{"id":"73cfdec4-3756-4396-9815-b9496fe863d4","arxiv_id":"2608.00461","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Staged equivalence-preserving optimizations cut JUNO's OMILREC reconstruction time from 1524.8 to 189.2 ms/event (8.06x) on an Intel Xeon, with numerical drift below 1.3e-14.","lead":"This paper makes the JUNO detector's event reconstruction code run eight times faster on a single CPU core while keeping the physics results essentially unchanged. It is a practical recipe for accelerating similar large-detector reconstruction software.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Verification envelope is too narrow: 16 FCN points and a z-axis scan cannot support the universal 'preserves reconstruction' claim, and reported exceptions already exceed the 4mm/7keV gate.","rationale":"The reader's weakest assumption correctly identifies the core issue: the verification gates are based on a small set of fixture points and one calibration scan, and the paper itself documents near-boundary events exceeding the stated gate. My stress-test reading aligns: the central claim is a universal equivalence claim, but the evidence only supports equivalence for a 'typical' subset. The FCN gate's 16 points are too few to bound the minimizer’s trajectory, and the acceptance gate’s aggregate metrics can hide rare but physics-relevant deviations (up to 150 mm/55 keV). This is not an accusation of fraud; the paper is transparent about the deviations and flags a z-bias drift. However, the verification envelope must be quantified and expanded before the universal claim can be accepted. I therefore recommend no change to the reader's CONDITIONAL verdict: the paper's contribution is credible but requires additional evidence (a volume-and-energy-spanning test or an explicit, well-justified exception policy) to move to full acceptance. The proposed concrete test would directly address this gap.","tokens_in":7916,"tokens_out":7597,"duration_ms":81034,"concrete_test":"Generate 10,000 Monte Carlo events with vertices uniformly distributed within the detector volume (r ≤ 17 m) and energies spanning 0.5–10 MeV. Run both the baseline and optimized (v1.7.1) reconstructions on identical events. Compare vertex and energy pairwise; report the maximum and 99.9th percentile of |Δvertex| and |ΔE|. Pre-specify a pass criterion: 99.9th percentile ≤ 4 mm and ≤ 7 keV, and no event exceeding 5× these thresholds. Additionally, run the existing acceptance metrics on an off-axis scan (e.g., along the x-axis) and on a 2D slice, and require all eight metrics to pass. If off-axis or higher-energy bins show bias/resolution degradation beyond tolerance, the coverage assumption fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the speedup preserves the reconstruction output. This is supported by three gates: an FCN unit test on 4 events × 4 stages (16 points), an end-to-end gate requiring 4 mm/7 keV, and an eight-metric acceptance gate on an ~861,000-event 68Ge z-axis calibration scan. The load-bearing assumption is that these samples cover the full production phase space. This is not established: the z-axis scan samples only vertices along one axis, the FCN fixture points are sparse in the high-dimensional parameter space Minuit explores, and the end-to-end gate's 4 mm/7 keV criterion is already violated by the paper's own data. Table 2 reports a max position difference of 5.1 mm at v1.3.0 and up to 150 mm/55 keV at v1.4.0+ for near-boundary events; §6.2 says these 'pass the physics-acceptance gate' because the gate is an aggregate metric that can hide rare large deviations. If such deviations are not limited to a negligible fraction and to a specific region, the universal 'preserves reconstruction' claim is not quantitatively supported. The paper explicitly qualifies the 4 mm/7 keV as holding for 'typical events,' but it does not define the size of the 'atypical' set or provide evidence of its proportion in the full production phase space (off-axis vertices, other energies, other event topologies).