{"id":"aac6eb82-6068-45dd-8125-81d193c73e71","arxiv_id":"2607.09800","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A matched high-precision reference trace forecasts invisible weight-update events in controlled low-precision runs, and a pre-frozen 10% fp32-master mask recovers ~61% of the RNE-vs-fp32 loss gap on three unseen targets.","lead":"Low-precision training can silently discard small optimizer updates, and this paper tests whether a high-precision \"reference trace\" can forecast those invisible updates and choose which weights deserve fp32 backup. In controlled runs the forecast is accurate and a pre-frozen 10% mask recovers ~61% of the low-precision loss gap, but the audit does not transfer to a modern decoder and the allocator has no net systems benefit.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The audit's counterfactual proxy (Eq. 4) fails the paper's own locked 64-cell transfer, so the 'empirically useful aggregate audit' claim rests only on toy regimes; the paper acknowledges this, so the reader's CONDITIONAL verdict should stand unchanged.","rationale":"The reader's weakest-assumption identification matches my own: the counterfactual reliability of the high-precision reference path is the load-bearing assumption for the audit contribution. The paper reports its own locked transfer where this assumption fails, and the modern-decoder bundle is uninformative, so the positive audit claim is confined to controlled toy systems. This is a genuine limitation but the paper explicitly scopes the claim to 'controlled regimes studied here,' so the internal argument remains coherent. No code or data is public, which exacerbates the difficulty of independently verifying even the toy-regime results, but the quantitative evidence presented for those systems is detailed and internally consistent. The prospective allocation result is separate and does not depend on the failed proxy; it is also transparently limited to fixed targets and no population inference. Thus the reader's CONDITIONAL verdict (rather than ACCEPT or REJECT) is appropriate, and my stress-test does not move it.","tokens_in":31918,"tokens_out":11974,"duration_ms":131932,"concrete_test":"Pre-register and run the projected-reference audit (Eq. 4) on a held-out realistic setting—e.g., a 350M-parameter decoder with a different optimizer or schedule—using the same locked gates and thresholds from Appendix B.1. If the same-code premise again collapses below ~1% and the informativeness/directional-added-value gates fail (as in the modern-decoder bundle), the audit claim is confirmed to be toy-regime-only; if the gates pass on this new regime, the counterfactual concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The projected-reference proxy in Eq. (4) assumes the high-precision path remains a useful counterfactual near the visibility boundary. The paper's own locked 64-cell transfer (Appendix B.1) shows this assumption fails dramatically: measured and projected current-code vectors first differ at update 3, the same-code premise holds in only 0.797% of 762,880,000 coordinate-events, and every predeclared uncertainty criterion fails (AUROC 0.722 vs 0.656 comparator, risk 0.105 vs 0.059 required, frozen-fraction interval containment 0.368, onset interval 33/64 cells). The separate modern-decoder bundle likewise finds the visibility trajectories uninformative: range <0.10 and all directional-added-value gates fail. Therefore the central audit claim—that reference-derived proxies give an empirically useful aggregate audit—is supported only in the two-layer GD and small char-GPT controlled testbeds, with no demonstrated path to a realistic training regime. The exact local event (Eq. 1) remains definitionally correct, and the allocation result (which uses absolute proposal contribution, not the proxy) is not directly affected. But the headline 'reference traces can audit aggregate visibility' is materially weakened: outside toy systems, the aggregate audit has no positive evidence and the paper's own locked test provides strong negative evidence. This is not an internal inconsistency—the paper states the limitation—but it means the audit's operational value is unestablished for any practical setting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies two uses of high-precision reference traces for low-precision training: (i) auditing when stored-weight updates are invisible under round-to-nearest, and (ii) selecting an exact-budget fp32-master tile mask before unseen target runs. The exact visibility event is Eq. (1), Q_m(q_i+d_i)=q_i, with the interval equivalence in Eq. (2). The aggregate audit uses a projected-reference proxy, Eq. (4), which is explicitly not a pathwise certificate and is bounded only by a posterior diagnostic, Eq. (8). Controlled two-layer and char-GPT experiments show good forecast accuracy for first-majority-crossing times, while a locked 64-cell transfer and a modern-decoder bundle fail all predeclared uncertainty/informativeness gates. The write-back policy experiment reports that switching from RNE to stochastic rounding recovers 89.2–92.2% of the RNE-vs-fp32 loss gap after 10,000 GPT-2 steps. The prospective allocation study freezes a source-trace absolute-proposal-contribution mask before three unseen targets and reports mean recovery 0.607, close to target-specific masks (0.608) and well above fixed composition-matched controls (0.179). A systems check shows reduced PyTorch peak allocation versus a blanket master but net overhead and lower throughput versus direct RNE, with no validated break-even. The paper is unusually careful