{"id":"edb5c2eb-da24-4dee-8d56-fc7cf9b65aad","arxiv_id":"2607.08907","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":7,"one_line_summary":"High herding fraction and strength produce a null-verified, rule-robust liquidity-stress crossover in a continuous double-auction book, with a large self-reinforcing reflexive component under price-momentum herding.","lead":"A carefully controlled agent-based order-book study finds a genuine liquidity dry-up when many strong herders chase price momentum: the book goes one-sided about a third of the time. The effect is a smooth crossover, not a crash, and is driven by self-reinforcing feedback rather than a parameter artefact.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper follows its stated phase-diagram discipline carefully: full φ×κ map, scrambled-sign null, band and numerical artefact checks, rule and horizon robustness, fine onset scan, and amplitude-matched multi-comparator mechanism decomposition. The strongest claim—a genuine, null-verified, imitation-specific liquidity-stress crossover that is smooth rather than a Dark Corner, with a large comparator-robust reflexive component under price-momentum herding—is internally well supported by the reported tables and figures. The reader's identification of the scrambled-sign null as the weakest causal premise is accurate and matches §3.4/§5.1/§6.1, but the paper already mitigates that premise with shadow/replay/telegraph controls that keep the momentum result robust. No internal inconsistency, numerical red flag, or unacknowledged confound rises to the level of overturning the claim. Therefore the ACCEPT verdict with high confidence should stand unchanged; the proposed phase-randomized null is a worthwhile verification step rather than a required fix.","tokens_in":21414,"tokens_out":562,"duration_ms":6027,"concrete_test":"Re-run the corner (φ=0.9,κ=1.0) momentum cell with a phase-randomized null that preserves the real herd signal's power spectrum and run-length distribution but destroys directional sign correlation (e.g., randomize Fourier phases of the mid-return series before sign extraction). If ϕ∅ remains ≈0 under this stronger null while the real closed-loop value stays ≈0.34, residual non-sign structure is ruled out; if ϕ∅ rises substantially, the imitation-specific claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest-assumption concern about residual non-sign structure in the scrambled-sign null is real but already substantially addressed by the paper's own multi-comparator design and does not overturn the central claim. The null (§3.4) randomizes only buy/sell signs while preserving order-type mix and φ-driven ZI dilution; ϕ∅=0 across all 42 null cells (§5.1). Residual timing/burst/mid-anchoring confounds are further isolated by the amplitude-matched shadow and replay legs and the synthetic telegraph continuum (§5.6–5.8), which show that momentum's reflexive excess (+0.29) survives every open-loop comparator while OFI's does not. The imitation-specific, null-verified liquidity-stress crossover and the signed momentum reflexivity therefore remain supported as stated. Scope limits (two rules, CDA only, higher-moment residual for OFI) bound generalization rather than falsify the reported feature.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper applies Bouchaud's phase-diagram discipline to a continuous double-auction order-book ABM with fundamental-anchored zero-intelligence liquidity and a mid-anchored chartist herding layer controlled by herd fraction φ and strength κ. A 7×6 grid (336 runs with scrambled-sign nulls) locates an emergent liquidity-stress crossover: the one-sided-book fraction ϕ∅ rises to ≈0.34 at (φ,κ)=(0.9,1.0), is identically zero across all 42 scrambled cells, and forms a smooth ramp rather than a discontinuous Dark Corner. The dry-up is rule-robust under an order-flow-imbalance signal, horizon-robust across a 16× momentum-window range, and has a monotone onset boundary φ*(κ)={0.55,0.45,0.36}. At matched directional-bias amplitude (mean |p_buy−0.5|≈0.269), price-momentum herding carries a large comparator-robust reflexive component (+0.29), while the OFI component is ≈0 and comparator-dependent. RMS-mispricing growth is identified as a placement artefact (largest at κ=0), and a companion two-market analysis finds no directional contagion across a signal-only herding link.","tokens_in":21734,"tokens_out":1348,"duration_ms":10971,"significance":"If the result holds, this is a substantial methodological and substantive contribution to market-microstructure ABMs. It is, to the authors' knowledge and mine, the first full application of Bouchaud's phase-diagram / artefact-ruling-out programme to an order-book vehicle with herding and liquidity as control axes. The scrambled-sign null, band and numerical-stability controls, rule- and horizon-robustness battery, fine onset scans, and amplitude-matched multi-comparator open-loop decomposition (shadow, replay, synthetic telegraph) are unusually thorough and reproducible; code and data are public. The signed, rule-dependent reflexive mechanism and the no-contagion corollary are falsifiable predictions that distinguish the work from exogenous-shock flash-crash calibrations and from sharp Dark-Corner claims in the macro lineage. The careful