{"id":"48642495-9d7c-44ea-a688-da3416a35757","arxiv_id":"2507.11654","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A custom silicon-photonic recurrent spectral filter, used as a two-node optical pre-processor with a simple digital equalizer, keeps 32-GBd PAM-4 C-band IM/DD transmission below FEC thresholds up to 100 km.","lead":"Researchers built tiny silicon chips that break a light signal into frequency slices before detection, then recombine the slices digitally, and showed this lets four-level optical pulses travel up to 100 km of fiber without error correction failure. The approach could extend low-cost optical links used in data centers to longer reaches at lower power.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 100-km two-node ROSS claim relies on sequential single-PIC emulation that omits the 3-dB input-splitting loss and fixed-synchronization constraints of a real multi-node receiver.","rationale":"The paper is a serious, well-characterized demonstration of a custom ROSS filter PIC, with measured transfer functions and system BER data. The strongest claim is conditional on the fidelity of the sequential emulation of a two-node receiver. The most concrete physical shortfall is the missing input power split: a real two-node ROSS receiver divides the received signal between two filters, so each branch sees 3 dB less power than the full-power sequential captures used in the paper. Since the paper reports BER below the 400ZR C-FEC at 100 km but does not quantify the margin, a 3-dB OSNR penalty could plausibly push it above threshold. The reader's weakest assumption identified the same emulation issue; I partially agree, adding the splitter-loss mechanism as the sharpest technical test. This does not overturn the paper's value—the PIC itself works and the trend is consistent—but it does mean the 100-km reach claim should remain conditional until the emulation is shown to be equivalent to a real multi-node receiver.","tokens_in":5233,"tokens_out":6961,"duration_ms":87013,"concrete_test":"Place a 3-dB optical attenuator (or a 1x2 coupler with one output used) after the first EDFA and before the PIC, and repeat the 50-km and 100-km sequential node measurements and the two-node offline emulation using the same synchronization method. If the 100-km BER rises above the 400ZR C-FEC threshold, or if the 50-km BER crosses the KP4 FEC threshold, then the missing splitter loss is load-bearing and the reported two-node performance is not representative of a real multi-node receiver. As a secondary check, redo the emulation with a fixed, non-optimized relative delay derived from a common clock to test whether synchronization freedom inflated the result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a two-node ROSS receiver achieves BER below FEC at 100 km, but the two-node receiver was not built. As stated in the paper, the two nodes were emulated by sequentially measuring DSO traces from a single PIC at different heater voltages and then synchronizing the traces offline. A real two-node ROSS receiver would require a 1x2 optical splitter at the input, so each ROSS node would receive only half the optical power (3 dB less than in the sequential full-power measurements). The paper does not include this splitting penalty or show that the 100-km BER margin over the 400ZR C-FEC threshold exceeds the resulting OSNR degradation. The offline synchronization of traces may also permit a favorable relative delay that a simultaneous receiver with fixed electrical paths would not provide, especially because the paper does not describe the synchronization procedure. Unless the emulation is validated by including the splitting loss or by a true multi-node PIC, the headline reach claim is not demonstrated for a realistic device.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports the design, fabrication, and system test of a silicon-photonics recurrent optical spectral slicing (ROSS) filter intended for IM/DD receivers. Two filter variants with different MZI arm delays (long and short ΔT) were fabricated and characterized, and a two-node ROSS receiver was emulated by sequentially capturing single-PIC DSO traces at different heater voltages and synchronizing them offline. For a 32-GBd PAM-4 C-band link, the authors report BER below the KP4 FEC threshold at 50 km and below the 400ZR C-FEC threshold at 100 km, with up to two orders of magnitude improvement over a neural-network equalizer without optical preprocessing.","tokens_in":5446,"tokens_out":3404,"duration_ms":42163,"significance":"If the result holds, this is a meaningful advance for short-reach IM/DD links: a compact, purpose-designed integrated filter with low feedback loss provides optical frequency diversity that enables a simple FFE to extend the reach to 100 km. The paper provides measured transfer functions, BER results over distance, and a benchmark comparison including a programmable-PIC ROSS and a neural-network equalizer. The design builds transparently on the authors' prior theory, and the experimental benchmarking against external FEC thresholds is independent of fitting parameters. The main strength is the demonstrated