in stating its limitations, including the failures of the aggregate audit and the absence of a systems benefit.","tokens_in":32237,"tokens_out":5875,"duration_ms":63510,"significance":"If the results hold, the exact coordinate-level event in Eq. (1) is a clean, machine-checkable definition of update visibility, and the allocation protocol is a useful template for prospective, pre-registered evaluation of precision-allocation policies. The paper also reports negative results—the failure of the projected-reference proxy under code divergence, the failure of visibility-based allocation, and the lack of systems benefit—with unusual honesty and detail. However, the significance as a general method is limited: the aggregate audit has positive evidence only in small controlled testbeds and has direct negative evidence in the paper's own locked transfer test; the allocation result covers one architecture at two depths with three fixed target seeds; and the systems comparison shows no net benefit. The exact event and the write-back policy contrast are solid, but the broader 'reference traces for auditing' claim needs either substantial additional transfer evidence or a much narrower framing.","major_comments":[{"comment":"The projected-reference proxy is the load-bearing component of the aggregate-audit claim, yet the paper's own locked transfer test shows it fails outside the controlled testbeds: measured and projected current-code vectors first diverge at update 3; the same-code premise holds in only 0.797% of 762,880,000 coordinate-events; and every predeclared uncertainty gate fails (AUROC 0.722 vs 0.656 comparator, risk 0.105 vs 0.059 required, frozen-fraction interval containment 0.368, onset interval 33/64). The modern-decoder bundle also finds visibility trajectories uninformative (range <0.10 and all directional-added-value gates fail). I acknowledge that the paper states these limitations explicitly, but the conclusion that reference-derived proxies 'give an empirically useful aggregate audit' is supported only in two-layer and small char-GPT regimes. The exact local event Eq. (1) is unaffected,","section":"§2.2, Eq. (4), Appendix B.1"},{"comment":"The paper makes extensive empirical claims backed by an elaborate artifact ledger and many hashes, but the Data availability and Code availability sections state that no public archive or repository exists and the Zenodo DOI remains private. A reader cannot verify any headline number—recoveries, gate outcomes, systems timings, or the locked 64-cell failure attribution. For a cs.LG paper whose contribution is primarily empirical and methodological, public release of code, data, and at least the locked protocols and validation outputs should be a condition of publication. This is a load-bearing reproducibility issue, not a stylistic one.","section":"Data availability / Code availability"},{"comment":"The prospective allocation result is based on three fixed target seeds and five fixed composition-matched controls, and the paper correctly says no population inference is intended. However, the source-P mean recovery of 0.607 is only 0.00143 below the privileged target-specific P, and the post hoc ECO comparator recovers 0.965 (mean) with unmeasured systems cost. The claim that a source trace can 'guide exact-budget protection' is therefore demonstrated only as a fixed-design proof-of-concept against fixed controls, not as an effective or production-viable allocation rule. This limitation is present in the text, but the title and abstract give it less weight than the evidence supports. I would ask the authors to make this scoping prominent in the abstract and conclusion.","section":"§5, Tables 17–18"}],"minor_comments":[{"comment":"The phrase '384x in time' should use the multiplication sign (384×) for consistency with the rest of the text.","section":"Abstract"},{"comment":"The sentence 'No low-precision observation enters Eq. (4)' is misleading because the target grid m enters through Q_m; clarify that no low-precision trajectory observation enters.","section":"§2.2"},{"comment":"Figure 2(c) labels the E4M3 SR never-moved fraction as exactly '0' in one place and '5.9e-09' for E5M2; the text says 'at most 1.18e-8' and 'exactly zero.' Please harmonize the reporting of these near-zero values.","section":"§4.2 / Figure 2"},{"comment":"The caption says 'No confidence interval or hypothesis test is reported.' This is good, but the same statement should appear in the main text where the mean 0.607 is first introduced, since some readers will otherwise interpret the three fixed targets as a sample.","section":"§5 / Figure 3"},{"comment":"The reproducibility statement claims a retained provenance bundle, but the Data/Code availability sections say no public record exists. These statements should be reconciled, ideally by making the artifacts public.","section":"Reproducibility statement"},{"comment":"The term 'audit' is used both for the exact event Eq. (1) and for the aggregate proxy Eq. (4). Given that the proxy is not a certificate, consider using 'forecast' or 'aggregate proxy' consistently to avoid conflating the two.