{"id":"53509b58-3dc6-495c-9090-8f1064727247","arxiv_id":"2609.36549","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A frozen medical deepfake detector improved its own accuracy by 9 to 11 points over 20 rounds by writing experience entries and image-analysis tools that were admitted only after passing acceptance tests on independent labeled validation images.","lead":"This paper wraps a frozen medical deepfake detector in a self-improvement loop that lets it write its own memory entries and image-analysis tools, admitting every change only after it passes a test on labeled validation images. Across 49 configurations the loop raised average test accuracy from 75.0% to 84.4% without retraining, worth reading for its controlled study of autonomous adaptation and its careful catalog of failures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"No operating-point-matched frozen baseline: the headline gains equal the real/fake recall mirror tradeoff of the frozen detector, so they may be pure threshold rebalancing rather than recursive self-improvement.","rationale":"The reader's weakest assumption (deployment label source) is real but acknowledged in Section 7 (b159): the loop requires a verified label stream, and the paper explicitly conditions the deployment claim on it. I do not treat that as the decisive gap. The threshold-matched baseline is more load-bearing because it is unaddressed and directly bears on the interpretation of every headline number. The recall pairs in Section 5.1 are consistent with moving along the frozen detector's existing ROC curve: each final-evaluation gain equals half the real-recall increase plus half the fake-recall decrease, and the paper itself attributes the improvement partly to threshold miscalibration. A threshold-matched frozen control would settle whether the RSI machinery is responsible or whether a fixed operating-point change suffices. This does not overturn the paper's empirical honesty or the value of its failure taxonomy, but it is a genuine condition on the central 'self-improvement' claim, so the reader's CONDITIONAL verdict stands.","tokens_in":30411,"tokens_out":8790,"duration_ms":106473,"concrete_test":"On the released 4,000-image final evaluation, evaluate the frozen MedForge-Reasoner checkpoint across a sweep of operating points without any learning, memory, or tools: vary a confidence threshold on the parsed verdict (or token-level probability on the conclusion line; if the API exposes no score, vary a fixed one-sentence directive such as 'when uncertain, clear real images'). Plot accuracy and real/fake recall per threshold. If any threshold point reaches 87.9% clean and 70.5% distorted, or the 84.4% four-set average, the RSI gains are consistent with pure operating-point rebalancing. If the best threshold point falls materially below those numbers while holding both sets, the self-improvement claim survives. This requires no retraining and uses only the released harness and scorer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is that no frozen baseline is evaluated at a matched operating point. The base detector's error structure is extremely one-sided: main-set real recall is 60% versus fake recall 99%; on the clean final evaluation real recall is 53.8%, which with the balanced 2,000-image design implies fake recall of about 99.2% at 76.5% accuracy. The headline improvements are exactly the mirror-image tradeoff along that same axis. On the clean final evaluation, ensemble reflection reaches 89.3% real recall and 86.4% fake recall, that is, 87.9% accuracy; on the distorted set it moves 22.1% to 58.8% real and 95.9% to 82.2% fake, that is, 59.0% to 70.5% accuracy. In both cases the accuracy gain equals the average of the real-recall gain and the fake-recall loss, the signature of an operating-point shift. The paper's own Section 7 explanation says the gains reflect missing evidence and miscalibrated thresholds rather than missing capability, yet no configuration reports the frozen detector at a similarly balanced operating point, for example via a confidence threshold or a fixed one-line directive. Without that control, the observed +9.4-point four-set average and +11.4/+11.5-point final-evaluation gains do not establish that reflection, memory, self-written tools, or acceptance testing add capability; they may be a threshold adjustment the frozen model could have made with no self-improvement loop at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents MedForge-RSI, a recursive self-improvement framework that adapts a frozen 8B medical deepfake detector (MedForge-Reasoner) at deployment time by letting the detector itself propose experience-memory entries and self-written image-analysis tools, with every persistent change admitted only after a replay check plus an A/B comparison on an independent 100-image validation pool. Over 20 rounds and 49 registered, pre-registered configurations, the authors report raising the four-test-set average from 75.0% to 84.4% and, on a 4,000-image final evaluation, clean accuracy from 76.5% to 87.9% and transmission-distorted accuracy from 59.0% to 70.5%, with gains concentrated in real-image recall. The paper additionally contributes a tool-content audit (61% of self-written implementations emit unparseable readings, 26% embed hardcoded numbers), a seed-replication study showing that only the ensemble-reflection configuration replicates (mean 84.0 vs. 79.5 control on the main set) while the single best run (87.0) does not, and a four-mode failure taxonomy with two documented verification-failure incidents. All trajectories, including rejected and rolled-back