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a staged, equivalence-preserving optimization of the JUNO OMILREC vertex/energy reconstruction, reducing single-thread reconstruction time from 1524.8 ms/event to 189.2 ms/event (8.06x) on an Intel Xeon 8358P and from 705.1 to 134.9 ms/event (5.22x) on an AMD EPYC 9654, with a further 8.6x figure at v1.12.0. The optimization is organized as a sequence of latency-removing transformations (data layout flattening, vectorizable geometry, hoisting of Minuit-invariant work, per-event precomputation, fit-phase loop splitting, reduced-precision fast paths, and a better minimizer seed), each gated by a frozen-reference regression test. The authors claim the likelihood is bit-identical through v1.0.7 and within 1.3e-14 relative drift thereafter, and that typical reconstructed vertices/energies agree with the baseline within 4 mm/7 keV. An eight-metric acceptance gate passes on ~861,000 68Ge calibration events, with a small z-bias drift flagged as under study. The work is framed as a transferable template for accelerating likelihood-based reconstruction in large detectors.","tokens_in":8290,"tokens_out":3752,"duration_ms":42889,"significance":"If the equivalence claim is established quantitatively, this is a practically important result for JUNO's offline computing: an order-of-magnitude single-thread speedup that drops into production without a re-validation campaign is valuable, and the profiling-led optimization patterns (flatten data, hoist invariants, specialize loops) are transferable to other per-channel maximum-likelihood fits. The paper's verification methodology is a genuine strength: the frozen-reference contract, per-commit drift ledger, and staged gates make the equivalence claim auditable in principle. The early releases being bit-identical and the explicit reporting of residual deviations are also positive features. However, the universality of the claim rests on a narrow verification envelope, and the paper's own data show deviations larger than the stated gate; these gaps must be closed before the headline claim is fully supported.","major_comments":[{"comment":"The end-to-end gate is defined as agreement within 4 mm and 7 keV, but Table 2 reports a maximum position difference of 5.1 mm at v1.3.0 and 150 mm/55 keV at v1.4.0+. If the gate is a per-event maximum, v1.3.0 already fails the stated criterion. If the gate is intended for 'typical' events only, the paper must define the exceptional set (e.g., a spatial region, a quantile, or a fraction of events) and show that this set is negligible in the production phase space. As written, the paper's own data contradict a literal reading of the gate, and this is load-bearing for the central 'preserves reconstruction' claim.","section":"§4.1 and §6.2, Table 2"},{"comment":"The acceptance test reports a monotonic z-bias drift 'within the 10 mm absolute tolerance', but the end-to-end physics gate is stated as 4 mm. These tolerances are inconsistent as written. Either the 10 mm is the tolerance for the acceptance gate (in which case the 4 mm gate is not the governing criterion for the large sample) or the drift actually violates the 4 mm gate and the acceptance claim needs re-evaluation. Please clarify which tolerance applies where and report the magnitude of the drift explicitly.","section":"§6.3 and §4.1"},{"comment":"The equivalence claim is supported by only 16 FCN fixture points (4 events x 4 stages), a 100-event benchmark, and a z-axis calibration scan. The paper itself shows that near-boundary events deviate by up to 150 mm and 55 keV, meaning the verification sample already contains cases outside the nominal gate. The z-axis scan samples vertices along a single axis only; it does not cover off-axis vertices, different energies, or other event topologies. No evidence is given that the fixture events and scan span the production phase space. The acceptance gate uses aggregate bias/resolution metrics that can hide rare large deviations. To support the universal phrasing, the authors should report tail quantiles (e.g., 99.9th percentile deviations) and phase-space coverage, and either restrict the claim accordingly or enlarge the verification set.","section":"§4.1 and §6.2"},{"comment":"The headline speedup of 8.06x is based on a single 100-event sample on two machines, with no run-to-run variance, no statistical uncertainty, and no detailed composition of the sample beyond 'nH-calibration run 12628'. A single 100-event sample is a narrow basis for a result quoted to three decimals. At minimum, report multiple independent runs (or the run-to-run spread) and specify the sample's phase-space coverage; if the benchmark time is known to be stable, state that explicitly with supporting data. This does not undermine the qualitative speedup but currently overstates the quantitative precision.","section":"§6.1 and §6.6 (Reproducibility)"}],"minor_comments":[{"comment":"Typo: 'framwork' should be 'framework'.","section":"Introduction"},{"comment":"Grammar: 'The agent development workflow are reported separately' should be 'is reported separately'.","section":"§4.3"},{"comment":"The v1.12.0 result is included in Table 1 and Figure 1 but §6.1 says it is 'reported separately'. Please reconcile whether v1.12.0 is part of the main progression or a separate extension.","section":"§6.1"},{"comment":"The phrase 'offers great guidance and help' is informal; consider 'provided guidance and support'.","section":"Author contributions"},{"comment":"The benchmark and drift data are only 'available on reasonable request'. Given the paper's auditability argument, making the per-release benchmark logs and drift.csv publicly available (even as supplementary material) would substantially strengthen the reproducibility claim.