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"This is one of the most honest empirical papers I have reviewed: the negative results are reported with unusual care, and the exact event and prospective protocol are methodologically sound. My recommendation is driven by two issues: (1) the title/abstract claim is materially broader than the positive evidence, especially in light of the paper's own locked transfer failure; and (2) no public code or data is available despite an intricate artifact ledger. If the authors narrow the claims to the controlled regimes and the exact-event/policy-contrast results, and release the artifacts, I would be willing to accept a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that it does two different jobs and they are not equally strong. The exact coordinate-freeze event (Eq. 1 and its interval equivalent) is simple, correct, and useful: it gives a precise definition of when a stored low-precision code fails to record an optimizer proposal. That part is solid. The second job — using a high-precision reference trace to audit aggregate visibility — is scrupulously honest about its limits, and those limits are severe.\n\nWhat is genuinely new and well done: the matched-trajectory audit protocol, the controlled validation across a 384× range of crossing times, the clean SR versus RNE write-back contrast where only the write-back policy changes, and a fully prospective allocation test with the mask frozen before any target run. The author also deserves credit for reporting the negative results at full strength: the locked 64-cell transfer fails, the modern-decoder trajectories are uninformative, the systems check shows no net benefit, and the ECO comparator is post hoc. The paper does not oversell.\n\nThe soft spots are real but mostly acknowledged. The projected-reference proxy in Eq. (4) is the load-bearing piece for the aggregate audit claim, and the paper's own locked transfer shows it fails: paths diverge by update 3, same-code premise holds in 0.8% of coordinate-events, every uncertainty gate fails. So the 'reference traces can audit visibility' headline is only supported in small controlled testbeds. That is not an internal inconsistency — the paper says so — but it means the audit contribution is much narrower than the title suggests. Also, no code or data is public, only a private Zenodo record and hashes; the allocation result rests on three fixed target seeds and one architecture; and the post hoc ECO result, while stronger, has no systems cost measured. These limitations cap significance, but they do not make the core exact-event math wrong.\n\nWho gets value: anyone working on low-precision training, exact-budget protection, or registered/reproducibility methodology. The paper is a good example of how to run predeclared gates and then report the failed gates without hiding them. I think it deserves a serious referee: the referee should push on whether the audit claim should be further scoped or reframed, and on whether the negative transfer results belong in the main text rather than appendices.\n\nRecommendation: send to peer review. Conditional accept or major revision, but not desk reject.","headline":"A careful, self-limiting study of a real low-precision failure mode: the exact freeze event is clean, the audit proxy is honest but fails transfer, and the allocation result is genuinely prospective. Worth refereeing.","tokens_in":32759,"tokens_out":2153,"would_cite":true,"duration_ms":26342,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A matched high-precision trace can audit when low-precision weight updates vanish and, frozen before training, can select the exact 10% of weights to protect, recovering 60.7% of the loss gap on unseen targets.","keywords":["invisible weight updates","round-to-nearest","stochastic rounding","low-precision training","reference traces","fp32 master weights","proposal contribution","gap recovery"],"falsifier":"Run the projected-reference proxy (Eq. 4) on a previously unseen model or seed with a predeclared ±15% band and compare its first-majority-crossing forecasts against measured code-equality crossings; the proxy is falsified if fewer than about 90% of forecasts fall in the band, or if measured and projected code vectors diverge by update 3 as they do in the paper's own locked 64-cell test.","tokens_in":31766,"feed_emoji":"🎯","tokens_out":8301,"duration_ms":72793,"temperature":0.7,"pith_summary":"This paper tries to establish that the moment a low-precision weight update becomes invisible is an exact, testable arithmetic event, and that a separate high-precision reference trace can be used in two distinct ways: to audit how many weights are frozen over time, and to rank which weights deserve an fp32 master copy. The audit forecasts aggregate visibility accurately in controlled, same-schedule testbeds, but the paper is explicit that the proxy is not a pathwise certificate: in a predeclared transfer test the measured and projected paths diverge by update 3, and in a modern decoder the trajectories are uninformative. The allocation use is stronger: one 600-step source trace, sealed before three unseen targets, selects an exact 10% fp32-master tile mask that recovers 60.7% of the round-to-nearest-versus-fp32 loss gap on average, nearly matching target-specific masks (60.8%) and far exceeding fixed controls (17.9%). The paper also shows that changing only the write-back rule from round-to-nearest to stochastic rounding recovers 89–92% of that gap at 10,000 steps, and that the tested selective-master implementation saves peak memory against a blanket master but is slower than direct write-back, so it currently has no net systems benefit.","feed_headline":"Short trace preselects 10% of weights to guard; 61% of gap saved","feed_subtitle":"One 600-step pilot, sealed before training, matches per-target tuning (60.8%) and beats fixed controls (17.9%) in a 162M-parameter decoder.","key_machinery":"The load-bearing object is the exact coordinate event Q_m(q_i + d_i) = q_i: the stored low-precision code is bitwise unchanged after quantization