potential of the custom filter, but the central claim relies on an emulated two-node receiver, which limits the strength of the headline conclusion.","major_comments":[{"comment":"The two-node ROSS receiver is emulated by sequentially measuring DSO traces from a single PIC at different heater voltages, as explicitly stated in the text. This emulation does not capture two essential behaviors of a real multi-node receiver: (i) the input signal must be split between the two nodes, incurring at least 3 dB of additional loss per node relative to the sequential full-power measurements, and (ii) the relative delay between the two nodes is fixed by hardware in a real receiver, whereas the offline synchronization of independently acquired traces may select a favorable relative delay. The paper should quantify the splitting penalty and repeat the key measurements with a real 1x2 splitter, or alternatively re-scope the abstract and conclusions to describe an emulated receiver. The synchronization procedure (e.g., cross-correlation, clock recovery) should also be described.","section":"System performance"},{"comment":"The BER versus distance curves in Fig. 4 have no error bars or repeated trials, and the best node pairs are selected from the 50-km heatmaps in Fig. 3. As a result, the robustness of the 100-km margin over the 400ZR C-FEC threshold to thermal drift, polarization re-alignment, or heater-voltage variability cannot be assessed. Please provide repeated BER measurements for the best node pair at each distance (including re-alignment) and report the spread; also explicitly state whether the same heater voltage pair was used at all distances or re-optimized per distance.","section":"Fig. 4 and associated text"},{"comment":"The abstract and conclusions state as a fact that a two-node ROSS receiver achieves BER below FEC at 100 km, but the two-node receiver was never physically built. Given the importance of the splitting-loss and synchronization issues, the claims should be re-scoped to “emulated two-node receiver” unless the emulation is validated by including a splitter in the power budget or by a true multi-node measurement.","section":"Abstract and Conclusions"}],"minor_comments":[{"comment":"The FFE description says “62 T-spaced-taps (31 symbols/filter)”; clarify whether this is 62 taps total or 62 taps per node, and how the two node outputs are combined.","section":"System performance"},{"comment":"The labels “long” and “short” are ambiguous in the figure and caption; add explicit legend entries or a statement identifying which curve corresponds to which ΔT design.","section":"Fig. 4"},{"comment":"The reference path is described as “bypassing an EDFA and the PIC”, but the setup in Fig. 2 shows a pre-amplified receiver; clarify whether the reference measurement uses the same EDFA and BPF without the PIC or a completely different photodetection chain.","section":"Experimental setup"},{"comment":"The statement that “the best node positions are consistent over transmission distance” is not directly supported because the heatmaps in Fig. 3 are shown only at 50 km; either provide supporting data or soften the claim.","section":"System performance"}],"recommendation":"major_revision","confidential_remarks":"This is a two-page conference-style manuscript. For a journal review, the missing details on synchronization, splitting loss, and repeatability are critical to the main claim. The authors are transparent about the emulation, which is commendable, but the paper should either validate the emulation or present the result as a proof-of-principle emulation study. The best-node-pair selection procedure also needs clarification to rule out overfitting to the 50-km heatmaps."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is not the programmable-PIC ROSS demo repackaged. The authors designed and fabricated dedicated silicon photonic filters with feedback loss below 0.1 dB and free choice of MZI delay, and the transfer-function measurements show sharp, tunable, high-extinction spectral slices that the programmable platform couldn't give them. Measured BER at 50 km with two nodes is about two orders of magnitude below the no-preprocessing case and well under KP4 FEC; the trend with distance is consistent, and the best node pairs stay the same across distances, which is a good sign that the effect is real. The comparison against their own prior programmable-PIC result is useful and credibly isolates the benefit of lower loss and flexible delay. I take the result as a genuine engineering advance: a practical route to make ROSS receivers viable.\n\nThe soft spot is exactly where the reader and the stress-test note point. The two-node receiver was never built. They took sequential DSO captures from a single PIC at different heater voltages and synchronized them offline. The paper says this explicitly, which is honest, but it has real consequences. A physical two-node receiver needs an input splitter, so each node gets 3 dB less power than in the sequential full-power measurement. That OSNR penalty is not in the reported BER. Also, offline synchronization can pick a favorable timing offset that a real receiver with fixed electrical paths won't get. The 100-km BER is below the 400ZR C-FEC threshold, but without knowing the margin, I can't tell whether it survives the 3-dB hit. This should be flagged as a load-bearing caveat, not a cosmetic one.