).total_seconds()) for a, _ in ar), default=9e9) for t in resume_ms] ar_paired = sum(1 for g in ar_offsets if g <= 0.5) ar_max_off = max(ar_offsets, default=0.0) ar_per_sess = Counter(sid for t, sid in ar) # ---- per-interaction phase timing (the client's own state machine) ----------- # One line per phase boundary of a turn: `State transition: A -> B, trigger: T`. # The interval spent in state B runs from that line to the conversation's next # one, so `prompting` = send pressed → first streamed chunk, `streaming` = the # generation itself, `suspended` = a dialog or the timeout banner is up. PLINE = re.compile(r"^\[(window\d+)\] (\d{4}-\d\d-\d\d \d\d:\d\d:\d\d\.\d{3}) \[\w+\] " r"\[ACPProgressStateMachine\] State transition: (\w+) -> (\w+), " r"trigger: (.+), sessionId: (sess-\d+)") ph_raw = [] for ln in (EV / "logs/phase_transitions.txt").read_text().splitlines(): m = PLINE.match(ln) if m: ph_raw.append((datetime.strptime(m.group(2), TS + ".%f"), m.group(3), m.group(4), m.group(5), m.group(6))) ph_by_sess: dict[str, list] = defaultdict(list) for t, _frm, to, trig, sid in sorted(ph_raw): ph_by_sess[sid].append((t, to, trig.split(":")[0], trig)) wait_first, fastfail, gen_ok, hold_banner, hold_perm = [], [], [], [], [] suspend_cause = Counter() prompts = 0 for sid, seq in ph_by_sess.items(): for _t, to, fam, _trig in seq: if to == "suspended": suspend_cause[fam] += 1 for (t1, state1, fam1, _tr1), (t2, state2, fam2, _tr2) in zip(seq, seq[1:]): secs = (t2 - t1).total_seconds() if fam1 == "user_message_chunk": prompts += 1 if state2 == "streaming": wait_first.append(secs) elif state2 == "error": fastfail.append(secs) if state1 == "streaming" and state2 == "completed" and fam2 == "chat_finish": gen_ok.append(secs) if state1 == "suspended": (hold_banner if fam1 == "resume_tool_call" else hold_perm).append(secs) # ---- report assertions ------------------------------------------------------- print("assertions against the report:") check("timeout events", len(rows), 33) check("all timeouts code 80408", dict(codes), {"80408": 33}) check("timeouts Sep 7/8/9", dict(per_day), {"2026-09-07": 1, "2026-09-08": 4, "2026-09-09": 28}) check("densest 75-min block", block_n, 20) check("quota events clustered", len(events), 18) check("quota payload variants", len(payloads), 3) check("80411 never fired (absent from every log extract)", sum("80411" in p.read_text(errors="replace") for p in (EV / "logs").rglob("*")), 0, ) # the code exists only in the dictionary (config/), never on an event line check("chat_finish lines", c_total, 245) # echo-artifact line removed by anonymize.py check("chat_finish successes", c_success, 225) check("largest session entries", big[0][3], "626") check("lifecycle requests", len(lc), 308) check("lifecycle sessions", len({r[0] for r in lc}), 10) check("lifecycle successes (S)", len(_succ), 211) check("lifecycle stalls (T)", len(_stall), 71) check("lifecycle timeout parks (trigger resume_tool_call)", len(_tto), 33) check("lifecycle dialog parks (trigger permission_request)", len(_tdi), 38) check("lifecycle first-attempt usable", _fau, 179) check("lifecycle first-chunk median seconds", _p50(_ttft), 11.3) check("lifecycle clean-success total median seconds", _p50(_tot), 120.0) check("lifecycle time-to-stall median seconds", _p50(_stv), 305.4) # ---- incident 2 (2026-09-10 collection round; evidence/incident_20260910/) ---- # Regenerate: python3 scripts/latency_from_logs.py raw_snapshots/20260910T155718/logs \ # --since 2026-09-10T00:00 --emit-csv evidence/incident_20260910/request_lifecycle.csv i2 = [ln.split(",") for ln in (EV / "incident_20260910/request_lifecycle.csv").read_text().splitlines()[1:]] i2_to = [r for r in i2 if len(r) > 7 and r[7] == "resume_tool_call"] i2_dlg = [r for r in i2 if len(r) > 7 and r[7] == "permission_request"] i2_s_ok = [int(r[9]) for r in i2 if r[2] == "S" and len(r) > 9 and r[9]] i2_to_ctx = [int(r[9]) for r in i2_to if len(r) > 9 and r[9]] check("incident-2 requests", len(i2), 62) check("incident-2 timeout parks (80408 banner)", len(i2_to), 5) check("incident-2 permission-dialog parks", len(i2_dlg), 2) check("incident-2 smallest context that timed out (tokens)", min(i2_to_ctx), 490789) check("incident-2 largest context