of the optimizer's proposal. The audit uses a projected-reference proxy (Eq. 4), which replays a high-precision trajectory onto the target grid to estimate the frozen fraction without any low-precision observation. The allocation study uses a third object: absolute proposal contribution |g_i u_i|, averaged per 768-coordinate tile over a 600-step fp32 reference, which ranks tiles so the top 10% of the eligible fp32-master payload can be protected in advance. Each carries a different burden: the exact event is a fact; the proxy is an empirical aggreg","core_discovery":"Central claim: a low-precision update becomes invisible when the quantizer leaves the stored code unchanged — an exact testable event — and a matched high-precision reference trace can both audit how many weights freeze over time and rank which weights deserve an fp32 master copy. The audit forecasts crossing times across an 11–4222-step range (median ratio 1.000, 94.5% within 15%) but is not a pathwise certificate: in a locked 64-cell test paths diverge at update 3, and in a modern decoder the trajectories are uninformative. The allocation use is stronger: one 600-step source trace, frozen before three unseen targets, picks an exact-10% fp32 mask that recovers 60.7% of the gap on average, v","pith_inferences":["Editorial inference: the near-tie between the pre-frozen source mask (60.7%) and privileged target-specific masks (60.8%) suggests a single short pilot run may suffice to pick protective masks for a whole family of seeds; a direct test would freeze one mask and evaluate across different batch orders, schedules, or widths within the same architecture.","Editorial inference: the audit's failure once measured and projected paths diverge suggests a practical hybrid: use the offline proxy for early planning, then switch to exact online code-equality monitoring once divergence is detected, instead of trusting the aggregate throughout training.","Editorial inference: the post hoc ECO-Adam comparator recovering 96.5% with unmeasured cost points to a question the paper leaves open — whether error-feedback compensation can replace selective master weights entirely, provided its memory and throughput costs are counted.","Editorial inference: because write-back-policy recovery declines between 3,000 and 10,000 steps, the benefit of stochastic rounding may shrink as runs lengthen; testing beyond 10,000 steps would show whether these short-horizon contrasts persist."],"forward_implications":["Coordinate-level update invisibility is exactly testable: for any optimizer and quantizer, the code-equality condition decides whether a stored weight actually changed, independent of loss behavior.","Reference-derived frozen-fraction proxies forecast aggregate visibility timing well in controlled, same-schedule grids (94.5% of forecasts within 15% over an 11–4222-step span), but the paper's own transfer tests show the proxy is not transportable as a calibrated predictor.","Changing only the post-optimizer write-back from round-to-nearest to stochastic rounding removes the deterministic never-moved population and recovers 89.2–92.2% of the RNE-vs-fp32 loss gap at 10,000 steps in a 124M-parameter transformer.","A single 600-step source trace, frozen before three unseen target seeds, selects an exact-10% fp32-master tile mask that recovers a mean 60.7% of the gap, within 0.0014 of target-specific masks and 0.428 above the best fixed control; a depth-17 migration recovers 66.4%.","The current selective-master implementation reduces peak allocation by 584.65 MB versus a blanket master but adds 65.85 MB and lowers arm-median timed-window throughput by 7.196% versus direct round-to-nearest, so as implemented it has no validated break-even."],"fun_headline_variants":["One 600-step trace picks 10% of weights; saves 61% of loss gap","10% mask from short trace matches per-target tuning: 60.7% gap recovery","Pre-trained trace beats fixed control: 61% vs 18% loss gap recovered","Audit trace fails to transfer; allocation trace succeeds: 61% gap saved"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The audit's projected-reference proxy assumes the high-precision trajectory stays a useful stand-in for the low-precision path near the rounding boundary; once the two paths diverge, every aggregate uncertainty claim loses support, although the exact whether-a-weight-changed event remains valid.","fun_headline_variants_meta":{"raw":{"variants":["One 600-step trace picks 10% of weights; saves 61% of loss gap","10% mask from short trace matches per-target tuning: 60.7% gap recovery","Pre-trained trace beats fixed control: 61% vs 18% loss gap recovered","Audit trace fails to transfer; allocation trace succeeds: 61% gap saved"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001033,"raw_usage":{"total_tokens":4312,"prompt_tokens":997,"completion_tokens":3315,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":741,"completion_tokens_details":{"reasoning_tokens":3222}},"tokens_in":741,"tokens_out":3315,"duration_ms":18988,"temperature":1.0,"reasoning_tokens":3222,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:49:10.577874+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the projected-reference proxy (Eq. 4) on a previously unseen model or seed with a predeclared ±15% band and compare its first-majority-crossing forecasts against measured code-equality crossings; the proxy is falsified if fewer than about 90% of forecasts fall in the band, or if measured and projected code vectors diverge by update 3 as they do in the paper's own locked 64-cell test.","supporting_citations":[],"review_version":2}