\n\nOther complaints are minor and standard for this kind of proof-of-concept: no error bars, best voltage pairs chosen post hoc from the 50-km heatmaps, one launch power, one fiber length per point. None of these undercut the fabrication story. The self-citation to their own theory and prior experiments is not a problem because the BER benchmark against external FEC thresholds is independent.\n\nWho it's for: people working on short-reach IM/DD, silicon photonics for communications, and optical pre-processing. It deserves a serious referee. A good referee should push for a true two-node measurement or at least a quantified account of the splitting loss and synchronization before the headline reach claim is treated as demonstrated. As is, I'd call it a solid proof-of-concept with an overstated headline.","headline":"Custom low-loss ROSS filter is a genuine fabrication advance, but the 100-km two-node claim rests on an emulation that omits the 3-dB input-splitting penalty.","tokens_in":6035,"tokens_out":2341,"would_cite":true,"duration_ms":29073,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.79.Sz","42.82.-m"],"model":"deepseek-v4-flash","headline":"The paper claims that a purpose-built silicon photonic spectral-slicing receiver can hold 32-GBd PAM-4 C-band transmission below forward-error-correction limits at 50 and 100 km, using only a simple linear equalizer after the optical…","keywords":["recurrent optical spectral slicing","silicon photonics","IM/DD transmission","PAM-4","C-band","optical equalization","feed-forward equalizer","power fading"],"falsifier":"Build or operate a chip with two physical ROSS nodes on the same die and measure 32-GBd PAM-4 at 50 and 100 km with both nodes running simultaneously; if the bit-error ratio at 100 km rises above the 400ZR C-FEC threshold or the 50-km result rises above KP4, the central claim fails. A cheaper check is to re-run the sequential-capture emulation with a different synchronization method and see whether the reported BER depends strongly on how the two traces are aligned.","tokens_in":5024,"feed_emoji":"📡","tokens_out":9378,"duration_ms":104273,"temperature":0.7,"pith_summary":"This paper reports a silicon-photonics receiver front end that recovers dispersion-limited reach in short-haul intensity-modulation/direct-detection (IM/DD) links without coherent detection. The device, a recurrent optical spectral slicer (ROSS), is an asymmetric Mach-Zehnder interferometer with delayed feedback whose transfer function can be tuned by a thermo-optic heater to carve the received spectrum into non-overlapping slices. The paper claims that combining two such nodes followed by a simple linear feed-forward equalizer keeps a 32-GBd PAM-4 C-band signal below the KP4 hard-decision FEC threshold ($2.25\\times10^{-4}$) at 50 km and below the 400ZR C-FEC threshold ($1.25\\times10^{-2}$) at 100 km. In the reported measurements this is roughly two orders of magnitude lower bit-error ratio than a neural-network equalizer applied to the same signal without optical pre-processing. The reason this matters is that it extends the reach of low-cost direct-detection links by moving part of the equalization into the optical domain.","feed_headline":"Spectral slicer pushes 32-GBd PAM-4 to 100 km in C-band","feed_subtitle":"Two tuned silicon filters beat neural-net equalization and carry IM/DD across 100 km in C-band.","key_machinery":"The load-bearing element is the ROSS node, implemented here as an asymmetric Mach-Zehnder interferometer with a feedback loop from output to input containing about 19 ps of delay and a thermo-optic phase shifter. Feedback makes the filter recurrent (infinite impulse response), producing sharp spectral features with up to 15 dB extinction; tuning the heater voltage selects different spectral slices from the same device. On the receiver side, two independently tuned nodes' detected outputs are recombined by a 62-tap feed-forward equalizer spaced one symbol period apart, and the frequency diversity from non-overlapping slices is what suppresses the power-fading penalty.","core_discovery":"On the paper's own terms, the central discovery is that a purpose-built recurrent spectral slicer, rather than a general-purpose programmable photonic circuit, makes the ROSS receiver practical. The fabricated filters show insertion losses of 9.8–10.4 dB, feedback-loop losses below 0.1 dB (versus more than 1 dB in the previous programmable implementation), and a rich family of sharp, high-extinction transfer functions that shift and reshape as the heater voltage changes. With two such nodes tuned to non-overlapping spectral slices, the paper reports BER below the KP4 threshold ($2.25\\times10^{-4}$) at 50 km and below the 400ZR C-FEC threshold ($1.25\\times10^{-2}$) at 100 km for 32-GBd PAM-4, with the short-delay (≈23 ps) design performing slightly better than the long-delay (≈44 ps) design.","pith_inferences":["A