changes, are released.","tokens_in":30614,"tokens_out":9573,"duration_ms":57156,"significance":"If the central claim held, this would be a notable result: a safety-critical perception system whose weights stay frozen yet which adapts to new generators and transmission distortions through verified artifact-level self-modification, with full trajectory disclosure and an unusually honest seed-replication analysis. The paper's strengths are real and worth naming: 49 registered single-switch configurations, pre-registration of batches, machine-checked disjointness of the validation pool from the 4,000-image final evaluation, complete release of trajectories, and a self-critical mechanism analysis (several intuitive mechanisms measurably hurt, and the best single run fails to replicate). However, the headline improvements are exactly the average of the real-recall gains and fake-recall losses of the frozen detector, which is the signature of a pure operating-point shift, and no frozen baseline at a matched operating point is reported. The central claim that the loop, memory, tools, and acceptance testing add capability beyond a threshold adjustment is therefore not yet established.","major_comments":[{"comment":"No operating-point-matched frozen baseline is reported, and this is load-bearing for the central claim. The base detector's error structure is one-sided (main-set real recall 60%, fake recall 99%; clean final evaluation 53.8% real, ~99.2% fake). Every reported gain lies exactly on the balanced-accuracy line: on the clean final evaluation, ensemble reflection moves real/fake recall from 53.8/99.2 to 89.3/86.4, and (89.3+86.4)/2 = 87.9 reproduces the reported accuracy; on the distorted final evaluation, (58.8+82.2)/2 = 70.5 reproduces the reported accuracy; the main-set control and ensemble points satisfy the same identity. That is, the +11.4/+11.5 point final-evaluation gains equal the average of the real-recall gain and the fake-recall loss, which is the signature of an operating-point rotation, not of added discriminative capability. The paper's own Figure 5 shows every learning configuration sliding along the real/fake recall tradeoff, and Section 7 (b154) concedes that the improved components reflect 'miscalibrated thresholds rather than missing capability.' No configuration evaluates the frozen detector at a matched operating point, e.g., a confidence threshold or a one-line leniency directive tuned on the same 100-image validation pool used by the acceptance test (§3.4, Appendix E). The tool ablation of §6.1 (Table 12) shows tools matter at the RSI-selected operating point, but it does not show the RSI outcome lies outside the frozen detector's achievable operating-point set. I recommend adding a matched-operating-point frozen baseline (threshold or fixed directive tuned on the validation pool) reported on all four sets and the final evaluation, with the real/fake recall pair, and an ROC-style comparison showing whether the RSI final state is outside the frozen detector's achievable tradeoff. Without this control, the claim of b163 that the loop adds capability is not distinguished from a threshold adjustment the frozen model could make with no self-improvement loop.","section":"§5.1, Table 1; §7 (b154); Figure 5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The headline magnitudes rest on single-run measurements even for the configuration that is claimed to be seed-replicated. Per Table 11, only ensemble reflection replicates, and the replication was measured on the main test set only (86.5 and 81.5; mean 84.0 versus 79.5 control). The final-evaluation figures that carry the abstract's and conclusion's main claims (clean 87.9, distorted 70.5) are from the base-seed run v12 and were not replicated; two-sided reflection (86.9/65.7) and case-based memory lack seed replication entirely (b160). Given the paper's own stated standard ('seed replication as a methodological requirement for self-improvement claims', §6.2, and condition 4 of b163), the +11-point final-evaluation and +9.4-point four-set numbers should be presented as single-run, seed-unconfirmed estimates, with the replicated estimate (roughly +4.5 points on the main set) given equal prominence in the abstract and conclusion. The acknowledgment in the Limitations section does not fully carry into how the headline numbers are framed.","section":"§6.2, Table 11; §4 (final evaluation)","authors_acknowledge":"The final evaluation is complete for six configurations; the self-reflection and case-based memory rows and the first two training-time rows are single-run or pending measurements and their final-evaluation cells are left blank accordingly. formance, as in the case of second-opinion arbi- tration, which reaches 87.5% at its best round and finishes at 78.5%.","authors_acknowledge_p":0.7},{"comment":"The deployment-assumption gap should be tightened where the quantitative claims depend on it. The loop requires 40 verified labels per round from the learning stream (drawn here from the detector's own training pool) plus a 100-image labeled validation pool drawn from the official held-out split. The paper acknowledges the label-source requirement (b159), so this is not presented as a new finding, but it does mean the demonstrated gains are conditional on a continuously labeled stream that a real deployment must supply externally; the abstract and conclusion state the four conditions of b163 without this fifth, arguably equally binding, condition.","section":"§4; Appendix E; §3.4","authors_acknowledge":"The learning stream draws labeled images from the training pool, and a de- ployment would require