{"id":"88de6306-153f-4f89-a083-37d0b157121b","arxiv_id":"2607.21116","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Proton-beam tests of HCCStar ASICs show triple-modular redundancy corrects register bit flips and place LP-path data loss at O(10^-10) of the 400 kHz HL-LHC readout.","lead":"HCCStar control chips for the ATLAS inner tracker were exposed to 20–80 MeV protons to test their triple-modular-redundancy error protection. The authors estimate data loss in the low-priority readout path at about one in ten billion readout transactions, and roughly ten corrected register-bit flips per bit per year.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"One ASLP event on the shared LP multidrop line can corrupt multiple ABCStar readout transactions, so 'at most one transaction per event' is not a conservative upper bound and the O(10^-10) fraction may be O(10^-9).","rationale":"The reader's conditional verdict already identifies the LP-path assumptions as fragile and specifically mentions the single-transaction-per-ASLP assumption in the rationale. My concern agrees with that but elevates it to the single most load-bearing issue because it directly undercuts the 'conservative upper bound' language and the precise order-of-magnitude headline, rather than merely the chip-to-chip variation. The register-level TMR evidence (CBF cross sections near 10^-13 cm2/p/bit, large suppression of ASIR) is internally consistent and not in question. The energy/flux scaling concerns are real but less sharp: the paper already uses a measured worst-case cross section and the assumed flux is consistent with the stated lifetime fluence. The multiplicity issue is an internal inconsistency: §3 describes shared multidrop lines, while §5 assumes one affected transaction per ASLP event. A single fault-injection test can resolve it. Even if the concern lands and the fraction becomes O(10^-9), the operational conclusion that LP data loss is negligible remains intact, so no change to the conditional verdict is warranted; the paper should simply correct the bound and the order-of-magnitude statement.","tokens_in":5908,"tokens_out":7974,"duration_ms":78692,"concrete_test":"Perform a fault-injection test on the HCCStar LP transmit path using the ITSDAQ setup: inject a single SEU/SET (e.g., via laser or microbeam) into the LP logic while monitoring all connected ABCStar response headers, and count the number of L0tag/BCID mismatches generated by that one ASLP event. If the count exceeds 1 (because the shared multidrop line fans out to multiple ABCs), multiply the 546 annual anomalies by that factor and recompute the data-loss fraction; if the resulting fraction is O(10^-9) or greater, the O(10^-10) headline is not conservative.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline quantitative claim — a data-loss fraction O(10^-10) from 546 LP anomalies per HCC per year — rests on the statement in §5: 'We therefore associate uncorrected SEEs on the LP control path (ASLP events, Table 1) with at most one affected readout transaction per event and use this as a conservative upper bound.' This is not conservative. Section 3 says the HCC distributes 'triggers, event data readout requests, register read and write commands, test pulse generation, and reset signals' to between 6 and 11 ABCStar ASICs 'via shared multidrop lines.' The LP low-priority readout request is one of these shared lines. A single SET on a shared multidrop line can in principle be received by every ABCStar on the hybrid, so one ASLP event can corrupt the L0tag/BCID matching for multiple readout transactions, not just one. The 'at most one transaction' mapping is therefore a lower bound on the impact, not an upper bound. Recomputing with, say, 11 affected ABCs raises 546 to roughly 6000 anomalies per HCC per year, giving a fraction ~1.5e-9 — O(10^-9), not O(10^-10). The broad conclusion that LP-path data loss is negligible would survive, but the paper's specific claim of a 'conservative upper bound' and the precise O(10^-10) number are not supported without a bound on this multiplicity. The chip-152 TMR-on LP events (13 vs 15) reinforce that the LP path may not be protected as effectively as registers, but this multiplicity issue is the more direct threat to the quantified claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports proton irradiation tests of HCCStar ASICs for the ATLAS ITk strip upgrade, performed