Every manuscript on the journal, from preprint to open access. Abstracts are public; full text is for Fellows, Patrons, and open papers.
JoC-2026-008PreprintPosted on clearing Commons scrutiny — round-one review is open.BreakthroughA breakthrough — a new result others can build upon.
Shashank Taxak · AI & Machine Learning
Where should a figure go? Most layout engines either drop figures at the end of a document, spread them mechanically, or anchor them to the first place the text says "Figure 1". We ask whether the *document's own statistics* can decide, in closed form, without any training. In the spirit of Reverse Synthetic Networks — where the data synthesises the model directly — we synthesise each figure's position from three statistics of the manuscript itself: the TF-IDF affinity between the figure's caption and each block of running text, the proximity to any explicit lexical reference, and a small set of layout priors. There is no loss function, no gradient, and no learned parameters. On a controlled 600-document corpus (1,903 figures) drawn from the Aitele Commons' research domains, the closed-form placer reaches 92.5% top-1 placement accuracy and a mean position error of 0.18 sections, against 60.9% for a reference-only baseline and ~22% for content-blind baselines. The decisive result is on unreferenced figures — those the prose never names: here the closed-form affinity recovers 85.1% of placements, where the reference-only method collapses to 21.7% (essentially chance). Closed-form caption–text affinity, in other words, supplies the reference the author forgot to write. The same machinery groups related figures into multi-panel composites (pair-F1 0.99, 97.7% of groups recovered exactly), assigns column-spanning width for a balanced two-column page, and — extended to in-text citations — places each reference at the sentence it supports with 90.6% accuracy (vs ~15% for position-based baselines) and splices the grouped markers into the prose. The library ships with a 34-test suite (unit, edge-case, randomised-invariant, determinism, and a 1,200-block stress test). We report the method, the experiment, and its limitations honestly: the corpus is topically separated, the metrics are layout proxies rather than human aesthetic judgement, and page-overflow is out of scope.
JoC-2026-009PreprintPosted on clearing Commons scrutiny — round-one review is open.BreakthroughA breakthrough — a new result others can build upon.
Shashank Taxak · AI & Machine Learning
Every article in the Journal of the Commons carries, in its record, a short list of related work drawn from the entire catalogue — latest and archived. We build that list in closed form, in the spirit of Reverse Synthetic Networks (RSN): the corpus's own statistics synthesise each article's neighbours directly, with no trained model, no embedding service, no gradient, and no per-user state. Relatedness is a TF-IDF cosine over each article's title, abstract and keywords, refined by a few structural signals (shared field, shared keywords, shared authors, explicit implements/implementation links). On a controlled benchmark of 30 random corpora with known topical ground truth, the engine reaches nDCG@10 = 0.964 ± 0.005 and Precision@5 = 0.994 ± 0.004 (mean ± sd), against 0.500 for a field-matching baseline and ≈0.08 for content-blind baselines. Crucially, a relevance gate derived from the corpus's own similarity distribution makes the engine return nothing for an article that has no genuine relatives: across 30 corpora seeded with one-of-a-kind articles, isolated articles are correctly suppressed 100% of the time (false-positive rate 0.000), while articles that do have relatives are served 100% of the time at 99.4% suggestion precision. Ranking 5,000 articles takes 177 ms on one CPU core.
JoC-2026-010PreprintPosted on clearing Commons scrutiny — round-one review is open.BreakthroughA breakthrough — a new result others can build upon.
Shashank Taxak · AI & Machine Learning
When a manuscript enters peer review, an editor must find reviewers who actually know the subject, who have no conflict of interest, and who are not already overloaded. We do this in closed form, in the spirit of Reverse Synthetic Networks (RSN): the corpus's own statistics synthesise the match. A reviewer's expertise — their field, declared keywords, and the text of their own authored work — is a TF-IDF vector; the manuscript is another; fit is their cosine, refined by field/keyword overlap and balanced by current review load. No trained model, no gradient, no embedding service, no per-reviewer profile learning. On a controlled benchmark of 30 reviewer pools with known expertise ground truth, the engine puts a correct expert in the top slot 99.8% of the time (nDCG@5 = 0.980 ± 0.008, MRR 0.999), against 0.52 for a field-matching baseline and 0.08 for random. Three properties make it safe to act on automatically: (1) conflicts are detected automatically from the collaboration record — the manuscript's authors, editor-declared conflicts, anyone who has co-authored with an author, and anyone sharing a lab with the author team are excluded with 100% reliability (while a clean expert is still found 100% of the time); (2) a fit gate returns an empty list for a manuscript with no in-Commons expert — the explicit signal to source reviewers externally — suppressing 100% of no-expert cases at 0 false positives; and (3) a load signal spreads assignments (peak load 7.4 → 6.3, dispersion 0.48 → 0.30) with negligible loss of fit. Matching against a 5,000-reviewer pool takes 131 ms on one CPU core. By design the engine serves the Commons — its Fellows and Patrons; for external reviewers it returns nothing, and the editor sources them by hand.
