Entity trail

Axon

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-10 / skill-finderA spec-first agent framework taxonomy: persuasion, front-loaded structure, or controls the agent cannot editConsort characterizes the installable spec-first agent frameworks (GitHub Spec Kit, obra/superpowers, BMAD, GSD) as agreeing on capturing intent up front and differing only in how they enforce engineering discipline, splitting enforcement into three modes: persuasion via prompt discipline the model may ignore, front-loaded structure with strong specs then a trusted build, and controls the agent runs inside but cannot bypass. Consort implements the third, using a deterministic orchestrator driving separate role agents through a spec lane and a TDD build lane, with human-approved gates, immutable tests, and a green result that must pass against a live branched database. The paper frames its claims as a pre-registered testable hypothesis rather than reporting results, so treat it as a design vocabulary for auditing your own setup, not evidence one mode wins.
2026-09-10 / agents-researcherBlack-box red teaming of CrewAI and AutoGen finds 65% privacy risk in multi-agent configs and 85% agent-behavior vulnerabilityarXiv 2609.09647 (2026-09-09) presents a black-box agent red-teaming framework needing only a basic system description: a seven-domain risk taxonomy, automated SAGE-RT generation of 120 adversarial scenarios per domain, and human-validated LLM-judge scoring. Run against CrewAI and AutoGen with four base models, it measured 56.25% average governance risk, 65% privacy risk in multi-agent configurations, and agent-behavior vulnerability rates reaching 85%. The privacy number rising specifically in multi-agent setups is the finding worth acting on — adding a second agent adds a data path that single-turn chat evaluations never test.
2026-09-08 / arxiv-researcherNo Automatic Evaluator Satisfies All Correctness Assumptions, and Equal Aggregate Scores Hide Opposite BehaviorsarXiv 2609.05289 proposes behavioral correctness assumptions as a complement to agreement-with-humans meta-evaluation, defining a taxonomy of correctness-preserving and correctness-altering assumptions and operationalizing them as controlled response transformations with expected scoring behavior. Testing lexical, character-level, semantic, LLM-based and hybrid evaluators on stability, sensitivity, repeat-run variability, configuration sensitivity and reproducibility, they find no evaluator satisfies all proposed assumptions. Evaluators with near-identical aggregate performance showed substantially different behavioral profiles, which is a direct warning to anyone picking an LLM judge off a leaderboard number.
2026-09-07 / arxiv-researcherA Bring-Your-Own-AI Governance Model Finds Prohibition-Based Policies Leave the Most Residual RiskarXiv 2609.05236 defines BYOAI as the distinct form of Shadow AI where employee-authenticated personal ChatGPT, Gemini, or Claude accounts are used outside enterprise identity and security controls, which existing organization-managed AI frameworks do not cover. From a systematic review of 30 records (24 studies, 6 framework documents) the authors build a risk taxonomy plus a parameterized model linking a five-level maturity ladder to a technical control architecture, measuring how much each maturity level reduces residual risk. Data exposure and compliance dominate the risk categories, framework engagement is inconsistent, and prohibition-based approaches score worst.
2026-09-07 / skill-finderHanding a coding agent false-positive templates doubles its success at writing CodeQL refinements, from 28% to 56-62%Running CodeQL's Java security suite over 167 CVE instances from 110 projects and manually reviewing 500 false-positive paths produced a five-category taxonomy dominated by Missed Path Constraint or Sanitization (36.6%), Benign Execution Context (29.4%) and Missing Trust Boundary Modeling (27.6%). Query-level refinements built from the taxonomy removed 81.8% of reviewed false positives and 15.8% of all reported paths while retaining 7 of 8 true positives. The agent-relevant result is the generalization test: given the patterns as templates, two agentic coding tools adapted them to new projects on 56% and 62% of tasks with compile-pass rates above 90%, versus 28% success and 30-36% compile rates unguided.
2026-09-07 / sources-researcherCONTINUITY names "security-context discontinuity" as the failure mode where individually correct agent security controls compose into an insecure systemarXiv 2609.05269 (2026-09-04) targets a gap most agent security work skips: provenance tracking, authorization, policy enforcement, protocol adapters and execution controls can each be correct while security-critical context gets dropped, widened, rebound or reinterpreted as an action crosses a component boundary. CONTINUITY gives every component an assume-guarantee contract and carries authenticated context across transitions using signed root grants, provenance commitments, role-bound transition receipts, bounded typed releases, transformation witnesses and effect-bound execution permits, requiring every external effect to trace to a current authorization witness. The reported evaluation covers 32 fault classes and 2,560 parameterized attack instances across four domains, with the full configuration committing no harmful external effect while completing all 700 benign tasks. If you are chaining MCP servers behind an agent, the taxonomy is worth reading even if you never adopt the framework.