at CSNS with 20–80 MeV protons. It measures corrected bit flips (CBFs) in TMR-protected registers, actual SEEs in registers (ASIR) and in the low-priority (LP) control path (ASLP), and derives cross sections as a function of energy. Using the measured ASLP cross section and an assumed HL-LHC hadron flux of O(10^7) cm^-2 s^-1 for 10^7 s/year, it estimates 546 LP-related readout anomalies per HCC per year, a data-loss fraction O(10^-10) relative to the 400 kHz LP readout, and O(10) corrected bit flips per bit per year. It concludes that TMR effectively reduces SEEs and that LP data loss is negligible.","tokens_in":1416,"tokens_out":1385,"duration_ms":41770,"significance":"The raw data are valuable: direct proton-beam measurements on a production-relevant ASIC, with a large accumulated fluence (~48% of HL-LHC lifetime), and a clear TMR-on/off comparison for registers. The cross sections can feed rate estimates for the ITk upgrade. The rate calculation is transparent and, conditional on the flux assumption, the arithmetic is straightforward. However, the paper's headline quantitative claims rest on two fragile steps: the interpretation of ASLP events as affecting at most one readout transaction, and the generalization from two chips. Neither is fatal to the general conclusion that LP-path SEEs are subdominant, but the specific O(10^-10) bound is not established.","major_comments":[{"comment":"The 'at most one affected readout transaction per event' mapping is not a conservative upper bound. Section 3 states that the HCC distributes triggers, readout requests, and control signals to 6–11 ABCStars via shared multidrop lines. An SET on the shared LP line can corrupt the L0tag/BCID matching for all ABCStars on the hybrid. With the maximum 11 ABCStars, 546 anomalies becomes ~6000 per HCC per year, a fraction ~1.5×10^-9, not O(10^-10). The paper should either bound the multiplicity from data (e.g., by correlating anomalies across input channels) or rephrase the result as a lower bound on data-loss impact. This is the central quantitative claim.","section":"§5, Eq. (5.4) and preceding paragraph"},{"comment":"The TMR-on ASLP count for chip 152 (13 events) is statistically indistinguishable from the TMR-off count (15 events) under Poisson statistics. Only chip 154 shows TMR suppressing ASLP (0 vs 32). The text acknowledges chip-to-chip variation, but the abstract/conclusion claim that 'TMR effectively reduces SEEs' is not established for the LP path. Table 1 lists no uncertainties for nonzero counts, so the reader cannot judge the significance. Provide Poisson confidence intervals or p-values, and state the chip-152 result as a caveat in the abstract/conclusion if the claim is retained.","section":"Table 1 and §5: 'From Table 1...' paragraph"},{"comment":"Cross sections for nonzero event counts are quoted without uncertainties. For example, chip 154 Off ASLP = 32 events gives a ~18% relative Poisson error; chip 152 On ASLP = 13 events gives ~28%. Figure 4 plots cross-section points with no error bars, so the purported energy dependence—especially the 'dramatic rise from 60 MeV onward'—cannot be assessed. Add statistical error bars and, if possible, a fit to a Weibull or similar response curve.","section":"Table 1 and Fig. 4"}],"minor_comments":[{"comment":"Missing space in 'HCCStartoreduce'.","section":"Abstract"},{"comment":"'FromtheEoS' missing space.","section":"§3"},{"comment":"Formatting of '#fluence' and '·' is inconsistent; use a consistent notation.","section":"Eqs. (5.1), (5.2)"},{"comment":"Axis labels for (a) and (b) are garbled in the text; ensure the LaTeX renders correctly.","section":"Fig. 4"},{"comment":"Grammar: 'The flux were from...' should be 'The flux ranged from...'.","section":"§4"},{"comment":"The statement that the total fluence is equivalent to 48% of the HL-LHC lifetime is based on dose; please specify the dose-to-fluence conversion used.","section":"§4"},{"comment":"Effective monitoring intervals are given as ~2.1 h, but fluences and therefore cross sections differ per chip; clarify how individual fluences are computed if monitoring intervals differ.","section":"Table 1"},{"comment":"Journal names 'Jinst'/'JINST' are inconsistent; follow the journal style.