that completed (tokens)", max(i2_s_ok), 452447) check("incident-2 timeout send-to-stall, min/max seconds", (round(min(float(r[5]) for r in i2_to), 1), round(max(float(r[5]) for r in i2_to), 1)), (125.3, 4860.7)) check("task-status lines (incl. .skipNonQuestTask twins)", len(tl), 1252) check("conversations with task lifecycle events", len({e[1] for e in task_ev}), 9) check("task lifecycle events after ≤2 s de-dup", len(task_ev), 524) check("task lifecycle statuses seen", dict(task_status), {"Running": 258, "Completed": 208, "ActionRequired": 33, "Error": 18, "Stopped": 7}) check("ActionRequired count == timeout resume count", len(ar), len(rows)) check("every timeout has an ActionRequired ≤0.5 s away", ar_paired, len(rows)) check("per-conversation timeout split (session-level, was window-level)", dict(ar_per_sess), {"sess-16": 1, "sess-11": 5, "sess-03": 23, "sess-04": 4}) check("closed Running intervals", len(run_iv), 257) check("Running spans still open at the extract horizon (log ends mid-run)", len(run_state), 1) check("max simultaneous Running tasks", conc_max, 2) check("overlapping Running pairs (two tasks at once)", conc_pairs, 10) check("watchdog: modal Running→ActionRequired span (s)", wd_mode, 61) check("banners at that modal span", wd_mode_n, 16) check("successful runs longer than the watchdog span", (ok_over, len(ok_span)), (155, 194)) check("phase transition lines", len(ph_raw), 885) check("conversations with phase data", len(ph_by_sess), 9) check("prompts sent (user_message_chunk)", prompts, 252) check("what suspended a turn", dict(suspend_cause), {"resume_tool_call": 33, "permission_request": 37}) check("banner holds that were resumed (sess-03's last never was)", len(hold_banner), 32) check("median wait for the first streamed chunk (s)", round(statistics_median(sorted(wait_first)), 1), 9.6) check("prompts answered with an error before any chunk", len(fastfail), 4) check("successful generations longer than the watchdog span", (sum(1 for g in gen_ok if g > wd_mode), len(gen_ok)), (149, 208)) check("total operator hold on banners (min)", round(sum(hold_banner) / 60, 1), 90.5) check("total operator hold on permission dialogs (min)", round(sum(hold_perm) / 60, 1), 122.0) # ---- cross-document consistency (the paper and the case report must not drift) -- # Both were corrected during the second collection round; a fix applied to one and # not the other is exactly the failure mode this guards, so the retired wordings are # forbidden outright and the shared figures must appear in both narratives. DOCS = [doc for doc in (EV.parent / n for n in ("report_qwenAliServiceQuality.md", "paper_ai_QoS.md", "methodology.md", "README.md", "abstract.md")) if doc.exists()] # The dataset repository ships the data layer only, so the narratives may be absent; # their absence is reported as a skip, never as a silent pass. NARRATIVES = [doc for doc in DOCS if doc.name in ("report_qwenAliServiceQuality.md", "paper_ai_QoS.md")] RETIRED = ["no per-request token usage is logged", "No per-request token usage exists client-side", "checked — zero hits", "same window3 session", "Deepest session in window", "62 requests across three", "requests, three conversations, columns", "largest *active* contexts", "is not documented and was not measured"] SHARED = ["490,789", "494,193", "452,447", "117,153", "episode"] drift = [] for doc in DOCS: body = doc.read_text(errors="replace") drift += [f"{doc.name}: retired wording still present: {phrase!r}" for phrase in RETIRED if phrase in body] for doc in NARRATIVES: body = doc.read_text(errors="replace") drift += [f"{doc.name}: shared figure missing: {figure!r}" for figure in SHARED if figure not in body] if NARRATIVES: check("no drift between the paper and the case report", drift, []) else: print(" SKIP no drift check: the paper and the case report are not in this " "checkout (data-layer repository)") # ---- release notice: one text, one version, present everywhere ---- NOTICE_DATE = "2026-09-11" RELEASED = "1.0.2" # the released version; CITATION.cff must carry this same string NOTICE_MARK = f"**Released {NOTICE_DATE} as v{RELEASED}**" PUBLISHED_NAMES = ("paper_ai_QoS.md", "report_qwenAliServiceQuality.md", "README.md", "abstract.md", "methodology.md", "anonymization.md", "ask_vendor.md", "DATA_REQUEST.md", "DATASET_CARD.md") def tracked_here() -> set[str] | None: """Paths git tracks, but only at the top level of a repository: `git ls-files` lists what is tracked *under* the current directory, so asking it from a staging directory would return an empty set and make every published file look unpublished.""" try: prefix = subprocess.run(["git", "-C", str(EV.parent), "rev-parse", "--show-prefix"], capture_output=True, text=True, timeout=30) if prefix.returncode != 0 or prefix.stdout.strip(): return None listing = subprocess.run(["git", "-C", str(EV.parent), "ls-files"], capture_output=True, text=True, timeout=30) except (OSError, subprocess.SubprocessError): return None return set(listing.stdout.split()) if listing.returncode == 0 else None TRACKED = tracked_here() # "Published document" means a document in the published set, not merely a file on this disk: # zenodo/README.md moved to the private workshop repository and is no longer a reader's file. PUB_DOCS = [doc for doc in (EV.parent / n for n in PUBLISHED_NAMES) if doc.exists() and (TRACKED is None or doc.relative_to(EV.parent).as_posix() in TRACKED)] def where(doc) -> str: """Relative path, never a bare name: zenodo/README.md and README.md share a name, and a failure that names only "README.md" sends the reader to the wrong file.""" return doc.relative_to(EV.parent).as_posix() notice_problems = [where(doc) for doc in PUB_DOCS if NOTICE_MARK not in doc.read_text(errors="replace")] if PUB_DOCS: check(f"release notice {NOTICE_MARK} present in every published document", notice_problems, []) # The version lives in two places, so one is checked against the other: a notice that names a # version CITATION.cff does not would send a citation to a snapshot that is not this one. cff = EV.parent / "CITATION.cff" cff_found = re.search(r'^version: "([^"]+)"', cff.read_text(errors="replace"), re.M) if cff.exists() else None if cff.exists(): check("CITATION.cff and the release notices carry the same version", cff_found.group(1) if cff_found else "unreadable", RELEASED) late = [doc.name for doc in PUB_DOCS if datetime.fromtimestamp(doc.stat().st_mtime).date() > datetime.strptime(NOTICE_DATE, "%Y-%m-%d").date()] if late: print(" ADVISORY: edited after the release date — a released snapshot is frozen, so " "these edits belong in the next version: " + ", ".join(sorted(late))) else: print(" SKIP release notice check: no published documents in this checkout (data layer)") # ---- authorship: one handle, no institution, anywhere in the public set ---- BYLINE = "independent, unaffiliated" # By relative path, not by name: zenodo/README.md and the root README.md share a name. BYLINE_DOCS = ("paper_ai_QoS.md", "DATASET_CARD.md", "abstract.md", "README.md") byline_docs = [doc for doc in PUB_DOCS if doc.parent == EV.parent and doc.name in BYLINE_DOCS] if byline_docs: check("authorship stated as independent and unaffiliated in every document that carries a byline", [where(doc) for doc in byline_docs if BYLINE not in doc.read_text(errors="replace")], []) zen = [p for p in (EV.parent / "zenodo" / "deposit-dataset.json", EV.parent / ".zenodo.json") if p.exists()] bad = [] for p in zen: creators = json.loads(p.read_text()).get("creators") or [] if len(creators) != 1 or creators[0].get("name") != "2makeitwork" \ or creators[0].get("affiliation", ""): bad.append(p.name) if zen: check("archive metadata carries the pseudonym with an empty affiliation", bad, []) # ---- published documents must not send a reader to a file that is not published ---- PRIVATE = ["PUBLICATION.md", "prior-works.md", "methodology_guidance"] DISCLOSED = ["not published", "not part of", "deliberately", "private", "ignored", "working notes", "kept local", "no reading obligation"] private_problems = [] for doc in [d for d in PUB_DOCS if d.name != "PUBLICATION.md"]: body = doc.read_text(errors="replace") for n, line in enumerate(body.splitlines(), 1): named = [p for p in PRIVATE if p in line] if named and not any(w in line.lower() for w in DISCLOSED): private_problems.append(f"{where(doc)}:{n} names {named[0]} without saying it is unpublished") if PUB_DOCS: check("no published document points a reader at