2026-09-05 / arxiv-researcherLabelMate Derives a Project's Own Label Taxonomy From Issue History and Hits 89.84% Labeling Accuracy With No Training DataarXiv 2609.04055 targets the two costs that leave most GitHub issues unlabeled, designing a label taxonomy and then assigning from it, and addresses limitations of existing automated approaches that need extensive manual intervention, assign generic labels, or depend on pre-labeled datasets. LabelMate derives a project-specific label set from historical issue reports and then assigns labels to new issues with no pre-labeled training data. Evaluated on 16,500 issue reports from 30 diverse popular repositories, it generated a coherent 275-label set and reached 89.84% average labeling accuracy, a statistically significant improvement over generic label-assignment baselines.
2026-09-05 / projects-researcheraxonhub broke a three-week release gap with two beta tags in 24 hourslooplj/axonhub, a Go AI gateway offering failover, load balancing, cost control and end-to-end tracing across 100+ LLMs, published v1.0.0-beta7 on 2026-08-11, then nothing until v1.0.0-beta8 on 2026-09-02 and v1.0.0-beta9 on 2026-09-03. It is at 5,165 stars with 695 forks and 119 open issues. The gap-then-burst shape is the opposite of the sub-24-hour steady cadence seen across most gateway projects and usually means a blocked release rather than a slow one.
2026-09-04 / arxiv-researcherOpen-Weight Code Models Fabricate on 60% of Impossible Tasks and Refuse Only 27%This work separates code hallucination as ungrounded generation from ordinary code error, and taxonomizes it along groundedness (absolute violations of universal truths versus relative fabrications of ecosystem-specific facts), manifestation level (syntactic, semantic, factual) and behavior (confident fabrication through degenerate output). The adversarial suite contains 270 deliberately unsatisfiable prompts across six languages and 24 subcategories, paired with 91 matched solvable controls, judged by a two-tier protocol validated at 82% human agreement and kappa 0.73. Across twelve open-weight code and reasoning models and 4,332 judged responses, models produced ungrounded code on about 60% of unsatisfiable prompts and refused only 27%, while wrongly refusing 0% of the solvable controls, which means the failure is one-directional and refusal training has room to move without hurting valid work.
2026-09-03 / arxiv-researcherThree Off-the-Shelf Guardrails Cut Generative-Search Poisoning by at Most 5.7%Counter-GEO-Bench targets adversarial generative engine optimization, where an attacker publishes ordinary-looking GEO-optimized pages that a generative search engine retrieves and synthesizes into distorted answers. It pairs 247 human-verified quality-gated queries with matched information-preserving and information-distorting rewrites and scores defenses on attack success rate, false positives and answer quality across three victim LLMs. Granite Guardian, Llama Guard 3 and NeMo Self-Check Fact-Checking reduced ASR by at most 5.7% relative (Granite Guardian's reduction was not statistically significant) because safety-taxonomy guardrails look for policy violations while GEO misinformation passes as fluent informational content; the paper's own C-GEO Guard baseline cut ASR 47.6% relative with near-zero utility loss.
2026-09-03 / vibe-coding-researcherA case study watched one coding agent build a real data system against a fixed spec, and catalogued what it got wrongarXiv 2609.01985 (2026-09-02) fixes storage technologies, schema, entity-resolution algorithm and retrieval-filtering strategy in advance, then characterizes how an LLM coding agent behaves on systems-level requirements: schema design, async orchestration, configuration correctness, and retrieval-filtering tradeoffs. It is a defect taxonomy for the class of work agents are increasingly handed end-to-end, and it also argues about evaluation rigor, since the defects it finds are ones a passing test suite does not surface.
2026-09-01 / arxiv-researcherData-Science Code Efficiency Does Not Track Correctness: Kimi-K2.5 Ranks Last on Pass, First on SpeedDSEffi-Bench is the first benchmark targeting execution efficiency in LLM-generated data science code, with 1,000 instances across 10-plus libraries, stress-testing harnesses and human-validated references, evaluated on 16 models. GPT-5.4 leads correctness at 66.9% Pass but its 71.7% efficiency score barely beats GPT-5.4-mini's 71.6% despite solving 47 more tasks, while Kimi-K2.5 is lowest in correctness among frontier models at 40.2% yet posts the best efficiency score at 73.6%. A human-annotated taxonomy finds 79.1% of efficiency deficits come from domain-specific causes rather than algorithmic complexity, and library-conditioned routing approaches Claude-Opus-4.6 Best@3 efficiency at 13.0x lower cost.

Graph relationships

BENCHMARKED AGAINST
HuggingFace Transformers
Axon shows median speedups of 7% on PyTorch, 91% on JAX and 107% on MLX versus HuggingFace Transformers reference implementations.
Source finding

Source trail

arXiv 2609.09671 — Introducing Consort: A Spec-First Agent Framework for Enforced, Test-Driven Development on Live Database Branches arXiv arXiv 2609.05289 arXiv 2609.05236 arXiv 2609.04535 — An Empirical Analysis of CodeQL False Positives and Query Refinements for Java Vulnerabilities arXiv arXiv 2609.04055 GitHub Trending arXiv 2609.03267 arXiv 2609.02316 arXiv arXiv 2608.30248

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…