{"id":"87f178f2-b6bc-4f3a-9270-8ef0ecdd6f1e","arxiv_id":"2607.19476","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A streaming detector that reads codec motion vectors from the compressed bitstream achieves anytime-valid false-positive control with a single fixed threshold and a priced deferral frontier at about 10^5 MACs per GOP.","lead":"This paper reframes AI-generated-video detection as a streaming problem, using the motion field already stored in the compressed bitstream as a CPU-only early-warning stage. A running-max score with one pre-calibrated threshold keeps the false-positive rate controlled at the moment the detector decides, and uncertain clips are escalated to a heavier model on demand.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"E3 paired gain inconsistent with flip counts: 60 vs 130 on ~27k clips gives ~+0.003, not +0.027; headline 0.75→0.78 needs code verification.","rationale":"Reader's verdict is CONDITIONAL and I concur. My stress-test did not find a flaw in Proposition 3; the monotone-max argument is elementary and correct. The soft spot is the arithmetic of the headline paired comparison. The same concern appears in the reader's rationale as a blocker, but their stated weakest_assumption is the calibration-null transfer; I would rank the E3 inconsistency first because it is a purely internal contradiction, not a judgment about distribution shift. The paper's own cross-dataset FPR drift is a real limitation but not a falsifier of the theorem, and the compute caveat is honestly disclosed; the E3 numbers, however, cannot all be true as printed. Code availability is claimed, so the check is feasible. If the code reproduces the +0.027 with N≈27k and 60/130 discordants, then the gain calculation must be wrong; if the code uses a smaller N (e.g., a paired subset), the text must say so. Either way the current text is not self-consistent, so I would not move the verdict to ACCEPT. REJECT is too strong because the theoretical contribution and the qualitative frontier (monotone up to a knee, then bend-back) are supported by other runs (E4, E5, E6) and by the proofs. Verdict remains CONDITIONAL (UNCHANGED) pending this check.","tokens_in":22238,"tokens_out":7692,"duration_ms":79046,"concrete_test":"Run the released E3 script on the matched cell and print N, the two discordant counts, McNemar statistic, and the paired bootstrap CI. Compute (b−c)/N and compare with the reported '+0.027' and Table 2's 0.75→0.78. If N≈27k and discordants are 60 vs 130, the gain is ~0.003 and the headline frontier point is wrong; if the gain is truly +0.027, the discordant counts or N must differ. This one check settles whether the 15%-deferral result is reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The anytime-validity theorem (Prop. 3) is proved correctly; monotonicity does collapse the union. The load-bearing problem is in the paired result that anchors the paper's headline frontier. In E3, on the matched ~27k-clip cell, the paper reports that deferring 15% lifts decision accuracy from 0.75 to 0.78, with paired gain +0.027 (CI [+0.016,+0.038], McNemar p=4×10^-7, '60 vs 130 flips'). For a paired binary outcome, the gain equals (b−c)/N with b,c the discordant counts. The quoted counts differ by 70, so on N≈27k the gain is 70/27000≈0.0026, an order of magnitude below the reported +0.027. A +0.027 gain requires nearly 730 net flips. Table 2's decision-accuracy rows (0.75 [0.73,0.77] alone vs 0.78 [0.76,0.79] at 15% deferral) repeat the same tension. The released code must resolve which quantity is mis-stated. If the true paired gain is ~0.003, the 15%-deferral operating point (≈0.753) no longer exceeds the pixel-everywhere baseline (0.76), so the claim 'above pixel-everywhere at 7× less compute' fails. This is not a challenge to the theoretical construction, but it is the central measured evidence for 'escalate rarely', and the paper should not be accepted until the discrepancy is resolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper recasts AI-generated video detection as a streaming perception problem. It scores a codec motion-vector field already present in the compressed bitstream, using a CPU parse of ~10^5 MACs per GOP, and aggregates per-GOP scores with a running maximum. A single end-calibrated threshold decides “generated” when the running maximum crosses it; clips ending below a deferral band are escalated to a pixel or VLM stage. The main formal contributions are Proposition 2 (monotonicity of the running maximum), Proposition 3 (one end-calibrated threshold controls the false-positive rate at the data-dependent stopping time under the calibration null, while per-prefix recalibration does not), and Proposition 4 / Corollary 5 (closed-form expected compute and cascade accuracy, with a frontier monotone exactly where the marginal deferral condition holds). Experiments on a matched ~27k-clip GenVidBench cell report codec AUC@N 0.64, sAUC(1) 0.80, and deferring 15% of clips lifting decision accuracy from 0.75 to 0.78 at ~2.6e9 expected MACs, claimed as 7x below pixel-everywhere; cross-dataset replication