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of JINST and the measurements are useful for the ITk community. The main technical issue is the unsupported 'conservative upper bound' on LP-event multiplicity, which should be corrected before publication. I also recommend adding statistical uncertainties to the cross sections and tempering the TMR-effectiveness claim for the LP path. These are fixable within the current manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the essentials. This paper reports the first proton SEE cross sections for HCCStar v1, including an energy scan from 20 to 80 MeV, and a rate estimate for LP-path data loss. The register TMR result is solid: two chips, CBF cross sections ~10^-13 cm2/bit with TMR on, actual SEEs orders of magnitude higher with TMR off. The paper is honest about chip 152, where TMR didn't reduce LP events (13 vs 15), a chip-to-chip variation they acknowledge. I trust the measurements.\n\nWhat's not solid is the claim that the LP loss fraction is O(10^-10) with a conservative upper bound. The stress-test note is correct: the LP line is a shared multidrop line to 6-11 ABCStars, so one SET can corrupt the readout transaction for more than one ABC. The paper's 'at most one affected transaction per event' is a lower bound on impact, not an upper bound. With eleven ABCs, 546 anomalies per year becomes ~6000, and the fraction is ~1.5e-9 — still negligible compared to 4e12 transactions, and the practical conclusion survives. But the paper should either fix the wording or bound the multiplicity, otherwise the headline number is not as conservative as advertised.\n\nTwo smaller issues. Table 1 has nonzero counts without uncertainties, and Fig. 4 has no error bars; for a paper whose main output is a rate, Poisson error bars should be there. Also, the 'worst-case' ASLP cross section is the 70 MeV point, but Fig. 4 suggests the cross section rises toward 80 MeV; I'd like a sentence confirming that the selected value is indeed the largest across the scan, or use the highest measurement.\n\nWho's this for? People in the ITk strip collaboration, and anyone doing ASIC radiation qualification. It's a useful, citable data point. It deserves a serious referee; the data are real and the soft spots are fixable. I'd send it to review with a request to address the multiplicity issue and add uncertainties.","headline":"Good, honest HCCStar SEE data with an overstated 'conservative upper bound'; the LP loss is more like O(10^-9), but the practical conclusion still holds.","tokens_in":6817,"tokens_out":4371,"would_cite":true,"duration_ms":38777,"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":"Proton-beam tests show the HCCStar chip's triple-modular redundancy cuts single-event upsets to a negligible data-loss rate.","keywords":["HCCStar","single event effects","triple modular redundancy","proton irradiation","ITk strip detector","HL-LHC","radiation tolerance","ASIC"],"falsifier":"Take a production batch of HCCStar chips, enable TMR, and expose them to a hadron spectrum that mimics the high-luminosity LHC (including higher-energy components). If the average LP-path anomaly rate exceeds the paper's conservative 546 per chip per year—or if most chips behave like the one that showed 13 LP events with TMR on—then the O(10^-10) data-loss fraction and the claim that TMR protects the LP path would fail.","tokens_in":5816,"feed_emoji":"☢️","tokens_out":5157,"duration_ms":47241,"temperature":0.7,"pith_summary":"The paper claims that the triple-modular-redundancy (TMR) protection built into the HCCStar readout chip works: under 20–80 MeV proton irradiation, corrected bit flips in registers occur at about 10^-13 cm^2 per proton per bit, and uncorrected single-event effects in the low-priority readout path are rare enough that the resulting data loss is only about 10^-10 of the normal 400 kHz readout rate. It estimates about 10 corrected bit flips per bit per year in the high-luminosity LHC environment. This matters because the HCCStar is the digital hub of each inner-tracker strip module, and the test provides quantitative evidence that the chip will not lose physics data during the upgrade's lifetime.","feed_headline":"Proton tests show HCCStar chip loses only 1 in 10 billion readouts","feed_subtitle":"Triple-modular redundancy keeps single-event upsets from corrupting inner-tracker strip data; expect ~10 corrected flips per bit per year.","key_machinery":"Triple Modular Redundancy (TMR): each critical register bit is stored in three copies and a voting circuit compares them, so a single flipped