#!/usr/bin/env python3 """The paper's headline finding, recomputed from the deposit alone. WHY THIS EXISTS --------------- The trajectory split is what the paper is for: of the 43 CITES-listed Arctic species, 40 carry a comparable Red List trajectory, and 20 of those are unchanged, 12 improved and 8 deteriorated. Every other number in the paper supports that one. Until now it could only be recomputed against the live database, which means a reviewer could not check it, and the three rows for it in the traceability table stood as PENDING CONVENTION. It turns out the whole series is already in the deposit: figure5_source_data.csv inside source_data.zip carries every global assessment for every species with its year and category. So the finding is reproducible from the deposited package with no credentials, no network and no database, which is the strongest form of the claim the paper makes about itself. This check recomputes it and holds three things to each other: the deposit, the traceability table, and the manuscript. If any one of them moves without the others, it fails and says which. THE SERIES, from 21 October 2026 -------------------------------- Everything above describes the paper as first submitted, when the trajectory was computed on global assessments only. The lead author decided on 21 October 2026 that European regional assessments are included. Per species the Global and Europe assessments now form ONE series ordered by (year, assessment_id), built by redlist_scale.series_rows; in a year with several assessments the higher assessment id stands (redlist_scale.pick_in_year), the transition count keeps both rows of a same-year pair, and the Mediterranean porbeagle assessment (redlist_scale.NOT_EUROPEAN_IDS) stays out. On that series 41 of the 43 species carry a comparable trajectory: 18 unchanged, 13 improved, 10 deteriorated, 2 without one (roseroot and Stejneger's beaked whale), with 43 transitions (22 improvements, 21 deteriorations). Exit 0 pass, 1 fail, 3 unmeasured. """ import collections import csv, io, pathlib, re, sys, zipfile from collections import Counter, defaultdict # The deposit ships no IUCN Red List category (IUCN's terms forbid it), so # every Source Data read goes through the shim that prefers a local rebuild # and stops with UNMEASURED when none has been made. sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent)) import iucn_series # noqa: E402 HERE = pathlib.Path(__file__).resolve().parent DEPOSIT = HERE.parent REPO = DEPOSIT.parent ZIP = DEPOSIT / "source_data.zip" MEMBER = "Source Data/figure5_source_data.csv" TRACE = DEPOSIT / "results_traceability.csv" # The manuscript the numbers are held to: the 21 October 2026 text with the # lead author's decision (global and European assessments) applied. DOC = (REPO / "response" / "from-tom-2026-10-21" / "skuggi" / "MANUSCRIPT-21Oct-European-TRACKED.docx") # The IUCN ladder, least to most threatened. Pre-2001 codes that carry a # direction are placed on it; the ones that do not (K, V, E, R, I, T, NR, DD, # NA) are absent on purpose, so a species whose only assessments are those has # no comparable trajectory rather than a guessed one. # # This table used to be written out here, and that is how the project came to # hold two of them. On 6 October 2026 it was found to disagree with # ingest/redlist_scale.py — the module that calls itself the one scale and that # the figure generators and the claims verifier both import — on two codes: # LR/cd, which this table read as NT and that module excluded, and RE, which # this table placed at 5 and that module keeps off the global ladder. The # manuscript reported this table's species split and that module's transition # count, in the same sentence. The crosswalk the Methods cite (ref 76) settles # LR/cd and says nothing about RE, so the module now carries the one table and # this file imports it. RE comes off, which moves the transition count from 38 # to 36 and leaves the species split untouched. # # The module is imported from beside this file, because a deposit that reaches # back into the repository is only as durable as the repository; build_deposit.py # copies redlist_scale.py in for exactly this reason, and the two copies are now # the same file. # Importing a sibling here writes __pycache__ INTO the deposit, and # verify_deposit_accounted.py rejects bytecode in a citable archive. Running # these scripts in place is the intended use, so the run must not leave # residue behind: the gate was failing on 7 October 2026 for no reason other # than that the verifiers had been run. sys.dont_write_bytecode = True sys.path.insert(0, str(HERE)) import redlist_scale as _rl # Every code the module can rank under the convention in force: the modern # ladder, the 2001 renames, and whatever the selected crosswalk adds. Built by # asking the module rather than by restating it, so this file cannot drift from # it again. RANK = {code: _rl.rank(code) for code