natural next step the authors leave implicit is to put two physical ROSS nodes on one chip and operate them simultaneously; their two-node receiver was emulated by synchronizing sequentially recorded traces from a single chip, so an integrated multi-node version would be the direct test of the 100-km claim.","If the frequency-diversity mechanism is the real source of gain, adding more nodes should improve reach or support higher baud rates, since more non-overlapping slices should further mitigate dispersion-induced fading.","The same recurrent-slicing front end could be applied to other direct-detection formats or access-network links where chromatic-dispersion power fading, rather than noise, is the bottleneck."],"forward_implications":["A two-node ROSS receiver can hold 32-GBd PAM-4 C-band transmission below the 400ZR C-FEC threshold at 100 km, extending IM/DD reach without coherent detection.","Because the digital stage after slicing is a linear feed-forward equalizer, the scheme avoids complex nonlinear or neural-network equalizers while reportedly beating them by about two orders of magnitude in BER.","The same chip design realizes different ROSS nodes by voltage tuning, so a single fabrication can serve multiple spectral-slice configurations.","The comparison with the programmable-PIC implementation shows that reducing feedback-loop loss and insertion loss is what unlocks the added order-of-magnitude BER gain."],"supporting_citations":[{"why":"This prior analysis defines the ROSS technique and predicts that non-overlapping spectral slices mitigate power fading, forming the basis for the node design used here.","marker":"[7]"},{"why":"This previous experimental ROSS demonstration on a programmable PIC supplies the baseline performance and the loss and granularity limitations that the custom design overcomes.","marker":"[8]"},{"why":"This reference documents the programmable photonic platform's unit-cell delay and coupling losses, which motivate the need for a purpose-built filter.","marker":"[10]"},{"why":"This source provides the KP4 hard-decision FEC threshold of $2.25\\times10^{-4}$ used to define the 50-km success criterion.","marker":"[11]"},{"why":"This implementation agreement defines the 400ZR C-FEC threshold of $1.25\\times10^{-2}$ used to claim 100-km operation.","marker":"[12]"}],"fun_headline_variants":["Recurrent spectral slicer hits 100-km C-band PAM-4","Low-loss recurrent slicer enables 100-km PAM-4 in C-band","Integrated slicer pushes IM/DD to 100 km with low loss","Silicon recurrent filter reaches 100-km C-band FEC limit","Recurrent slicer equalizes 32-GBd PAM-4 over 100 km"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire 100-km claim rests on treating two separately recorded, time-synchronized traces from the same chip as if they came from a real two-node receiver; if sequential captures are not equivalent to simultaneous operation of two physical ROSS nodes, the reach result is not yet demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Recurrent spectral slicer hits 100-km C-band PAM-4","Low-loss recurrent slicer enables 100-km PAM-4 in C-band","Integrated slicer pushes IM/DD to 100 km with low loss","Silicon recurrent filter reaches 100-km C-band FEC limit","Recurrent slicer equalizes 32-GBd PAM-4 over 100 km"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00042,"raw_usage":{"total_tokens":2075,"prompt_tokens":777,"completion_tokens":1298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":393,"completion_tokens_details":{"reasoning_tokens":1192}},"tokens_in":393,"tokens_out":1298,"duration_ms":10759,"temperature":1.0,"reasoning_tokens":1192,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:04:38.790927+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build or operate a chip with two physical ROSS nodes on the same die and measure 32-GBd PAM-4 at 50 and 100 km with both nodes running simultaneously; if the bit-error ratio at 100 km rises above the 400ZR C-FEC threshold or the 50-km result rises above KP4, the central claim fails. A cheaper check is to re-run the sequential-capture emulation with a different synchronization method and see whether the reported BER depends strongly on how the two traces are aligned.","supporting_citations":[{"cited_title":"Experimental Investigation of a Recurrent Optical Spectrum Slicing Receiver for Intensity Modulation/Direct Detection sys- tems using Programmable Photonics","cited_arxiv_id":null,"evidence_quote":"This previous experimental ROSS demonstration on a programmable PIC supplies the baseline performance and the loss and granularity limitations that the custom design overcomes."},{"cited_title":"Information-theoretic tools for optical communications engineers","cited_arxiv_id":null,"evidence_quote":"This source provides the KP4 hard-decision FEC threshold of $2.25\\times10^{-4}$ used to define the 50-km success criterion."},{"cited_title":"https://www.oiforum.com/wp-content/up- loads/OIF-400ZR-03.0.pdf","cited_arxiv_id":null,"evidence_quote":"This implementation agreement defines the 400ZR C-FEC threshold of $1.25\\times10^{-2}$ used to claim 100-km operation."}],"review_version":1}