refusal to over-claim a tipping line is itself a strength.","major_comments":[{"comment":"§5.6 / Table 7: The OFI reflexive component is reported as ≈0 and comparator-dependent, yet the headline contrast with momentum's +0.29 is load-bearing for the 'signed, rule-dependent' mechanism claim. The baseline-flow shadow (0.418) is a ~2–3× outlier among matched-amplitude drives and rests on only 3 independent driving series; the paper itself notes that seed-level SEM understates uncertainty and that clustering widens the interval. A higher-moment-matched open-loop comparator (or a larger driving-series ensemble with clustered errors) is needed before the asymmetry can be stated as firmly as in the abstract and §6.2. Without it the OFI side remains under-resolved rather than cleanly unsigned.","section":null},{"comment":"§3.4 and §5.1: The scrambled-sign null is the central causal instrument for the 'imitation-specific' claim. It randomises only buy/sell signs while preserving order-type mix and φ-driven ZI dilution. Residual non-sign structure of real herding (timing, burstiness, mid-anchoring interaction with the closed-loop signal) is partially isolated by the later shadow/replay/telegraph legs, but those legs are run only at the corner and only after amplitude matching. A brief explicit check that null and real share comparable higher-moment order-arrival statistics (or a null that also scrambles arrival timing) would close the residual channel the authors themselves flag as the weakest premise.","section":null}],"minor_comments":[{"comment":"§3.5: The continuous-reference convention (carry forward last traded price when mid is undefined) is essential for fair real-vs-null comparison; it should be stated once in the main text with a short sensitivity note, not only in the methods paragraph.","section":null},{"comment":"Figure 1 / Table 1: The tabulated zeros for low-φ cells are roundings; the fine-scan nonzero tail (ϕ∅=0.002 at (0.30,1.0)) is mentioned in text but not shown. A footnote or inset would prevent a reader from misreading a hard threshold.","section":null},{"comment":"§5.7: The two-market no-contagion result is a clean corollary; the coupling form and the intensity-modulation variant are described briefly. A one-sentence statement of the exact signal-injection equation would aid replication.","section":null},{"comment":"Notation: the order parameter is written both ϕ∅ and ϕ∅ (and occasionally ϕ∅) inconsistently across abstract, tables and body; standardise.","section":null},{"comment":"Limitations §6.4 correctly note that ϕ∅ is still rising at the corner and that φ\to1 is deliberately avoided; a single sentence in the conclusion reminding the reader that the map does not reach a plateau would keep the scope clear.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is unusually careful and self-critical for the ABM literature; the public code repository is a genuine plus. The two major points are fixable within the existing experimental design and do not threaten the central null-verified crossover or the momentum reflexivity result. Fit for a serious q-fin / market-microstructure journal is good; the companion two-market paper should be cross-referenced clearly if it is under simultaneous review."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is that this is careful computational microstructure work that actually follows Bouchaud’s phase-diagram discipline instead of just citing it. On a 7×6 grid with scrambled-sign nulls, Novotny finds a genuine imitation-specific liquidity-stress crossover: one-sided-book fraction ϕ∅ rises to ~0.34 at (φ,κ)=(0.9,1.0), is zero across all 42 null cells, is rule- and horizon-robust, and has a clean monotone onset boundary. He correctly refuses to call it a discontinuous Dark Corner and correctly demotes the RMS-mispricing gradient to a placement artefact (largest at κ=0).\n\nWhat is new is the combination: full 2D map of herding fraction × strength on a continuous double-auction book, explicit null, band and numerical artefact checks, fine onset scan, and an amplitude-matched multi-comparator open-loop decomposition. The mechanism result is the interesting part. At matched directional bias (~0.269), price-momentum herding carries a large, comparator-robust reflexive component (+0.29; buying begets buying), while the OFI rule’s component is ~0 and comparator-dependent. The no-contagion corollary for signal-only cross-market links follows cleanly from that self-referential reading. Code and data are public.\n\nSoft spots are real but scoped. The scrambled-sign null isolates directional correlation well, and residual burst/timing confounds are further checked by shadow, replay, and telegraph continuum; the stress-test concern does not overturn the central claim. OFI’s higher-moment residual is acknowledged, not papered over. Scope is narrow (CDA only, two rules, free ZI parameters), and ϕ∅ is still rising at the corner, so the map is incomplete beyond φ=0.9. None of that falsifies the reported crossover or the momentum reflexivity.