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper's core speedup is credible and the methodology is a step forward, but the equivalence claim is currently overreaching relative to the evidence: the verification set is small, the z-axis scan is not phase-space representative, and the reported worst-case deviations (5.1 mm, 150 mm) conflict with the stated 4 mm gate. These are fixable with additional analysis (tail quantiles, phase-space coverage, clarified tolerances) rather than a fundamental flaw. I would encourage the editor to request a major revision focused on those points. There is also a question of fit with EPJ C scope: the paper is mostly software engineering with a physics-validation component; that is acceptable for a detector journal, but the authors should make the physics-impact case more explicit. The AI-assistance disclosure is transparent and appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading: the speedup is real and the paper is unusually honest about its limits. The authors don't oversell—they report 8.06x on one Xeon, 5.22x on an EPYC, and they explicitly flag the boundary-event deviations and the z-bias drift as open items. That transparency is the paper's best feature.\n\nWhat's new: most HEP acceleration work is GPU or SIMD rewrites that change the numerics and require re-validation. This paper shows a single-thread, equivalence-preserving optimization of a production reconstruction, gated at every step against a frozen baseline reference. The methodology—FCN drift ledger, end-to-end physics gate, large-sample acceptance gate—is a genuinely transferable template. The profiling diagnosis (latency-bound at ~10% of scalar peak, not compute-bound) is clean and actionable. For typical events, the equivalence claim holds: bit-identical through v1.0.7, FCN drift at 1.3e-14, and the physics gate met.\n\nThe soft spots are real but not disqualifying. The benchmark is one 100-event sample on two machines, with no error bars; code and data are not public, only \"available on reasonable request.\" The verification envelope is narrower than the claim: 16 FCN fixture points and a z-axis calibration scan do not sample the full phase space off-axis, other energies, or other topologies. The more serious issue is that Table 2 shows up to 150 mm / 55 keV differences for near-boundary events after v1.4.0, and the paper says these \"pass the physics-acceptance gate\" only because the gate is an aggregate metric. That is a meaningful caveat: the 4 mm / 7 keV gate holds for \"typical\" events, but the atypical set is neither quantified nor bounded. The authors say so themselves, which is honest, but it means the universal \"preserves the reconstruction\" claim is not yet fully supported.\n\nStill, this is a solid piece of software engineering for a large experiment. The authors know their code, they measured carefully, and they disclose what they haven't fixed. The paper deserves a serious referee, and the main revisions should be: quantify the fraction and location of boundary events, report timing variance, and either release the code or provide a more detailed public artifact.\n\nFor you: if you work on HEP reconstruction or offline software, this is worth citing. I'd bring it to a reading group as an example of how to do equivalence-gated optimization well. Peer review: yes, send it out—but expect a request for the atypical-event analysis.","headline":"A credible, transparent report of an 8x single-thread speedup for JUNO's OMILREC, with a real verification methodology but a verification envelope that is narrower than the universal equivalence claim.","tokens_in":8680,"tokens_out":1015,"would_cite":true,"duration_ms":14731,"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":"By profiling the bottleneck as latency rather than arithmetic, the paper's staged, equivalence-gated optimizations cut JUNO's OMILREC reconstruction time from 1524.8 to 189.2 ms/event on a single CPU thread while keeping the likelihood and","keywords":["JUNO","OMILREC","event