some verified label source, such as adjudicated clinical flags.","authors_acknowledge_p":0.72}],"minor_comments":[{"comment":"The abstract's 'falls to 59% on unseen generators and 59% under transmission distortion' mixes bases: 59.0% is the final-evaluation distorted base, while the four-test-set distorted set's base is 64.5%. Please state which base each claim refers to, or the numbers read as inconsistent with Table 1.","section":"Abstract vs. §4/Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The silent-verifier narrative ends mid-sentence: 'the score-drop check The main scoreboard compared fields explicitly...' appears to be a truncated sentence. Please complete it.","section":"Appendix Q","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The recommended four-mechanism combination (v48/v49) is marked 'run.' with no results anywhere in the paper; since the contribution list and Section 7 reference it, state explicitly that no results for it are available and that all headline figures are from ensemble reflection (v12) and two-sided reflection (v19).","section":"Table 2 (v48/v49) and §5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The text says ten of eleven configurations lose accuracy without their tools 'by 4.3 points on average'; the mean across the eleven rows of Table 12 is 4.27 including the zero row, which matches only if the zero row is included. Please state the computation so readers are not misled about the ten that lose.","section":"§6.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Given the operating-point issue, consider reporting a table of (real recall, fake recall, accuracy) for the frozen control under several thresholds alongside the RSI states in Table 1; this would make the rebalancing component of the gains directly visible and would strengthen the paper's otherwise excellent disclosure.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Read this one. It is the most careful empirical study of recursive self-improvement in a perception task I have seen, and it has one load-bearing gap. What is new: a frozen 8B medical deepfake detector gains roughly 9 to 11 accuracy points through two knowledge stores, experience memory and self-written tools, under independent-validation acceptance testing; there are 49 pre-registered single-switch configurations, a measured four-mode failure taxonomy, a negative seed-replication result, and a promised release of all trajectories. That is real work. The tool-calibration analysis is genuinely good: 61% of model-written tools emitted no parseable reading, 26% embedded hardcoded numbers unrelated to their computation, and the percentile-calibration layer plus tool ablations show the tools contribute about 4.3 points on average. Credit where due: the validation pool is disjoint from the test sets, the 4,000-image final evaluation is disjoint from training, circularity is low, and the authors disclose the resample contamination and the missing replications rather than hiding them.\n\nSoft spots, in order of importance. First, the operating-point concern lands. The frozen detector's errors are extremely one-sided: on the clean final evaluation real recall is 53.8% and fake recall is about 99.2% at 76.5% accuracy. Ensemble reflection moves real recall from 53.8% to 89.3% and fake recall from 99.2% to 86.4%, which is an accuracy gain of 11.4 points on a balanced set. The arithmetic is exactly the mirror tradeoff along the same axis, and the distorted set shows the same pattern. No configuration evaluates the frozen base at a matched operating point, via a confidence threshold or a fixed one-line directive. Without that control, the headline gain may be a threshold adjustment available with no self-improvement loop at all. The authors say in Section 7 that the gains reflect missing evidence and miscalibrated thresholds rather than missing capability, but they never isolate the threshold component. That is the central ambiguity. Second, seed replication covers only ensemble reflection; the best single-seed configuration does not replicate, two-sided reflection and case-based memory are single-seed, and the recommended four-mechanism combination was still running at submission. The authors state all of this, but by their own methodological standard only one configuration fully supports the claim. Third, learning batches draw from the training pool, so a deployment label source is assumed; they acknowledge this in Section 7.\n\nWho this is for: anyone working on deployment-time adaptation, verifier design, or evaluation standards for self-improvement claims. The failure taxonomy, operating-point reversal, silent verifier failure, and single-run variability, is a contribution on its own. Recommendation: it deserves a serious referee. The referee should ask for a matched-operating-point frozen baseline and completed seed replications for the principal configurations. With those, the claim becomes either much weaker or much stronger; right now it is genuinely ambiguous.","headline":"Strong empirical study of deployment-time self-improvement with a real missing control: no frozen baseline at a matched operating point, so the headline gains may be threshold rebalancing.","tokens_in":31376,"tokens_out":2968,"duration_ms":38889},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A medical deepfake detector whose weights are permanently frozen can improve itself at deployment time, raising average accuracy from 75.0% to 84.4% across four test sets without any gradient update.","keywords":["recursive self-improvement","medical deepfake detection","frozen weights","acceptance testing","tool