on AIGVDBench gives AUC 0.62. The paper is careful to state its limitations: Proposition 3 is valid under the calibration null, Proposition 4 is conditional on the deferral condition being tested, and motion-bias controls are reported.","tokens_in":22457,"tokens_out":5144,"duration_ms":54150,"significance":"The core theoretical observation is correct and clean: for the monotone running maximum, the union over prefixes collapses to the final-prefix event, so a threshold calibrated once at the end controls the false-positive rate at the stopping time. I verified the proof of Proposition 3; it is elementary but non-obvious and directly useful. The empirical apparatus is unusually careful for this area: an identity floor, leave-one-generator-out evaluation, per-generator spread, motion-bias control, re-run external baselines (ReStraV, VideoVeritas), and a released harness that asserts Propositions 2–4 and Corollary 5 on synthetic data. If the measured frontier survives verification, this is a valuable contribution: a cheap always-on front-end with honest guarantees and a priced cascade. However, the practical impact rests on two load-bearing empirical claims that need scrutiny: the paired accuracy gain at the 15% deferral point, and the wall-clock interpretation of the compute savings. The manuscript itself repeatedly and honestly flags the conditional nature of the guarantees, which is a strength, not a defect.","major_comments":[{"comment":"The reported paired gain of +0.027 with “60 vs 130 flips” on ~27k matched clips is arithmetically inconsistent. For a paired binary outcome, the accuracy gain equals (b−c)/N where b and c are the discordant counts; (130−60)/27000 ≈ 0.0026, an order of magnitude below the reported +0.027. A +0.027 gain requires roughly 730 net flips. This is not a rounding issue: it is the central evidence for the headline operating point of 0.78 versus pixel-everywhere 0.76. If the true paired gain is ~0.003, the 15%-deferral decision accuracy is ~0.753, below the pixel baseline, and the claim “above pixel-everywhere at 7× less compute” fails. The released code must resolve whether the flip counts, the sample size, or the gain is mis-stated. I could not resolve this from the manuscript text.","section":"E3 / Table 2 / Figure 3"},{"comment":"The headline “5 orders of magnitude less compute” and “7× less compute” are based on MAC counts for a purpose-built bitstream parser, but the released reference implementation reads motion vectors through the decoder’s side-data export and pays a full frame decode (~0.4 s per GOP), compared with ~28 ms for the pixel forward pass. As the paper itself admits, this largely erases the wall-clock advantage. The MAC accounting is defensible for a deployed platform that already decodes the stream, but the abstract and Figure 3 present the MAC-based number as the operating-point benefit. Please either provide a measured parser benchmark or qualify the headline to “added compute” with the decode treated as a shared cost, so readers can see the wall-clock trade-off at the claimed operating point.","section":"§4 Cost and §J Compute accounting"}],"minor_comments":[{"comment":"The caption states that Figure 2 endpoints (0.674/0.782) “sit consistently above” Table 2’s AUC@N values (0.64/0.76). Please clarify explicitly why this happens (e.g., different prefix truncation or population definition) so the reader does not perceive an inconsistency.","section":"Figure 2 / Table 2"},{"comment":"The McNemar p-value of 4×10−7 is consistent with 60 vs 130 discordant counts, but not with the reported gain of +0.027. After the code is checked, please ensure the reported CI, p-value, flip counts, and gain all refer to the same paired analysis and same N.","section":"E3"},{"comment":"The statement “the released harness reproduces every number in this paper except the VLM MAC estimate” is strong; please add a commit hash or version identifier and, after resolving the E3 discrepancy, confirm that the reproduction statement still holds.","section":"Data and Code Availability"},{"comment":"The limitation that Proposition 3 holds only under the calibration null is stated clearly, and the cross-dataset FPR drift to 0.068 is reported. Consider carrying this qualifier into the abstract, since the phrase “anytime-valid” could otherwise be read as unconditional.","section":"§7 Limitations"}],"recommendation":"major_revision","confidential_remarks":"The theoretical construction is sound and I would like to see the paper published after the paired-gain discrepancy is resolved. The arithmetic inconsistency in E3 is exactly the kind of issue that can be a typo or a deeper error; the released code must settle it. If the true paired gain is ~0.003, the central empirical claim collapses and the paper would need substantial re-framing. I am therefore recommending major revision rather than rejection because the formal results and most of the experimental apparatus are solid, but I cannot accept the manuscript in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The theoretical core is correct and the empirical reporting is unusually honest; but the paper's headline paired result in E3 is internally inconsistent — the quoted flip counts cannot produce the quoted gain — and the released code has to settle which quantity is mis-stated before this should be accepted.