{"id":"c561f1fc-6510-451b-8e0c-2080fb38570b","arxiv_id":"2607.22433","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VIREL compresses floating-point time series by predicting residuals in route-local integer/lattice coordinates, beating exact and error-bounded baselines on smooth streams.","lead":"This paper introduces VIREL, a database-page compressor that predicts floating-point time series using separate per-resolution integer lanes and reports 6–9.6× exact and 12.1× error-bounded compression on smooth streams. It matters because storage and scan cost in time-series databases could drop by 20–30% over current codecs if these gains generalize.","discovery_kind":"new_method","skeptic_critique":null,"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VIREL, a floating-point page codec that changes the representation domain that carries temporal prediction state. Instead of predicting over IEEE 754 words, erased words, decimal fields, or a page-wide integer stream, VIREL admits values to exact or error-bounded integer/lattice coordinates (Eqs. 4–9), routes them by source resolution, and optionally applies lattice-step normalization (affine q=dz+r for exact; q/d for error-bounded) before computing residuals. The authors claim large ratio gains over existing page codecs under equal 1,024-value page resets: 6.02–7.03× exact on the canonical 14-stream suite and 8.06–9.65× on 74.70 million values, plus 12.11× error-bounded at ε=10⁻³, with bit-exact and pointwise error verification, ablations, multicore scaling, and Apache TsFile integration.","tokens_in":20429,"tokens_out":26137,"duration_ms":538459,"significance":"If the results hold, this is a solid systems contribution: it identifies the representation domain of temporal prediction state as a first-class page-codec design choice, and it demonstrates that route-local lattice-coordinate residuals with cost-based normalization are effective on smooth, decimal/instrument-lattice workloads. The evaluation is unusually thorough: equal-reset protocols, complete encoded-size charging, verbatim losslessness and pointwise bound checks, a 48-stream and a 15-stream corpus, ablation isolating the key mechanisms, and an end-to-end TsFile comparison. The main weaknesses are the absence of a failure-mode evaluation on non-lattice data and the lack of a public artifact, both of which are needed to support the generality and reproducibility of the central empirical claims.","major_comments":[{"comment":"The central mechanism depends on the workload being lattice-like. §IV-A explicitly routes values that do not pass exact/EB admission (Eqs. 4–9) to the raw XOR lane, and §I asserts that 'many floating-point time series are generated on decimal or instrument lattices.' However, all evaluated corpora (ELF canonical suite, UCI, GeoLife, T-Drive, Serf) are smooth, decimal-like streams; the paper does not measure the failure mode where few values are admitted, e.g., random-mantissa scientific data or pseudo-random binary64. The generality claim is therefore unquantified. Please add experiments on non-lattice workloads reporting admission rate, compression ratio, and throughput against the same baselines, and discuss the boundary of applicability.","section":"§I, §IV-A, §VI"},{"comment":"No code artifact is linked, and the manuscript omits implementation details needed to reimplement the headline compression-first profile: the 'fixed five-predictor integer family' in §IV-B is not enumerated, and the candidate divisor derivation and REFIT order in §IV-C are only described at a high level. Since the central claims are empirical and the harness is custom, a public artifact or a detailed reproducibility appendix is necessary for the results to be verifiable. Please provide an artifact, or at minimum a precise specification of the enumerated predictors and the descriptor search.","section":"§VI-A, §IV-B, §IV-C"},{"comment":"The primary exact results are described as 'independent 1,024-value calls,' yet §IV-D states that the Upper profile 'stores repeated stream metadata once per page group and applies Zstd-22 to the resulting length-delimited group.' If a 'page group' spans multiple pages, metadata is amortized across pages, which would not be an equal-reset comparison with baselines charged complete per-page output. Please define 'page group' and state its size in the independent-page experiment; if it comprises multiple pages, re-run the main ratio with strictly per-page metadata or clearly separate the page-group experiment from the equal-reset claim.","section":"§IV-D, §VI-A"}],"minor_comments":[{"comment":"The sentence 'VIREL-Exact-Fast reaches 6.0243× and VIREL-Exact-Upper reaches 7.0287×, emitting 22.4% fewer bytes than