reconstruction","maximum-likelihood fit","software optimization","high-performance computing","numerical equivalence"],"falsifier":"Run both the baseline and optimized reconstructions on a large sample of simulated events uniformly distributed through the detector, including the boundary region, and compare per-event output. If any production-type event outside the frozen fixtures differs by more than 4 mm in vertex or 7 keV in energy (or shows likelihood drift above 10^-13), the equivalence claim fails. A positive control is to take the near-boundary events where the improved seed already moves the result by up to 150 mm/55 keV and check whether they now violate the physics gate.","tokens_in":7890,"feed_emoji":"⚡","tokens_out":6637,"duration_ms":73608,"temperature":0.7,"pith_summary":"The paper tries to prove that a large maximum-likelihood reconstruction can be made roughly eight times faster on a single CPU core without changing the physics it outputs. It diagnoses the production algorithm as latency-bound—running near 10% of scalar floating-point peak because of virtual-function calls, pointer chasing, and repeated recomputation—and then applies a staged set of optimizations that remove indirection rather than change the mathematics. Each step is checked against a frozen, never-regenerated reference likelihood, so the speedup is auditable instead of asserted. The result is a drop-in replacement: bit-identical output through the first seven releases and, afterward, output within a 4 mm/7 keV physics gate and a likelihood drift below the 10^-13 contract.","feed_headline":"Eightfold faster JUNO reconstruction preserves physics output","feed_subtitle":"Profile-guided, equivalence-gated optimizations cut per-event time from 1525 to 189 ms on a single CPU core.","key_machinery":"The load-bearing mechanism is a staged, equivalence-gated optimization library. The stages are: flattening per-photomultiplier data into contiguous arrays, bulk vectorizable geometry passes, hoisting of minimizer-invariant work out of the repeated fit calls, per-event precomputation, fit-phase loop splitting and indexing, and reduced-precision fast paths. The equivalence contract is enforced by a never-regenerated golden fixture pack of likelihood values at a frozen set of test points (4 events x 4 fit stages), a per-commit drift ledger, an end-to-end 4 mm/7 keV gate, and an eight-metric acceptance gate on the full calibration scan. Each optimization must pass this contract before being acce","core_discovery":"The central claim is that equivalence-preserving, single-thread optimization can retire the latency of a per-LPMT maximum-likelihood fit without altering its results. The production OMILREC reconstruction, which scans all 17,612 photomultiplier tubes in roughly 470 minimizer evaluations per event, was accelerated from 1524.8 to 189.2 ms/event (8.06x) on an Intel Xeon core and from 705.1 to 134.9 ms/event (5.22x) on an AMD EPYC core, with a further 8.6x figure after additional precomputation. The likelihood remained bit-identical through releases v1.0.2–v1.0.7, then stayed within a relative drift of 1.3e-14 against a 10^-13 contract. Typical reconstructed vertices and energies agree with the","pith_inferences":["The authors leave implicit that the frozen fixture pack should be expanded to boundary and simulated events before production rollout; their own data show the improved seed already shifts some boundary events by up to 150 mm/55 keV.","A broader lesson: code running near 10% of scalar peak because of pointer chasing and virtual dispatch may have order-of-magnitude single-thread headroom without changing algorithms—profiling first is the reusable move.","The different minima found by the improved seed at boundaries suggest the baseline itself may sometimes settle into worse local minima; comparing physics results on a large sample would tell whether the optimized code is not only equivalent but occasionally more accurate.","The workflow demonstrates that an AI coding agent can be safely used for aggressive optimization when every candidate change is checked against an immutable reference, which may lower the maintenance cost of optimized production code in other experiments."],"forward_implications":["JUNO reconstruction throughput multiplies by roughly eight on existing single-thread resources, since the gain composes with the existing multi-threaded production scheduling.","The profiling diagnosis directs optimization effort from 'do less mathematics' to 'remove indirection and expose the loop,' a pattern that applies to any likelihood-based reconstruction