a private file without saying so", private_problems, []) # ---- counts belong to the script's output, not to prose ---- # scripts/analyze.py prints one line per check, so the total is whatever a run reports. Quoting # it in a document is a claim with a shelf life: the dataset card said "50", then "52", while the # script was already at 56, and no gate failed. A count is therefore allowed only where the # passage says it was measured and on which date - the same date the notices carry, so bumping # the revision date forces the counts to be re-measured rather than left behind. COUNT_RE = re.compile(r"\b(\d{2,3})\s+(?:assertions?\b|published figures?\b|checks\b)") DATE_RE = re.compile(r"measured|re-runs|attempt", re.I) def section_of(lines: list[str], at: int) -> str: """The markdown section a line belongs to, by heading boundaries.""" start = 0 for i in range(at - 1, -1, -1): if lines[i].startswith("#"): start = i break end = len(lines) for i in range(at, len(lines)): if lines[i].startswith("#"): end = i break return "\n".join(lines[start:end]) count_problems = [] for doc in PUB_DOCS: body_lines = doc.read_text(errors="replace").splitlines() for n, line in enumerate(body_lines, 1): found = COUNT_RE.search(line) if not found: continue section = section_of(body_lines, n - 1) if NOTICE_DATE not in section or not DATE_RE.search(section): count_problems.append(f"{where(doc)}:{n}: {found.group(0)!r} quoted in a section that " f"does not date the measurement (need {NOTICE_DATE} + 'measured')") if PUB_DOCS: check("no published document quotes a check or figure count unless its section dates the measurement", count_problems, []) # ---- a published heading must not route text to a venue ---- # The released v1.0.0 tag carried "## One-line entry (repository subtitle, listing, tweet)" and # "## Short entry (about 300 characters: post summary field, search result, card lead)" in # abstract.md: notes for whoever posts the text, sitting in a file a stranger downloads. Venue # routing in a heading is the signature, so the heading is what is checked - prose may name a # medium freely. VENUE_WORDS = ("subtitle", "listing", "tweet", "summary field", "post summary", "search result", "card lead", "post body", "collection note", "use everywhere", "use verbatim", "for listings", "to paste", "paste this", "headline to use") editorial_problems = [] for doc in PUB_DOCS: for n, line in enumerate(doc.read_text(errors="replace").splitlines(), 1): low = line.lower() if line.startswith("#") and "(" in line and any(w in low for w in VENUE_WORDS): editorial_problems.append(f"{where(doc)}:{n}: heading routes text to a venue: " f"{line.strip()[:64]}") if PUB_DOCS: check("no published heading carries editorial routing instructions", editorial_problems, []) # ---- a published script may only tell someone to run something that exists ---- # scripts/preflight.sh invoked `python3 tools/verify_docs.py` unconditionally. tools/ is local-only, # so a clean clone of the release tag finished with NOT READY TO PUBLISH — a red light caused by # tooling nobody published. Same defect as the dataset card's un-runnable command, one layer down, # which is why this checks scripts too and asks git rather than the disk: the author's disk has # everything, so presence proves nothing about what was shipped. CALL_RE = re.compile(r"(?:python3|bash)\s+((?:scripts|tools|zenodo|analysis)/[\w./-]+)") def tracked_at_root() -> set[str] | None: """What git tracks, but only when this checkout is the top level of its own repository.""" root = EV.parent try: prefix = subprocess.run(["git", "-C", str(root), "rev-parse", "--show-prefix"], capture_output=True, text=True, timeout=30) if prefix.returncode != 0 or prefix.stdout.strip(): return None listing = subprocess.run(["git", "-C", str(root), "ls-files"], capture_output=True, text=True, timeout=30) except (OSError, subprocess.SubprocessError): return None return set(listing.stdout.split()) if listing.returncode == 0 else None TRACKED = TRACKED if TRACKED is not None else tracked_at_root() LOCAL_WORDS = ("local-only", "not published", "not in this repository", "never published") script_problems = [] scripts_dir = EV.parent / "scripts" if scripts_dir.is_dir(): # Iterate what