\n\nThis is for people who care about ABM integrity, liquidity dry-ups, and reflexivity measurement. Math and simulation design look solid; citations are fair to the Mark-0 lineage, Fosset et al., and Gao et al. I would send it to peer review and would cite the phase-diagram result and the matched-amplitude decomposition. Worth engaging.","headline":"Solid, disciplined phase-diagram work on order-book herding: a null-verified liquidity dry-up crossover with a signed reflexive mechanism under momentum, not a Dark Corner.","tokens_in":22325,"tokens_out":564,"would_cite":true,"duration_ms":6057,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Herding produces a smooth, null-verified liquidity dry-up in an order book, driven by a reflexive price-momentum loop, not a discontinuous crash.","keywords":["agent-based modelling","market microstructure","herding","liquidity crises","reflexivity","order book","phase diagram"],"falsifier":"Re-run the high-herding corner with a control that preserves higher-moment burst structure of the herd signal while still destroying directional correlation; if the one-sided-book fraction remains high under that control, the imitation-specific claim fails.","tokens_in":22238,"feed_emoji":"📉","tokens_out":767,"duration_ms":6763,"temperature":0.7,"pith_summary":"Agent-based market models often throw up apparent instabilities that may be nothing more than parameter quirks. This paper maps the full two-dimensional phase diagram of a continuous double-auction order book whose agents mix fundamental-anchored zero-intelligence liquidity with mid-anchored chartist herders, controlled by herd fraction and herd strength. It finds a genuine liquidity-stress crossover: when both controls are high the book is one-sided roughly a third of the time. The dry-up vanishes completely under a scrambled-sign null that keeps the same order mix but destroys directional correlation, so the effect is imitation-specific rather than a mere thinning of liquidity. The same corner stress reappears under an order-flow-imbalance herding rule, survives a sixteen-fold change of momentum horizon, and has a clean monotone onset boundary that falls as herding strength rises. Mechanism decomposition at matched directional-bias amplitude shows that price-momentum herding carries a large, comparator-robust reflexive component—buying begets buying—while the order-flow rule does not. The result is a characterised liquidity phenomenon, not a price bubble or a sharp tipping line, and a companion two-market test finds no directional contagion across a pure signal link.","feed_headline":"Herding dries the order book one-third of the time","feed_subtitle":"A phase-diagram map finds a smooth, null-verified liquidity crossover driven by reflexive momentum","key_machinery":"The scrambled-sign null together with amplitude-matched open-loop shadow/replay decomposition: the null randomises only herd buy/sell signs while preserving order-type mix and zero-intelligence dilution, isolating directional imitation; the matched-amplitude comparison then measures how much extra dry-up the closed loop produces over equally forceful open-loop drives, revealing a robust self-reinforcing component under price momentum.","core_discovery":"A null-verified, imitation-specific liquidity-stress crossover sits in the high-herding corner of the herd-fraction × herd-strength plane. The fraction of one-sided-book events rises to about 0.34 at the corner, is identically zero across the entire scrambled-sign null plane, is rule- and horizon-robust, forms a smooth crossover rather than a discontinuous Dark Corner, and under price-momentum herding is amplified by a large comparator-robust reflexive component of roughly +0.29 at matched directional-bias amplitude.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Herding flips books one-sided a third of the time","Null-verified liquidity dry-up in high-herding corner","Reflexive momentum herding dries books at 0.34 rate","Smooth herding crossover makes one-sided books rise","Phase map shows herding stress, zero in scrambled null"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The scrambled-sign null is assumed to isolate directional imitation as the sole cause of the dry-up, so that any real-minus-null difference is truly due to correlated herding rather than residual timing or burst properties of the closed-loop signal.","fun_headline_variants_meta":{"raw":{"variants":["Herding flips books one-sided a third of the time","Null-verified liquidity dry-up in high-herding corner","Reflexive momentum herding dries books at 0.34 rate","Smooth herding crossover makes one-sided books rise","Phase map shows herding stress, zero in scrambled null"]},"model":"grok-4.5","effort":"low","cost_usd":0.004532,"raw_usage":{"total_tokens":1458,"prompt_tokens":958,"num_sources_used":0,"completion_tokens":88,"cost_in_usd_ticks":45320000,"prompt_tokens_details":{"text_tokens":958,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":412,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":958,"tokens_out":88,"duration_ms":4442,"temperature":1.0,"reasoning_tokens":412,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T05:53:22.179502+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the high-herding corner with a control that preserves higher-moment burst structure of the herd signal while still destroying directional correlation; if the one-sided-book fraction remains high under that control, the imitation-specific claim fails.","supporting_citations":[],"review_version":1}