in (list(_rl.RANK) + list(_rl.EQUIVALENT) + list(_rl.CROSSWALK_LIN2025)) if _rl.rank(code) is not None} # Extended to 2025 on the lead author's instruction of 1 October 2026. It costs # the split nothing, because every 2025 global assessment repeats the category # its species already held; the one number it moves is the transition count. WINDOW = (1982, 2025) # SCOPE was "Global" until 21 October 2026; the series is now global and # European, selected by redlist_scale.series_rows / in_series. def trajectories(): """(improved, unchanged, deteriorated, comparable, species without one).""" with zipfile.ZipFile(ZIP) as z: rows = list(csv.DictReader(io.StringIO(iucn_series.read_member(z, MEMBER).decode()))) everyone = {r["scientific_name"] for r in rows} series = {sp: [(int(r["year"]), r["status"]) for r in recs if r["status"] in RANK] for sp, recs in _rl.series_rows(rows, window=WINDOW).items()} imp = unch = det = 0 comparable = [] for sp, a in series.items(): if len(a) < 2: continue d = RANK[a[-1][1]] - RANK[a[0][1]] comparable.append(sp) if d < 0: imp += 1 elif d == 0: unch += 1 else: det += 1 return imp, unch, det, len(comparable), sorted(everyone - set(comparable)) # Added 6 October 2026. Round u5's rejection lens objected that V and E are # discarded "as carrying no direction, which is indefensible: these are the # pre-1994 threatened categories". The justification was wrong (the real reason # is superseded criteria) and the decision needed measuring rather than # defending. The Methods and Table 3 now report the sensitivity, so this gate # recomputes it: reading V as VU and E as EN must leave the four totals # identical and move exactly these four species. # # 21 October 2026: on the global-and-European series this is no longer true. # Reading V as VU and E as EN moves TWO species (the bowhead whale, deteriorated # to improved, and the polar bear, deteriorated to unchanged) and so changes the # totals - from 13 improved, 18 unchanged, 10 deteriorated and 2 without a # trajectory to 14, 19, 8 and 2. The earlier sentence ("leaves the four totals # identical and moves four species") described the global series. The numbers # are computed by response/compute_european_extras.py on the same series. VE_MOVERS = {"Balaena mysticetus": ("deteriorated", "improved"), "Ursus maritimus": ("deteriorated", "unchanged")} VE_ALT_SPLIT = (14, 19, 8, 2) # improved, unchanged, deteriorated, unknown def ve_sensitivity(): """(published split, split with V=VU and E=EN, {species: (before, after)}).""" with zipfile.ZipFile(ZIP) as z: rows = [r for r in csv.DictReader(io.StringIO(iucn_series.read_member(z, MEMBER).decode())) if _rl.in_series(r) and WINDOW[0] <= int(r["year"]) <= WINDOW[1]] everyone = {r["scientific_name"] for r in rows} def classify(extra): rank = dict(RANK) rank.update(extra) series = {sp: [(int(r["year"]), r["status"]) for r in recs if r["status"] in rank] for sp, recs in _rl.series_rows(rows).items()} out = {} for sp in everyone: a = series.get(sp, []) if len(a) < 2: out[sp] = "unknown" continue d = rank[a[-1][1]] - rank[a[0][1]] out[sp] = "improved" if d < 0 else ("deteriorated" if d > 0 else "unchanged") return out def split(c): n = Counter(c.values()) return (n["improved"], n["unchanged"], n["deteriorated"], n["unknown"]) base, alt = classify({}), classify({"V": 2, "E": 3}) moved = {sp: (base[sp], alt[sp]) for sp in base if base[sp] != alt[sp]} return split(base), split(alt), moved # Added 5 October 2026. The manuscript reported the current category # distribution as "4 CR, 5 EN, 9 VU, 3 NT, 21 LC, and 1 DD species" and the # deposit gives 4 CR, 5 EN, 7 VU, 4 NT, 22 LC, 1 DD. Three of the six counts # were wrong, and both sequences sum to 43, so no gate that checks totals or # shares could see it. It had no traceability claim either, which is the # condition AGENTS.md names: a number the paper states and the harness does not # recompute is how the two drift apart. C23 to C28 now carry it. # # 21 October 2026: RE joins the list. One species' latest assessment of either # scope is Regionally Extinct, so the six categories above sum to 42 and the # forty-third species needs its own claim. CATEGORIES = ["CR", "EN", "VU", "NT", "LC", "DD", "RE"] DIST_CLAIMS = dict(zip(CATEGORIES, ["C23", "C24", "C25", "C26", "C27", "C28", "C148"])) def current_distribution(): """Most recent assessment per species (global or European), counted by category. The tie-break is the part that matters, and this function had the same accidental one as verify_claims_offline.py until 6 October 2026: "keep the row with the later year", which on a tie keeps whichever row the CSV listed first. Seven species carry two global assessments in their final year, and for two of them the rows disagree on the category - Bubo scandiacus 2021 (LC against VU) and Odobenus rosmarus 2025 (NT against VU). In both the row that came first is the one the deposit marks is_latest=False, so both verifiers reported a