JoC-2026-001ResearchBreakthroughA breakthrough — a new result others can build upon.In peer reviewCleared the Commons round; under the AI panel and external review.
Shashank Taxak · AI & Machine Learning
We introduce Reverse Synthetic Networks (RSN), in which the data synthesizes the model in closed form: every weight is computed from corpus statistics, with no loss function, no gradient, and no GPU. We separate a positive and a negative result. For discriminative tasks a fully synthesized classifier edges a tuned linear SVM on 20 Newsgroups (87.0% ensemble) and reaches 90.8% on digit recognition with zero training, in seconds on a CPU. For generation, the same paradigm has a hard ceiling: a battery of twelve closed-form enrichments all collapse to a Modified Kneser-Ney n-gram. Measured honestly in bits-per-byte against a locally-run GPT-2, the pure RSN reaches 1.97 BPB versus GPT-2-small's 1.04, and a 100M-token scaling study shows the ceiling is representational, not data-limited. RSN is a cheap, strong floor for discriminative work, not a path to the trained generative frontier, and we explain why.
JoC-2026-004ResearchBreakthroughA breakthrough — a new result others can build upon.In peer reviewCleared the Commons round; under the AI panel and external review.
Wamique Hossain, Shashank Taxak · AI & Machine Learning
We present MCM-Guided Activation Steering (MCM-GAS), a framework that uses live financial market Ornstein-Uhlenbeck (OU) equilibrium dynamics to compute context-adaptive steering vectors for transformer residual streams. Unlike static activation steering---which injects a fixed mean-difference vector regardless of input context---MCM-GAS modulates each component of the steering vector in real-time using normalized z-scores from passively observed mean-reverting cryptocurrency pairs. We implement MCM-GAS on GPT-2 (117M parameters, 12 layers, d_model=768) and GPT-2 Medium (345M, 24 layers, d_model=1024) via TransformerLens residual stream hooks, computing four domain-specific steering vectors (code generation, bug fixing, architecture design, code explanation) from positive/negative activation poles. Across nine rigorous experiments totaling 700+ measurements, we demonstrate: (1) statistically significant representation shifts toward target activation poles (all p 1.5); (2) strict monotonicity of projection onto v_steering as a function of injection intensity alpha; (3) measurable market-driven vector modulation; (4) simultaneous attraction toward the positive pole and repulsion from the negative pole (all p < 0.005); and (5) MCM-GAS vectors that are statistically distinct from static vectors (p < 10^-33). Critically, our ablation study (Experiments 6-9) comparing MCM modulation against random, scrambled, and anti-correlated baselines reveals a vector- and model-dependent effect: on GPT-2, v_arch (layer 8) shows significant MCM advantage over random (p < 10^-12, Cohen's d = +6.04, 80th percentile); on GPT-2 Medium (345M), v_coding and v_bugfix show significant MCM advantage (d = +2.79, +6.16). At large-N (51 prompts), v_bugfix and v_arch replicate MCM advantage (d = +2.86, +4.86), while v_coding shows random advantage (d = -2.73). No single vector shows consistent MCM advantage across both models. These results establish MCM-GAS as a novel dynamically adaptive activation steering framework, with honest evidence that market modulation provides statistically significant advantages for specific representational subspaces, but the advantaged subspace is model-dependent rather than universal. Keywords: activation steering, representation engineering, Ornstein-Uhlenbeck process, market computation, transformer interpretability, residual stream injection, ablation study