copy is corrected and logged as a corrected bit flip. The paper measures cross sections for these corrected flips (TMR on) and for actual single-event effects in registers and the LP control path (TMR off), then scales the worst-case LP cross section by an assumed hadron flux of about 10^7 particles per cm^2 per second over 10^7 seconds per year to get expected event rates.","core_discovery":"The central result is a set of proton-beam cross sections for the HCCStar ASIC. With triplication enabled, the register cross section is about 10^-13 cm^2 per proton per bit, recorded as corrected bit flips; with triplication disabled, the actual single-event-upset cross section rises to about 10^-12 cm^2 per proton per bit. In the low-priority (LP) control path, which relays trigger decisions, the worst measured cross section is 5.46×10^-12 cm^2 per proton, used as a conservative upper limit to estimate 546 LP-related readout anomalies per chip per year—a 10^-10 fraction of the 400 kHz LP readout. The paper also reports that no chip failed after a fluence equivalent to roughly half the expe","pith_inferences":["The chip-to-chip variation (one chip showed no LP events with TMR, the other showed as many as without TMR) implies that a larger sample is needed to confirm TMR effectiveness on the control path; a firmware-based parity or timeout check on LP responses could serve as a cheap safety net.","The same scaling method could be applied to the ABCStar's register and data paths, allowing a system-level single-event-effect budget for the whole module rather than per-chip.","A natural testable extension is to expose HCCStars to a mixed hadron field or higher proton energies to see whether the Weibull-like rise continues, which would sharpen the lifetime estimate before final production."],"forward_implications":["If the central estimate holds, LP-path single-event upsets will discard fewer than one in ten billion readout transactions, so physics data loss from the control path is negligible.","About 10 corrected bit flips per bit per year means the TMR circuits will be busy but manageable; register scrubbing or occasional resets will keep configuration intact.","No chip failures at roughly 24 MRad accumulated dose supports the HCCStar's design margin for the upgrade's radiation budget.","The energy-dependent cross-section rise above 60 MeV suggests that higher-energy hadrons may dominate the real environment, so continued monitoring at operating energies is prudent."],"fun_headline_variants":["HCCStar chip loses 1 in 10 billion readouts under protons","Proton beam test: HCCStar data loss at 10^-10 level","TMR protects HCCStar: 10^-10 readout loss in proton test","HCCStar ASIC survives protons: 10^-10 data loss, ~10 flips/yr","Radiation test: HCCStar readout loss just 10^-10"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The headline rates are computed by multiplying measured 20–80 MeV proton cross sections by an assumed hadron flux of ~10^7 cm^-2 s^-1 for 10^7 s/year in the high-luminosity LHC; if that flux or the beam's representativeness is wrong, the 10^-10 and O(10) numbers scale linearly, and the LP-protection conclusion rests mainly on one of the two chips tested.","fun_headline_variants_meta":{"raw":{"variants":["HCCStar chip loses 1 in 10 billion readouts under protons","Proton beam test: HCCStar data loss at 10^-10 level","TMR protects HCCStar: 10^-10 readout loss in proton test","HCCStar ASIC survives protons: 10^-10 data loss, ~10 flips/yr","Radiation test: HCCStar readout loss just 10^-10"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00015,"raw_usage":{"total_tokens":992,"prompt_tokens":664,"completion_tokens":328,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":236}},"tokens_in":408,"tokens_out":328,"duration_ms":3269,"temperature":1.0,"reasoning_tokens":236,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T08:24:14.758005+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a production batch of HCCStar chips, enable TMR, and expose them to a hadron spectrum that mimics the high-luminosity LHC (including higher-energy components). If the average LP-path anomaly rate exceeds the paper's conservative 546 per chip per year—or if most chips behave like the one that showed 13 LP events with TMR on—then the O(10^-10) data-loss fraction and the claim that TMR protects the LP path would fail.","supporting_citations":[],"review_version":1}