superseded assessment as current, and because they shared the same accident they agreed with each other and the disagreement was invisible. Two deposited verifiers agreeing on a wrong rule is worse than one being wrong. The rule lives in redlist_scale.pick_in_year: within a year the higher assessment id is the later publication and it stands, with is_latest as a cross-check where it exists. It is NOT the flag alone - the flag marks a species' current assessment, so it is absent from mid-series ties and cannot settle them. The two verifiers call the same function, which is what keeps them in step; sharing an accidental rule is what let them agree while both were wrong. """ with zipfile.ZipFile(ZIP) as z: rows = list(csv.DictReader(io.StringIO(iucn_series.read_member(z, MEMBER).decode()))) # Before anything is counted: the tie-break guard has to be seen working. # Series are ordered by (year, assessment_id), so no value here depends on # the alphabet; the final-year tie is settled by pick_in_year, which refuses # when is_latest contradicts the higher id. if not _rl.self_test_tiebreak(): raise SystemExit("the tie-break guard cannot be trusted; nothing checked") latest = {} for who, recs in _rl.series_rows(rows, window=WINDOW).items(): ymax = int(recs[-1]["year"]) final = [r for r in recs if int(r["year"]) == ymax] latest[who] = (ymax, _rl.pick_in_year(final)["status"]) return collections.Counter(s for _, s in latest.values()) def traced(): out = {} for r in csv.DictReader(open(TRACE)): out[r["claim_id"]] = r["value"] return out def self_test(): """The ladder has to be able to tell the three verdicts apart.""" assert RANK["LC"] < RANK["VU"] < RANK["CR"], "the ladder is not ordered" assert "DD" not in RANK and "K" not in RANK, \ "a category with no direction is on the ladder" assert RANK["LR/lc"] == RANK["LC"], "the legacy least-concern code moved" # The two codes the project held two answers about. Asserted here so a # future edit to either table fails this gate rather than moving a number. assert RANK["LR/cd"] == RANK["NT"], \ "LR/cd is not Near Threatened; the Methods cite a crosswalk that says it is" assert "RE" not in RANK, \ "Regionally Extinct is on the global ladder; it is a regional outcome" assert _rl.convention() == _rl.CONVENTION_LIN2025, \ ("the scale module is not on the convention the Methods state; this gate " "would recompute the paper's headline finding under a different scale") print(" self-test: PASS (the ladder is ordered, categories carrying no " "direction are off it, LR/cd reads as NT and RE is not on it)", file=sys.stderr) def genuine_transitions(): """Category transitions on the global-and-European series, by the one scale. Until 21 October 2026 this counted global assessments only and the manuscript reported 36 transitions (18 improvements, 18 deteriorations). The lead author then decided that European regional assessments are included: the Global and Europe assessments of a species form one series ordered by (year, assessment_id), both rows of a same-year pair stay in the sequence as the published method always counted them, and the Mediterranean porbeagle assessment stays out. redlist_scale.py is the one scale: a non-comparable assessment breaks the chain rather than being skipped over, so no transition is manufactured across a gap. """ sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent)) import redlist_scale as rl with zipfile.ZipFile(ZIP) as z: rows = list(csv.DictReader(io.StringIO(iucn_series.read_member(z, MEMBER).decode()))) out = [] for name, recs in rl.series_rows(rows, window=WINDOW).items(): out.extend(rl.transitions([(int(r["year"]), r["status"]) for r in recs])) return out, sum(1 for t in out if t[-1] == "improved") def main(): self_test() if not ZIP.exists(): print(f"UNMEASURED: {ZIP.name} is not in the deposit.", file=sys.stderr) return 3 imp, unch, det, comparable, without = trajectories() trans, ups = genuine_transitions() faults = [] t = traced() for claim, got, what in (("C19", imp, "improved"), ("C20", unch, "unchanged"), ("C21", det, "deteriorated")): want = t.get(claim, "") if want != str(got): faults.append(f"the traceability table has {claim} ({what}) as " f"{want!r}; the deposit computes {got}") # --deposit-only compares the deposit with the traceability table and stops # there. The pass that writes the manuscript sentence calls it this way, # because the full check holds the manuscript to a sentence that pass has not # written yet: a gate that blocks the edit satisfying it is a circle, not a # check. After the pass runs, the full gate holds the three to each other. # The current distribution, claim by claim and then in the sentence. dist = current_distribution() if sum(dist.values()) != comparable + len(without): faults.append(f"the distribution covers {sum(dist.values())} species; " f"the study set is {comparable + len(without)}") for cat in CATEGORIES: claim = DIST_CLAIMS[cat] want, got = t.get(claim, ""), dist.get(cat, 