the strongest evaluated exact baseline' is ambiguous: the 22.4% figure applies to Exact-Upper versus Falcon (1 − 5.4527/7.0287), not to both profiles. Please rephrase to attribute the percentage explicitly.","section":"Abstract"},{"comment":"The text uses both '1,000-value resets' (legacy, decimal-precision section) and '1,024-value pages' for the primary results. Please use consistent terminology or explicitly state that the 1,000-value protocol is a legacy harness used only where noted.","section":"§VI-A"},{"comment":"The binary32 results (20.38%/24.19% exact; 3.41% EB) are mentioned in prose but not shown in any table. Please add a table or appendix with these comparisons.","section":"§VI-B"},{"comment":"The 'Plain' row reports ratios below 1 (0.9697/1.5398). Please add a sentence explaining that this is due to TsFile/page overhead, to avoid confusion.","section":"Table XI"},{"comment":"The REFIT deterministic order ('lowest valid scale for Fast, fixed descriptor order for exact affine lanes, and largest selected divisor for EB') is not fully specified. Please list the exact descriptor order used for affine lanes.","section":"§IV-C"},{"comment":"The 'SElfStar artifact' is cited by name but no URL is given; please add a reference or footnote.","section":"§VI-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound and the evaluation is unusually careful, with equal resets, complete-size charging, and explicit losslessness/error checks. The main obstacles are the missing failure-mode evaluation on non-lattice data and the lack of a code artifact; both are addressable within the manuscript's scope. The 'page-group' ambiguity in §IV-D should be resolved because it could affect the equal-reset claim if it implies cross-page metadata amortization. I recommend major revision rather than rejection; the central idea is defensible and the requested additions are concrete."},"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that floating-point time-series compression improves when temporal prediction state is carried as route-local lattice-coordinate residuals rather than IEEE 754 words or a single integer stream.","keywords":["floating-point compression","time series","page codec","lattice coordinates","error-bounded compression","lossless compression","route-local prediction","temporal prediction state"],"falsifier":"Encode a page of doubles with random mantissas — for example, generated by hashing or from irrational constants — using VIREL and a strong XOR-based codec under identical 1,024-value page resets; if VIREL's admission rate is near zero and its compression ratio does not exceed the XOR baseline, the lattice premise fails for that regime.","tokens_in":1504,"feed_emoji":"🗜️","tokens_out":1815,"duration_ms":56520,"temperature":0.7,"pith_summary":"VIREL claims that the right place to keep temporal prediction state in a floating-point page codec is neither the IEEE 754 word nor one page-wide integer stream, but per-route lattice-coordinate residuals: values are first certified as integer coordinates that reconstruct bit-exactly (or within an error bound), then grouped by source resolution into separate lanes, each with its own history, and optionally normalized by a lattice step (q=dz+r storing z, or q'=q/d for divisible error indices) before residual coding. The payoff, if the claim holds, is that residuals become smaller before any entropy coding, so pages store fewer bytes without changing the fidelity contract. The paper reports 6.02x–7.03x exact compression on canonical pages, 8.06x–9.65x on 74.7M real values, and 12.11x at epsilon=10^-3, with all pointwise error bounds preserved. The deeper claim is that representation domain, not just predictor sophistication, determines the compressibility of smooth sensor and financial time series.","feed_headline":"Predicting floats as lattice coordinates cuts time-series bytes 6–12x","feed_subtitle":"Per-resolution integer lanes keep prediction state clean, so smooth sensor streams shrink more than with XOR- or word-domain codecs.","key_machinery":"Route-local lattice-coordinate residual prediction. The mechanism has four parts: (1) conservative admission via binary64 reconstruction checks — exact mode uses R_X(q,s)=fl(double(q)/10^s) and error-bounded mode uses R_B(o,q,delta)=fl(o+fl(q*delta)), so only values that reproduce the input word or satisfy the bound enter the integer domain; (2) a resolution router that splits the page into up to three or four integer lanes, each with its own previous-value predictor, so a fine-resolution exception does not reset the coarse lane's history; (3) lattice-step normalization — exact affine lanes store z for q=dz+r, and EB lanes store q'=q/d only for divisible