calibration","vision-language model","deployment shift","false-alarm reduction"],"falsifier":"Run the loop on fresh learning streams whose labels carry realistic clinical noise (for example, adjudication disagreement rates) and whose validation split is drawn from a genuinely disjoint deployment pool rather than the official held-out test split; if the roughly ten-point gains collapse, the deployment claim fails. The paper itself offers a partial check: calibration+freeze, its best single run at 87.0, averaged 79.3 across three seeds and was statistically indistinguishable from the no-learning control, so a direct falsifier is to replicate the headline ensemble and recommended configurations across more than two seeds and confirm that the 4,000-image final-evaluation gain persists.","tokens_in":31625,"feed_emoji":"🩺","tokens_out":7257,"duration_ms":39372,"temperature":0.7,"pith_summary":"This paper claims that a medical deepfake detector with permanently frozen weights can meaningfully improve itself at deployment time, with no gradient updates and no additional training compute. Over 20 rounds, the detector examines its own verified errors, accumulates structured experience entries, and writes and repairs its own image-analysis tools, while an independent acceptance test admits only changes that pass replay and A/B validation checks. The claim is that this loop raises average accuracy over four test sets from 75.0% to 84.4%, lifts real-image recall from 60% to 80% (91% with case-based memory), and raises clean and transmission-distorted accuracy on a 4,000-image held-out evaluation from 76.5% to 87.9% and from 59.0% to 70.5%. The authors argue the gains concentrate on the error type that reinforcement learning left unfixed: the detector falsely flagged 40% of authentic scans, and this false-alarm rate drops to 20%, or to 10% with case-based memory.","feed_headline":"Frozen AI detector gains 9 accuracy points, no retraining","feed_subtitle":"A self-improvement loop with acceptance tests lifts a deepfake detector from 75.0 to 84.4 percent, weights frozen.","key_machinery":"The machinery is a two-store adaptation loop wrapped in a safety architecture. The experience memory holds structured entries of the form (context, cue, caution), recording the imaging setting in which a lesson applies, the visual cue to seek, and the error it guards against, while the toolbox holds image-analysis tools written by the detector itself under a strict contract: a single self-contained file using only Pillow and NumPy, deterministic numeric output, and no verdict labels. Because self-written tools are unreliable instruments—61% emit output with no parseable reading and 26% embed hardcoded numbers unrelated to the computation—a calibration layer converts each raw reading into a percentile within historical fake- and real-image reading distributions, rendering directional evidence or “NO SIGNAL” when the distributions overlap. Every proposal to either store must pass a two-stage acceptance test: replay on the learning batch requiring non-negative net correct change, then an A/B comparison on a disjoint validation split allowing at most one correct-count drop, with a verifier health check that distrusts test outcomes when the replay baseline fails to reproduce 70% of known-correct predictions. The loop's write access excludes the model weights, the scoring code, the official system prompt, and all evaluation sets, which is what makes the improvement bounded self-refinement rather than unchecked self-modification.","core_discovery":"On its own terms, the central discovery is that recursive self-improvement works in a safety-critical perception setting when every persistent change is gated by an acceptance test on an independent validation split and when tool readings are calibrated against labeled history rather than trusted as reported. The frozen detector MedForge-Reasoner, an 8B vision–language model that reaches 99.2% accuracy in its training distribution but only 75.0% under deployment conditions, improves to 84.4% over four test sets, and on a 4,000-image final evaluation clean accuracy rises from 76.5% to 87.9% and distortion-affected accuracy from 59.0% to 70.5%. The decisive result is attribution: controlled single-switch comparisons across 49 registered configurations identify acceptance testing as the largest single contributor (a silent scoring bug that disabled it cost 17.5 points of final accuracy), tool-reading calibration as the most reliable positive mechanism (+6.5 points), and seed replication as the criterion that separates genuine mechanisms (ensemble reflection, +4.5 points over control) from single-stream luck (calibration+freeze, whose best run of 87.0 fails to replicate).","pith_inferences":["If the result holds, the binding resource for perception self-improvement is verified label supply rather than model capability, which suggests the same loop could transfer to other imaging domains where ground-truth labels are cheap or automatic, such as satellite or industrial inspection.","The finding that acceptance testing matters more than reflection quality, and that proposal-diversity mechanisms replicate while verdict-formation mechanisms (arbitration, self-consistency voting) hurt, predicts a testable ranking for other frozen-model self-improving systems: admission gating and evidence calibration should beat proposal cleverness.","Both a silent verifier bug and an operating-point reversal went unnoticed until hand inspection, which implies that a real clinical deployment needs externally held health checks on the acceptance machinery itself; the paper's released