\n\nWhat is actually new: recasting AIGV detection as streaming perception with prefix-indexed metrics (Definition 1: anytime AUC, latency-budgeted sAUC, decision-latency distributions) is a real reframing, and Proposition 3's point — one end-calibrated threshold on a monotone aggregate is anytime-valid, per-prefix recalibration is not — is correct and useful. I checked Propositions 2–4 and Corollary 5; the proofs are elementary and right. The paper is also honest where it matters: it labels the calibration null, tests the deferral condition rather than assuming it, reports the identity floor, the wide per-generator spread (AUC 0.45–0.90), and a motion-bias control against the known \"generated moves less\" shortcut. It re-runs ReStraV and VideoVeritas under its own protocol instead of quoting their papers, and the code/harness are claimed released. That is real evidence.\n\nNow the soft spots, in order of size.\n\nFirst, E3 arithmetic. On the matched ~27k-clip cell the paper reports 15% deferral lifting accuracy 0.75→0.78, paired gain +0.027 (McNemar p=4e-7, \"60 vs 130 flips\"). For a paired binary outcome the gain equals (b−c)/N with b,c the discordant counts; 70 net flips on ~27k clips is ~+0.003, an order of magnitude below +0.027, which would need ~730 net flips. This is not a quibble: if the real gain is ~0.003, the 15% operating point (~0.753) no longer clears pixel-everywhere (0.76), and the abstract's \"above pixel-everywhere at 7× less compute\" collapses. The stress-test note lands, and the released code must resolve it.\n\nSecond, the anytime-validity guarantee is a theorem about a fixed null. The paper's own cross-dataset run shows the stopping-time FPR drifting to 0.068 at target 0.05. The abstract's unqualified \"anytime-valid\" overstates; the limitation section discloses the shift, so this is a framing fix.\n\nThird, the 5-orders-of-magnitude compute claim counts the parse alone. The reference implementation exports motion vectors through the decoder and pays a full frame decode (~0.4 s per GOP, Sec. J), which largely erases the wall-clock advantage over the pixel stage's 28 ms. Disclosed, but the headline number is parse-only.\n\nNone of this sinks the central contribution — the reframing, the monotone-aggregate guarantee, and the measured frontiers are worth serious referee time. The paper is for people building efficient or streaming AIGV detectors and anyone applying anytime-valid inference. Send it to review, and make the E3 verification a condition of acceptance.","headline":"The streaming reframing and the anytime-validity trick are real and correct; the headline paired gain in E3 is arithmetically inconsistent with its own flip counts and must be resolved from the released code before acceptance.","tokens_in":23173,"tokens_out":4314,"would_cite":true,"duration_ms":38483,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that AI-generated video can be detected as it streams by reading the motion field the codec already wrote, and proves that one end-calibrated threshold—not per-prefix recalibration—controls false alarms at the data-depende","keywords":["AI-generated video detection","compressed-domain analysis","motion vectors","streaming perception","anytime-valid inference","cascade deferral","running maximum","false-positive control"],"falsifier":"Take a held-out collection of real videos whose codec and motion statistics differ from the calibration set, run the gate with the end-calibrated threshold, and count the fraction of real clips that trigger 'generated' at their stopping time; a rate clearly above the declared alpha—the paper already reports 0.068 at alpha 0.05—would falsify the practical claim, even though Proposition 3 itself remains true under the calibrated null.","tokens_in":21926,"feed_emoji":"📹","tokens_out":6462,"duration_ms":57968,"temperature":0.7,"pith_summary":"The paper recasts AI-generated video detection as a streaming decision rather than a one-shot offline scoring problem. It argues that the motion field a video codec writes into the compressed bitstream is enough to drive an always-on cheap detector, and that because the detector's running score is a monotone maximum, a single threshold calibrated once on real clips controls the false-positive rate at whatever time the detector decides to stop—while re-calibrating the threshold separately at each prefix breaks that control. It then prices a cascade: clips whose final score lands in an uncertainty band are deferred to a more expensive pixel or vision-language model, with expected compute and accuracy given in closed form. Measured on a matched public benchmark, the cheap stage reaches an AUC of 0.64 at five orders of magnitude less compute than a pixel CNN, and deferring 15% of clips lifts accuracy from 0.75 to 0.78 at seven times less compute. If correct, detection cost becomes