iterating over tens of thousands of channels inside a minimizer.","The bit-identical early releases and gate-bounded later releases allow the optimized code to enter production without a fresh physics re-validation campaign.","The residual runtime is concentrated in the time-based fit stages (roughly 65% of the total), which are the natural next target; SIMD and GPU offload of the inner LPMT loop remain available.","The transferable recipe—diagnose the bottleneck with hardware counters, then optimize under a frozen-reference equivalence contract—offers a template for similar reconstructions in other large detectors."],"supporting_citations":[{"why":"Defines the data-driven charge-and-time likelihood that OMILREC maximizes; the algorithm being optimized.","marker":"[3]"},{"why":"Supplies the numerical minimizer whose repeated function evaluations dominate the runtime that is profiled and optimized.","marker":"[5]"},{"why":"Provides the histogram-interpolating accessor layer identified as a primary source of pointer chasing and indirection latency.","marker":"[6]"},{"why":"Supplies the reproducibility practice that motivates the frozen-reference equivalence-gate methodology.","marker":"[16]"},{"why":"Represents the contrasting GPU-rewrite approach whose numerical results differ, against which the equivalence-preserving contribution is positioned.","marker":"[11]"}],"fun_headline_variants":["JUNO reconstruction 8x faster, physics unchanged","Equivalence-preserving speedup: JUNO's OMILREC 8x quicker","8x speedup for JUNO vertex/energy reconstruction, identical results","Bit-identical likelihood, 8x faster: JUNO reconstruction revamp","JUNO reconstruction tuned to 8x speed without altering output"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The equivalence guarantee is certified only on the small frozen fixture pack (4 events x 4 fit stages) and a single 68Ge calibration scan; the assumption that these samples represent the full production event space—including near-boundary events that already show 150 mm/55 keV shifts when the seed changes—is the load-bearing premise.","fun_headline_variants_meta":{"raw":{"variants":["JUNO reconstruction 8x faster, physics unchanged","Equivalence-preserving speedup: JUNO's OMILREC 8x quicker","8x speedup for JUNO vertex/energy reconstruction, identical results","Bit-identical likelihood, 8x faster: JUNO reconstruction revamp","JUNO reconstruction tuned to 8x speed without altering output"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3177,"prompt_tokens":977,"completion_tokens":2200,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":721,"completion_tokens_details":{"reasoning_tokens":2101}},"tokens_in":721,"tokens_out":2200,"duration_ms":13724,"temperature":1.0,"reasoning_tokens":2101,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:55:20.550051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both the baseline and optimized reconstructions on a large sample of simulated events uniformly distributed through the detector, including the boundary region, and compare per-event output. If any production-type event outside the frozen fixtures differs by more than 4 mm in vertex or 7 keV in energy (or shows likelihood drift above 10^-13), the equivalence claim fails. A positive control is to take the near-boundary events where the improved seed already moves the result by up to 150 mm/55 keV and check whether they now violate the physics gate.","supporting_citations":[{"cited_title":"Data-driven simultaneous vertex and energy reconstruction for large liquid scintillator detectors","cited_arxiv_id":"2211.16768","evidence_quote":"Defines the data-driven charge-and-time likelihood that OMILREC maximizes; the algorithm being optimized."},{"cited_title":"James, M","cited_arxiv_id":null,"evidence_quote":"Supplies the numerical minimizer whose repeated function evaluations dominate the runtime that is profiled and optimized."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the histogram-interpolating accessor layer identified as a primary source of pointer chasing and indirection latency."},{"cited_title":"Reproducibility and Replication of Experimental Particle Physics Results","cited_arxiv_id":"2009.06864","evidence_quote":"Supplies the reproducibility practice that motivates the frozen-reference equivalence-gate methodology."},{"cited_title":"Allen: A high level trigger on GPUs for LHCb","cited_arxiv_id":"1912.09161","evidence_quote":"Represents the contrasting GPU-rewrite approach whose numerical results differ, against which the equivalence-preserving contribution is positioned."}],"review_version":1}