trajectories, including rejected and rolled-back changes, are the natural audit substrate."],"forward_implications":["A deployed detector can track newly released generators and transmission distortions without retraining; each admitted change costs roughly 140 labeled validation images and the entire study uses fewer labels than a single SFT run of the base detector.","The false-alarm rate on authentic scans drops from 40% to 20% (10% with case-based memory), attacking the failure mode most damaging to clinical trust rather than the already-saturated fake-image side.","The gain generalizes beyond the learning stream: unseen-generator accuracy rises from 77.0% to 88.0–89.5%, and the improvement is monotonic across disjoint test sets, indicating the acquired capability is not memorization of the learning stream.","The division of labor persists: gradient updates set the base perceptual ability, verified self-improvement adapts it, and periodic lightweight retraining on the accumulated experience and tools is left as future work.","Six design guidelines transfer: test before admitting, calibrate tool readings against labeled history, write access only on enumerable artifacts, diversify proposals, protect recovery rather than only stability, and replicate across seeds."],"supporting_citations":[{"why":"Supplies the frozen base detector MedForge-Reasoner, the 90K benchmark with its 5:1:3 train/RL/test split, and the 99.2% in-distribution baseline that the self-improvement loop operates on.","marker":"(Chen et al., 2026c)"},{"why":"Provides the harness formalism (memory, tools, orchestration code), the deployment-time-by-verification taxonomy in which the paper situates itself, and the named failure modes the study instantiates.","marker":"(Chen et al., 2026a)"},{"why":"The closest comparison, improving the harness of a frozen 8B vision–language model for long-video understanding; its stated limitation of per-video rather than persistent memory is the gap the two-store design fills.","marker":"(Xu and Chen, 2026)"},{"why":"The bounded, human-auditable evolution design that the protected-core architecture and enumerable rule lists follow.","marker":"(Chen et al., 2026b)"},{"why":"Origin of persistent skill libraries from which the tool-creation channel descends.","marker":"(Wang et al., 2024)"},{"why":"Documents model collapse in recursively generated data, one of the failure modes the acceptance test and retirement rules are designed to avoid.","marker":"(Shumailov et al., 2024)"},{"why":"The protected-core design principle of keeping weights and scoring fixed while evolving the surrounding harness.","marker":"(Shi et al., 2026)"},{"why":"The experience-memory lineage (ExpeL) extended here with per-entry acceptance testing and usage-statistics-based retirement.","marker":"(Zhao et al., 2024)"}],"fun_headline_variants":["Frozen detector self-improves from 75% to 84.4% accuracy","Self-improving AI detector gains 9.4 points, weights untouched","Recursive self-improvement lifts frozen medical deepfake detector","Tool calibration and acceptance tests boost frozen AI detector"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that deployment supplies a continuous stream of correctly labeled images (40 per round) plus a labeled validation split whose distribution matches the evaluation target; the paper acknowledges in Section 7 that its learning stream draws from the training pool and that a real deployment would need a verified label source such as adjudicated clinical flags, and without such labels the loop cannot run and its gains are tuned toward a validation proxy.","fun_headline_variants_meta":{"raw":{"variants":["Frozen detector self-improves from 75% to 84.4% accuracy","Self-improving AI detector gains 9.4 points, weights untouched","Recursive self-improvement lifts frozen medical deepfake detector","Tool calibration and acceptance tests boost frozen AI detector"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"cost":0.000168028,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000168028,"upstream_inference_prompt_cost":0.000124068,"upstream_inference_completions_cost":4.396e-05},"total_tokens":1248,"prompt_tokens":1091,"completion_tokens":157,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":79}},"tokens_in":1091,"tokens_out":157,"cost_source":"openrouter_usage","duration_ms":2373,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":79,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-01T00:30:17.875625+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run the loop on fresh learning streams whose labels carry realistic clinical noise (for example, adjudication disagreement rates) and whose validation split is drawn from a genuinely disjoint deployment pool rather than the official held-out test split; if the roughly ten-point gains collapse, the deployment claim fails. The paper itself offers a partial check: calibration+freeze, its best single run at 87.0, averaged 79.3 across three seeds and was statistically indistinguishable from the no-learning control, so a direct falsifier is to replicate the headline ensemble and recommended configurations across more than two seeds and confirm that the 4,000-image final-evaluation gain persists.","supporting_citations":[{"cited_title":"M ed F orge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning","cited_arxiv_id":null,"evidence_quote":"The closest comparison, improving the harness of a frozen 8B vision–language model for long-video understanding; its stated limitation of per-video rather than persistent memory is the gap the two-store design fills."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}