0) if want != str(got): faults.append(f"the traceability table has {claim} (current {cat}) " f"as {want!r}; the deposit computes {got}") # The V/E sensitivity the Methods and Table 3 now report. ve_base, ve_alt, ve_moved = ve_sensitivity() if ve_base != (imp, unch, det, len(without)): faults.append(f"the V/E check recomputes the published split as " f"{ve_base}, which is not {(imp, unch, det, len(without))}") if ve_alt != VE_ALT_SPLIT: faults.append(f"reading V as VU and E as EN is recorded as giving " f"{VE_ALT_SPLIT} (improved, unchanged, deteriorated, " f"unknown) on the global-and-European series; the deposit " f"gives {ve_alt} against {ve_base} unread") if ve_moved != VE_MOVERS: faults.append(f"the species that move between classes when V and E are " f"read as VU and EN are recorded as {sorted(VE_MOVERS)}; the deposit moves {sorted(ve_moved)} " f"({ve_moved})") deposit_only = "--deposit-only" in sys.argv if deposit_only: print(" note: --deposit-only, so the manuscript was not compared", file=sys.stderr) elif DOC.exists(): sys.path.insert(0, str(REPO / "response")) import live_text live = re.sub(r"\s+", " ", live_text.from_docx(DOC)) total = comparable + len(without) pct = lambda k: f"{100 * k / total:.1f}" sentence = (f"Of {total} species, {imp} improved, {unch} were " f"unchanged, {det} deteriorated and {len(without)} lack a " f"comparable trajectory. Across {WINDOW[0]}\u2013{WINDOW[1]}, " f"the record contains {ups} improving and " f"{len(trans) - ups} deteriorating category transitions.") if sentence not in live: faults.append("the manuscript no longer states what the deposit " f"computes. The deposit gives: {sentence}") # The transition count ANYWHERE in the paper, not only in that one # sentence. On 7 October 2026 the count was corrected from 34 to 36 in # Section 4.7 and this gate went green, while Figure 4's caption one # page away still ended "...is not one of the 34 counted in Section # 4.7". The caption was invisible to a check that looks for a sentence. # # A number the paper repeats is a number that must be repaired # everywhere, so every count-of-transitions in the accepted text is # held to the deposit, caption and prose alike. The notes are stripped # first: a margin note quotes the number it is withdrawing. prose = re.sub(r"\s+", " ", re.sub(r"\[(?:TOM|MINT|SET|TOM og MAGN\u00daS|TOM/MINT|TOM/SET)[^\]]*\]", "", live)) for m in re.finditer(r"(\d+)\s+(?:category |genuine )?transitions?", prose): if int(m.group(1)) != len(trans): faults.append( f"the paper says {m.group(1)} transitions where the deposit " f"computes {len(trans)}: ..." + prose[max(0, m.start() - 70):m.end() + 40].strip() + "...") for m in re.finditer(r"the (\d+) counted in Section", prose): if int(m.group(1)) != len(trans): faults.append( f"a cross-reference says {m.group(1)} where the deposit " f"computes {len(trans)}: ..." + prose[max(0, m.start() - 70):m.end() + 40].strip() + "...") if dist["DD"] != 0 or dist["RE"] != 1: faults.append(f"the manuscript's sentence names one Regionally " f"Extinct species and none Data Deficient; the deposit " f"gives {dist['RE']} RE and {dist['DD']} DD") stated = ("the current distribution includes " + ", ".join(f"{dist[c]} {c}" for c in CATEGORIES[:3]) + f", {dist['NT']} NT and {dist['LC']} LC species, and one " + "Regionally Extinct (RE)") if stated not in live: i = live.find("the current distribution includes") faults.append( "the manuscript's current distribution is not what the deposit " f"computes.\n deposit: {stated}\n paper : " + (live[i:i + 120] if i >= 0 else "the sentence is absent")) # The two shares the Abstract and the Discussion carry, on the same base. hold = unch + imp for share, where in ((f"{100 * hold / total:.1f}%", "hold or improve"), (f"{100 * imp / total:.1f}%", "improve outright")): if share not in live: faults.append(f"the manuscript does not state {share} anywhere, " f"which is what the deposit gives for {where} of " f"all {total} species") else: print(" note: the manuscript is not built, so only the deposit and the " "traceability table were compared", file=sys.stderr) if len(without) != 2: faults.append(f"{len(without)} species have no comparable trajectory; " f"the paper reports 2 ({', '.join(without)})") if faults: print(f"\nTHE HEADLINE FINDING ({len(faults)}):\n", file=sys.stderr) for f in faults: print(f" {f}", file=sys.stderr) return 1 print(f"transitions: {len(trans)} category transitions between " f"comparable global and European assessments ({ups} improvements, " f"{len(trans) - ups} deteriorations) in {WINDOW[0]}-{WINDOW[1]}") print(f"trajectories: {comparable} comparable, {unch} unchanged, {imp} " f"improved, {det} deteriorated, {len(without)} without a comparable " f"trajectory ({', '.join(without)}) — recomputed from the deposited " f"Source Data, with no database and no network, and matching both the " f"traceability table and the manuscript") return 0 if __name__ == "__main__": sys.exit(main())