indices, turning physical steps into","core_discovery":"The central discovery is that decimal- and instrument-generated floating-point time series carry unused structure: once a value is admitted as an integer lattice coordinate q under a checked binary64 reconstruction, consecutive coordinates often differ by far smaller residuals than adjacent IEEE 754 words, especially when the physical step is normalized away (q=5z+4 stores z, so +5 becomes +1). The paper argues that a page codec should therefore keep its temporal prediction state in route-local lattice coordinates: separate lanes for different source resolutions preserve each lane's history when high-precision exceptions appear, and cost-based lattice-step normalization removes sparse physic","pith_inferences":["Because the gain depends on how many values pass lattice admission, workloads with mostly random mantissas would see VIREL fall back to raw XOR; a cheap admission-rate histogram emitted per page could let a storage engine choose between VIREL and word-domain codecs dynamically.","The route-local state idea is generic: any predictive codec that works in a transformed coordinate space — decimal, quantized, or learned — could adopt per-resolution lanes to keep fine exceptions from corrupting coarse histories.","A direct testable extension is to feed VIREL synthetic streams with controlled fractions of lattice-representable values and map the break-even point where route-local prediction stops paying for its metadata cost.","If the representation-domain claim generalizes, future storage engines could choose prediction domain from column metadata, such as sensor make/model or declared decimal precision, rather than per-page search."],"forward_implications":["Exact pages of smooth sensor and financial streams compress to roughly one-sixth to one-seventh of their raw size, and the compression-first profile beats all evaluated exact baselines on the canonical 14-stream suite with equal 1,024-value resets.","On 48 larger real streams totaling 74.7M values, the exact profiles reach 8.06x and 9.65x compression, winning all 48 streams against a per-stream oracle of six exact baselines except one 0.54% loss.","Error-bounded compression gains without consuming error budget: at epsilon=10^-3 the EB profile reaches 12.11x while preserving every pointwise bound, with q/d factoring alone contributing an 18.75% byte reduction over the routed unfactored variant.","The page-streamable design — independent frames, deterministic decoding, bounded memory — transfers to a file format, writing 27.9–30.1% fewer complete-file bytes than the strongest exact baselines and scanning faster on full, range, and aggregate reads.","Ablations attribute the largest single share (53.05% byte growth when removed) to integer-domain residual prediction rather than routing or normalization, confirming the representation-domain claim."],"fun_headline_variants":["Lattice coordinates beat XOR for float time-series compression","Float compression: predict in lattice space, not IEEE words","VIREL: route-local lattice residuals shrink float streams 6–12x","Lattice residual prediction compresses float time series up to 12x","Cuts float time-series bytes 6–12x via route-local lattice residuals"],"cache_read_input_tokens":22528,"weakest_assumption_plain":"The workload's values mostly lie on decimal or instrument lattice coordinates that pass bit-exact or error-bounded reconstruction; if they do not, the coordinate path is empty and VIREL falls back to raw XOR encoding, erasing its advantage over word-domain codecs.","fun_headline_variants_meta":{"raw":{"variants":["Lattice coordinates beat XOR for float time-series compression","Float compression: predict in lattice space, not IEEE words","VIREL: route-local lattice residuals shrink float streams 6–12x","Lattice residual prediction compresses float time series up to 12x","Cuts float time-series bytes 6–12x via route-local lattice residuals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3209,"prompt_tokens":915,"completion_tokens":2294,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":2202}},"tokens_in":659,"tokens_out":2294,"duration_ms":18483,"temperature":1.0,"reasoning_tokens":2202,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T03:31:48.279326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Encode a page of doubles with random mantissas — for example, generated by hashing or from irrational constants — using VIREL and a strong XOR-based codec under identical 1,024-value page resets; if VIREL's admission rate is near zero and its compression ratio does not exceed the XOR baseline, the lattice premise fails for that regime.","supporting_citations":[],"review_version":2}