Deposon: An Auditable, Conservation-Guaranteed, Game-Theoretically Tested Scattering Layer over LLM Reasoning Paths
cs.AI · cs.LG
Quick verdict
Deposon adds an auditable T+R+A=1 ledger to LLM reasoning paths; rigorously honest, but the conservation law is definitional and beats no real baseline.
Summary
The paper introduces Deposon, a 'scattering layer' over LLM reasoning paths: each node of an LLM-generated concept-decomposition graph is bound to a two-parameter state, and each candidate path's energy splits into transmission, reflection, and irreversible dissipation, with T+R+A=1 holding by construction (audited at 2.2×10⁻¹⁶, machine epsilon). Evidence is organized into three strength tiers and governed by pre-registered kill protocols. Positive results: potential trajectories are monotone on 22/22 graphs; the median non-potential residual is 1.594×10⁻²⁹; the median empirical coordination ratio is ECR=1.333. Equally prominent negative results: the synthetic 100%-versus-7%/10% gain depends on adversarially weighted decoy edges and free type labels; on real benchmarks the layer ties a six-keyword rule filter (GSM8K 0.85 vs 0.87) and trails chain-of-thought (85.0% vs 97.0%, McNemar p=4.9×10⁻⁴); convex fusion never improves; and all three formalized dynamical-equivalence propositions are falsified, downgrading the potential-game reading to approximate (cyclic median residual 0.669). The authors relocate the value proposition to machine verifiability, though the audited identity is true by definition. Main limitations: the conservation guarantee is a normalization tautology, real-benchmark samples are small (n=100/99), graphs come from a single LLM vendor with possible same-source contamination, and no experiment demonstrates the ledger catching a real fault.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A constructive three-channel scattering definition makes T+R+A=1 hold for arbitrary parameters, with a per-path audit residual of 2.2×10⁻¹⁶ that any third party can recheck in double precision.
- On real benchmarks (GSM8K, StrategyQA) the scattering layer is statistically indistinguishable from a trivial six-keyword rule filter, so its only differential value is machine verifiability (E9.5).
- Convex-combination fusion with a semantic prior never exceeds either single arm on any measured λ setting; any fusion gain must be nonlinear (§2.5).
- The reverse dynamics exhibits monotone potential trajectories on 22/22 graphs (GT-5b) and near-zero non-potential residuals (GT-6), supporting only a consistency-level potential-game reading, with median ECR=1.333 (GT-4).
- All three formalized dynamical-equivalence propositions (P1a/P1b/T-P1c) are falsified under pre-registered kill protocols over 61 graphs, 338 tasks, and 6,760 states, and the potential-game claim is downgraded to approximate on cyclic graphs (P2, median residual 0.669).
Strengths
Pre-registration discipline executed through the kills
Verdict functions were frozen as pure functions before any run with SHA-256 anchors (e.g., GT_FORMALIZATION_v1.md = aeefb8ef6972, §3.3), and the kill rules were applied against the paper's own headline claims: H-A1 (field beats random) died on its disjunctive kill line (4 reversed graphs, 16+/4−/2, p=0.0118), and P1a/P1b/T-P1c were all falsified over 61 graphs, 338 tasks, 6,760 states (§3.3).
Full number traceability and mechanical verdicts
Appendix A maps every cited number to a specific field of a frozen JSON (e.g., the 2.220446049250313×10⁻¹⁶ audit residual to deposon_v19_benchmark_fixes.json → physics_audit), verdicts are evaluated mechanically rather than argued, and a same-seed rerun of T-P1c is claimed byte-identical with an 11-test-locked verdict function.
Honest attribution controls that dismantle the paper's own apparent gains
Careful statistical hygiene
Weaknesses
The headline 'conservation guarantee' is a definitional tautology
No demonstrated accuracy or safety benefit; residual value claim unvalidated
Synthetic headline rests on a constructed strawman baseline
Game-theoretic positives are weak after the paper's own kill-tests
External validity and provenance limits
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Deposon (arXiv 2609.09001, cs.AI/cs.LG) proposes a 'scattering layer' over LLM reasoning paths. Each node of an LLM-generated concept-decomposition graph is bound to a two-parameter state (g_couple, g_aether); each candidate path's energy splits across transmission, reflection, and irreversible dissipation with T+R+A=1 holding by construction, audited at 2.2×10⁻¹⁶ (§2.2). Contributions C1–C4 pair a conservation/auditability claim with an unusual volume of pre-registered negative results: a tie with a six-keyword rule filter on real benchmarks (E9.5, §2.4), convex-fusion dilution (§2.5), and the falsification of all three formalized dynamical-equivalence propositions (§3.3). The author explicitly relocates the value proposition: the layer is 'not more accurate, but it keeps the books' (§5).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Deposon scattering layer?
It is a bookkeeping mechanism for LLM reasoning: each node of an LLM-generated concept graph gets a two-parameter state, and each candidate reasoning path has its 'energy' split among transmission, reflection, and irreversible dissipation. The three shares always sum to one by construction, so any third party can recompute the per-path energy ledger with ordinary double-precision arithmetic.
Does Deposon improve LLM reasoning accuracy?
No. In the paper's own pre-registered experiments it trails plain chain-of-thought on a GSM8K subset (85.0% vs 97.0%, McNemar p=4.9×10⁻⁴) and is statistically indistinguishable from a trivial six-keyword rule filter on both GSM8K (0.85 vs 0.87) and StrategyQA (0.899 vs 0.899). The authors explicitly claim its value is machine-checkable audit records, not accuracy.
Is the T+R+A=1 conservation law a real physics result?
No. The transmission, reflection, and dissipation weights are defined as shares of a normalization constant (Λ=1+g_eff+g_aether), so they sum to one identically — the 2.2×10⁻¹⁶ audit residual merely confirms double-precision arithmetic. The physics vocabulary (scattering, Feshbach resonance) is inspirational; the paper itself states the weights are constructive definitions, not consequences of an S-matrix, and the resonance channel is dormant in all main experiments.
What did the paper's pre-registered kill protocols find?
They falsified all three formalized claims that the reverse dynamics implements best response (P1a/P1b/T-P1c, tested over 61 small graphs, 338 tasks, 6,760 states), killed the headline field-beats-random claim (H-A1, 4 reversed graphs), and downgraded the potential-game reading to approximate on cyclic graphs (median residual 0.669). What survived: potential-trajectory monotonicity on the original 22-graph protocol and consistency-level (not formal) evidence only.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure images are not present in the extraction (only captions of Figures 1–5), so visual quality and plotted content could not be assessed.
- The repository (github.com/zeroandcat/Deposon) and its frozen JSONs were not accessed; reproducibility statements reflect the paper's claims, not independent verification.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.