is *published*, not what is on this disk: the first version of this check walked # scripts/ and reported an untracked local helper that no reader has, which is precisely the # mistake the check exists to catch. if TRACKED is not None: candidates = sorted(EV.parent / rel for rel in TRACKED if rel.startswith(("scripts/", "zenodo/")) and rel.endswith((".sh", ".py"))) else: candidates = sorted(list(scripts_dir.glob("*.sh")) + list(scripts_dir.glob("*.py")) + list((EV.parent / "zenodo").glob("*.sh"))) for script in candidates: if not script.is_file() or script.name == "analyze.py": continue # this file's own strings are not instructions lines = script.read_text(errors="replace").splitlines() for n, line in enumerate(lines, 1): for found in CALL_RE.finditer(line): target = found.group(1) shipped = (target in TRACKED) if TRACKED is not None else (EV.parent / target).exists() window = " ".join(lines[max(0, n - 4):n + 3]).lower() if not shipped and not any(w in window for w in LOCAL_WORDS): script_problems.append(f"{script.name}:{n} runs {target}, which is not shipped " f"and is not marked local-only here") if TRACKED is not None or scripts_dir.is_dir(): check("no published script tells a reader to run something that is not published", script_problems, []) # ---- the release note describes the artifact, not the machinery that publishes it ---- # A GitHub release body is reader-facing text. Workflow commentary in it - what we chose not # to attach and why, which integration has to be enabled first - is addressed to the person # running the release, and once published it sits on a public page. So the note is authored # in RELEASE_NOTES.md, scanned here, and pushed with # gh release edit --notes-file RELEASE_NOTES.md # The list was narrowed on 2026-09-11. It used to forbid the words "zenodo", "integration" and # "archived" outright, which was right when the only such sentence was pipeline instruction and is # wrong now that a published archive with an identifier exists: naming the archive and its # identifier is what a reader of a release needs. What stays forbidden is the mechanics - what has # to be enabled, what a draft does not trigger, which page to click, which token to revoke. MACHINERY = ["must be enabled", "never archived", "binary asset", "draft release", "edit page", "access token", "revoke", "integration watches", "sync_draft", "deposit.sh", "before the release is published", "not published as a"] release_notes = EV.parent / "RELEASE_NOTES.md" if release_notes.exists(): note_hits = [w for w in MACHINERY if w in release_notes.read_text(errors="replace").lower()] check("RELEASE_NOTES.md describes the release, not the publishing machinery", note_hits, []) # ---- CITATION.cff must be a citation file the format can read ---- # The file invented a `versions:` key and declared cff-version 1.3.0, neither of which exists: # checked against the official schema (github.com/citation-file-format, fetched 2026-09-11), whose # additionalProperties is false, so an invented key is a validation error and the invented history # block would be dropped by every converter. Keys are now prose comments, and the check fails on # any top-level key outside the schema's list. CFF_KEYS = {"abstract", "authors", "cff-version", "commit", "contact", "date-released", "doi", "identifiers", "keywords", "license", "license-url", "message", "preferred-citation", "references", "repository", "repository-artifact", "repository-code", "title", "type", "url", "version"} cff_file = EV.parent / "CITATION.cff" if cff_file.exists(): cff_text = cff_file.read_text(errors="replace") cff_top = [m.group(1) for m in re.finditer(r"^([A-Za-z][A-Za-z-]*):", cff_text, re.M)] check("CITATION.cff uses only keys the Citation File Format defines", sorted({k for k in cff_top if k not in CFF_KEYS}), []) check("CITATION.cff declares a cff-version the schema accepts", re.findall(r"^cff-version: (\S+)", cff_text, re.M), ["1.2.0"]) # ---- summary tables ---------------------------------------------------------- out = ["# Analysis summary (generated by scripts/analyze.py from evidence/)", "", "## Timeout (80408) events", "", f"- total: **{len(rows)}**, all `reasonForCode: 80408`", f"- per day: " + ", ".join(f"{d}: {n}" for