a codec parse instead of a large-model forward pass.","feed_headline":"Detect AI video from codec motion—no pixel decode","feed_subtitle":"One monotone score keeps false alarms in check at any stopping time; 15% deferral lifts accuracy to 0.78 at 7x less compute.","key_machinery":"The running maximum M_t = max_{i <= t} s_i of per-chunk scores, whose monotonicity collapses the union of crossing events {exists t: M_t >= tau} into a single final event {M_N >= tau}; plus the deferral band W = [tau - w, tau) and the closed-form cascade accounting of Proposition 4, which expresses expected compute and expected error as linear functions of the deferral probability and the deferral gain.","core_discovery":"The paper's central claim is Proposition 3: with M_t the running maximum of per-GOP scores, if a single threshold tau is calibrated so that P0(M_N >= tau) <= alpha under the real-clip distribution, then the alarm time sigma = inf{t <= N : M_t >= tau} satisfies P0(sigma <= N) = P0(M_N >= tau) <= alpha, and P0(M_t >= tau) <= alpha for every prefix t. Monotonicity collapses the union of crossing events into the single final event. The paper also shows the converse construction fails: choosing thresholds tau_t with P0(M_t >= tau_t) = alpha at each prefix gives a union probability at least alpha, with strict excess whenever an early crossing missed by the final test has positive probability. A co","pith_inferences":["The same monotone-score-plus-end-threshold construction should transfer to other streaming binary detection tasks (deepfake audio, fraud, content moderation), since the proof uses only monotonicity, not video-specific structure.","The five-orders-of-magnitude compute claim counts only the parse; a purpose-built bitstream parser is needed to turn that into wall-clock savings, because the reference implementation pays a frame decode to export motion vectors.","Calibration transfer, not the theorem, is the operational risk: re-calibrating on diverse real video samples, or adding a second monotone feature, could keep the stopping-time false-positive rate near target under distribution shift.","The paper's own ablation suggests a monotone transform of a cumulative average could recover some of the accuracy gap while preserving the anytime guarantee—an extension worth testing, since the running max is the weakest offline aggregate measured."],"forward_implications":["An always-on detector can run on CPU inside the streaming path, adding roughly 10^5 MACs per chunk instead of a pixel-domain forward pass.","A single end-calibrated threshold gives a legitimate anytime-valid guarantee; re-calibrating at each prefix is not just redundant but inflates the stopping-time false-positive rate (measured 1.7–2.8x under distribution shift).","A compute budget maps directly to a deferral width, and the measured compute-accuracy frontier is monotone only where stage 2 beats stage 1 on the deferred band; beyond that knee the frontier bends back.","The result is not tied to the specific motion features: any monotone stage-1 score inherits both guarantees, so the same construction can carry a stronger cheap detector.","When real and generated clips are matched in length, the anytime curve is flat—the first GOP is as informative as the full clip—so streaming latency buys little there."],"fun_headline_variants":["Anytime AI-video detection from bitstream motion—no pixel decode","Streaming AI video detection: one threshold for any stop time","AI video detection on the fly: read the motion field, skip decoding","One monotone score: any time AI video detection, no decode"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The practical guarantee holds only while deployed real clips have the same running-maximum distribution as the calibration sample; under the paper's own cross-dataset shift the stopping-time false-positive rate drifts from about 0.04 to 0.068 at a target of 0.05.","fun_headline_variants_meta":{"raw":{"variants":["Anytime AI-video detection from bitstream motion—no pixel decode","Streaming AI video detection: one threshold for any stop time","AI video detection on the fly: read the motion field, skip decoding","One monotone score: any time AI video detection, no decode"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001351,"raw_usage":{"total_tokens":5374,"prompt_tokens":844,"completion_tokens":4530,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":4453}},"tokens_in":588,"tokens_out":4530,"duration_ms":27550,"temperature":1.0,"reasoning_tokens":4453,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:40:23.134964+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out collection of real videos whose codec and motion statistics differ from the calibration set, run the gate with the end-calibrated threshold, and count the fraction of real clips that trigger 'generated' at their stopping time; a rate clearly above the declared alpha—the paper already reports 0.068 at alpha 0.05—would falsify the practical claim, even though Proposition 3 itself remains true under the calibrated null.","supporting_citations":[],"review_version":1}