d, n in sorted(per_day.items())), f"- densest 75-minute block: **{block_n} events** (the resume-button loop)", ""] out += ["## Quota/account events", "", f"- log lines: {len(qtimes)} → clustered (≤2 s apart): **{len(events)} events**", f"- distinct codes seen: {', '.join(codes_seen)}", f"- distinct message wordings: {len(payloads)}", ""] for p in payloads: out.append(f" - `{p[:100]}`") out += ["", "## chat_finish census", "", f"- {c_total} lines total: {c_success} `success:200`, {c_total - c_success} error payloads,", f" **zero attributable to the {len(rows)} timeout events** (no chat_finish line exists for them)", ""] out += ["## Three largest sessions (transcript entries)", "", "| session | entries | jsonl bytes |", "|---|---|---|"] for r in big: out.append(f"| `{r[1]}` | {r[3]} | {r[2]} |") out += ["", "## Task lifecycle (`quest.log` — the client's `task` = one conversation)", "", f"- {len(tl)} log lines → **{len(task_ev)} events** after ≤2 s de-dup " f"({task_windows} events replicated into ≥2 IDE windows, " f"{tl_dupes} `.skipNonQuestTask` twins dropped)", "- statuses: " + ", ".join(f"{k} {v}" for k, v in sorted(task_status.items())), f"- `Running` spans: {len(run_iv)} closed" + (f", {len(run_state)} still open at the extract horizon" if run_state else "") + f"; median {statistics_median(run_dur):.0f} s, " f"90th percentile {run_dur[int(0.9 * len(run_dur))]:.0f} s, max {max(run_dur):.0f} s", f"- **max simultaneous `Running` tasks: {conc_max}** (never 3+; " f"{conc_pairs} overlapping pairs of conversations), " "`error.code.task.running.quota.exceeded` fired 0 times, so that ceiling is " "only bounded ≥2", f"- `ActionRequired` (banner up, nothing generating): {len(ar)} events, 1:1 with " f"the {len(rows)} timeout resume lines (largest offset {ar_max_off * 1000:.0f} ms)", "- conversation-level split of those timeouts, which supersedes the " "window-level correlation in report §4.2: " + ", ".join(f"`{s}` {n}" for s, n in sorted(ar_per_sess.items())), f"- watchdog measured off the state machine: **{wd_mode} s of silence** before the " f"banner ({wd_mode_n} of {len(wd_span)} banners at exactly that span; the rest are " "mid-run stalls after some output), and " f"{ok_over} of {len(ok_span)} *successful* runs ran longer than that span — so it " "caps inactivity, not total duration", ""] out += ["", "## Per-interaction phase timing (`[ACPProgressStateMachine]`)", "", f"- {len(ph_raw)} transitions over {len(ph_by_sess)} conversations: " f"{prompts} prompts sent, {len(wait_first)} answered with a first streamed chunk, " f"{len(fastfail)} rejected before any chunk, {len(gen_ok)} generations finished", f"- **wait for the first chunk** (`prompting` → `streaming`, the UI's " f"'Thinking…'): median **{statistics_median(sorted(wait_first)):.1f} s**, " f"90th percentile {sorted(wait_first)[int(0.9 * len(wait_first))]:.0f} s, " f"max {max(wait_first):.0f} s, min {min(wait_first):.1f} s", f"- generation time (`streaming` → `chat_finish:success`): median " f"{statistics_median(sorted(gen_ok)):.0f} s, " f"90th percentile {sorted(gen_ok)[int(0.9 * len(gen_ok))]:.0f} s, max {max(gen_ok):.0f} s; " f"{sum(1 for g in gen_ok if g > wd_mode)} of {len(gen_ok)} exceed the " f"{wd_mode} s watchdog span", "- what suspended a turn: " + ", ".join(f"{k} {v}" for k, v in sorted(suspend_cause.items())) + " (`resume_tool_call` = the 80408 banner, a third stream confirming 33)", f"- operator hold: banners {len(hold_banner)} resumed, median " f"{statistics_median(sorted(hold_banner)):.0f} s, total {sum(hold_banner) / 60:.1f} min; " f"permission dialogs {len(hold_perm)}, median " f"{statistics_median(sorted(hold_perm)):.0f} s, total {sum(hold_perm) / 60:.1f} min", ""] # The footer carries both: a content hash that machines compare, and a second-precision # timestamp a person reads. It is appended at the very end so the hash covers the whole # document; scripts/check_release_sync.py compares that hash and ignores the time, so a # rebuild with the same statistics is recognised as identical rather than reported as drift. out += ["", "## Request lifecycle (frozen log snapshot; scripts/latency_from_logs.py)", "", f"- {len(lc)} requests across {len({r[0] for r in lc})} sessions — S {len(_succ)}, T {len(_stall)} (timeout {len(_tto)} / dialog {len(_tdi)}), F/U {len(lc) - len(_succ) - len(_stall)}", f"- first-attempt usable (0-retry S) **{_fau}/{len(_succ)} ({100*_fau//len(_succ)}%)**", f"- first-chunk wait median {_p50(_ttft)}s (n={len(_ttft)}); clean-success turn total " f"median {_p50(_tot)}s (n={len(_tot)}); send-to-stall median {_p50(_stv)}s (n={len(_stv)})", ""] out += ["", "## Incident 2 — collection round of 2026-09-10 (`evidence/incident_20260910/`)", "", f"- {len(i2)} request episodes (55 distinct submitted prompts) in 3 conversations on the observation day: " f"{sum(1 for r in i2 if r[2] == 'S')} completed, {len(i2_to)} timeout parks, " f"{len(i2_dlg)} dialog parks, {sum(1 for r in i2 if r[2] in ('F', 'U'))} failed/cancelled", f"- the five timeouts all sit in one conversation at **{min(i2_to_ctx):,} tokens**; " f"the largest context that still completed that day is **{max(i2_s_ok):,} tokens** — " f"a {max(i2_s_ok):,}-versus-{min(i2_to_ctx):,} separation on the day " "(round 1 pooled data does NOT separate cleanly; see report §4.2)", f"- send-to-stall spanned {min(float(r[5]) for r in i2_to):.0f}s to " f"{max(float(r[5]) for r in i2_to):.0f}s (the upper value is a banner left parked " "for 81 minutes)", "- paired experiment: the same conversation stalled five times at 494,193 tokens, then " "after the client's context-compaction control reported 117,153 tokens " "(`Compression API call completed for session compact`) a request carrying 85,115 " "tokens completed in 10.1 s — report §9", "- three negative results from this round: tab visibility (a hidden conversation " "completed in 2.58 s), server-side cache residency across a 27.7-minute idle gap " "(17.11 s versus 17.58 s at the same size), and inflating an agent's context by " "asking it to read files (one such turn added 851 tokens)", ""] OUT.parent.mkdir(parents=True, exist_ok=True) # The stamp records when the content last changed, not when the script last ran: a verification # run must not make the working tree dirty, which is what preflight.sh's clean-tree check relies # on. The hash covers the whole document, footer excluded, so two copies with the same hash hold # the same statistics whatever their stamps say. body = "\n".join(out).rstrip("\n") # canonical: trailing newlines are not part of the content content_sha = hashlib.sha256(body.encode()).hexdigest() previous = OUT.read_text(errors="replace") if OUT.exists() else "" MARKER = "\nContent hash `sha256:" idx = previous.rfind(MARKER) previous_body = previous[:idx].rstrip("\n") if idx != -1 else None keep_stamp = previous_body is not None and hashlib.sha256(previous_body.encode()).hexdigest() == content_sha if not keep_stamp: OUT.write_text(body + f"\n\nContent hash `sha256:{content_sha}` - generated " f"{datetime.now():%Y-%m-%d %H:%M:%S} - rerun: " "`python3 scripts/analyze.py`\n") print(f"\nwrote {OUT.relative_to(EV.parent)}" if not keep_stamp else f"{OUT.relative_to(EV.parent)} unchanged (content hash matches, stamp kept)") # ---- a count quoted in the documents must equal this run's count ---- # The dated-measurement rule above keeps a quoted number from going stale silently; this keeps # it from being wrong. Deliberately only in the source repository: a data-layer checkout skips # the checks that read the paper and the case report, so its smaller total is correct behaviour, # not drift. Printed as a note rather than a PASS line, so "how many assertions passed" keeps # meaning "how many check() calls succeeded" and nothing else. if NARRATIVES: quoted = {doc.name: [int(m.group(1)) for m in re.finditer(r"\b(\d+) assertions pass", doc.read_text(errors="replace"))] for doc in PUB_DOCS} wrong = {name: nums for name, nums in quoted.items() if any(n != checks_done for n in nums)} if wrong: fails.append("documented assertion count") print(f" FAIL {wrong} quoted in the documents, but this run performed {checks_done} " f"checks - update the number where it is quoted", file=sys.stderr) elif any(quoted.values()): print(f" note: the assertion count quoted in the documents equals this run " f"({checks_done} checks)") if fails: print(f"FAILED: {fails}", file=sys.stderr) sys.exit(1) print("all report numbers verified against evidence")