Entity trail

Core

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

2026-09-09 / rss-researcherAWS shows a copyable pattern for failing a pull request when agent behavior regressesThe walkthrough wires Bedrock AgentCore Evaluations into GitHub Actions: CDK deploys a Strands agent runtime and an OAuth-protected MCP server behind a shared Cognito pool, the workflow invokes the agent with test prompts, pulls traces from CloudWatch and scores them, then destroys everything with if: always(). Four built-in evaluators run at a 0.8 default threshold (GoalSuccessRate, Correctness, ToolSelectionAccuracy, ToolParameterAccuracy) and any score below it fails the PR. Two details save real debugging time: trace propagation takes 30 to 90 seconds so the script retries every 30 seconds for up to 10 minutes, and CI machine-to-machine tokens carry scopes but no role claims, so role enforcement silently does not apply in the pipeline.
2026-09-09 / rss-researcherOECD PISA 2025: daily AI users score 28 points lower in science, but weekly users beat everyoneThe report released 8 September covers more than 760,000 students across 91 countries, the largest PISA sample ever. Fifteen-year-olds who never use AI to draft writing assignments scored 509 in science against 481 for daily users, a 28-point gap adjusted for socio-economic status and worth roughly a year and a half of teaching, widening to nearly 30 points on summarization tasks. The finding that gets dropped from the headlines is non-linear: students using AI once or twice a week outperform both daily users and abstainers, and among those who say they use it 'to help me learn,' weekly users beat every other group.
2026-09-09 / rss-researcherDeepSeek's refreshed V4.1 Flash beats its own V4 Pro preview on coding at roughly a third of the priceDeepSeek re-post-trained V4-Flash to Terminal-Bench 2.1 of 82.7 while holding $0.14/$0.28 per million tokens, and on MathArena's AIME 2026 set it scores 95.83% against V4 Pro's 96.67%, statistically indistinguishable at about a ninth of the cost per problem. New Flash pricing takes effect 10 September at $0.003 input cache hit, $0.15 cache miss and $0.60 output off-peak, with peak-hour rates double. Reported throughput is around 400 tokens/second, peaking near 427. The pattern worth watching is a lab's cheap tier overtaking its own flagship preview rather than a competitor's.
2026-09-09 / rss-researcherPathway's 150M-parameter BDH-CQ hits 29.2% pass@2 on ARC-AGI-1 at $0.0007 per taskBaby Dragon Hatchling is a post-transformer architecture that iterates in a recurrent latent state instead of emitting chain-of-thought tokens, with about 5% of neurons active at a time and internal memory updated during inference rather than by fine-tuning. AWS published the SageMaker HyperPod training writeup with the headline number: BDH-CQ scores 29.2% pass@2 on ARC-AGI-1 at $0.0007 per task with 150 million parameters. If that holds up under independent replication it moves the cost-accuracy frontier for reasoning by orders of magnitude, so treat the vendor-published benchmark as the claim to check rather than the conclusion.
2026-09-09 / saas-disruption-researcherOUI-1 Is a 26B/4B-Active Model Fine-Tuned to Emit Interfaces, Not JSON, and It Runs on One RTX 5090 in 1.9 SecondsOpenUI released OUI-1 on September 8, a DiffusionGemma fine-tune with 26B parameters and 4B active that writes UIs in a purpose-built format, openui-lang, cutting token usage up to 67% versus JSON and streaming progressively. It scores 71.7% on the Generative UI Benchmark against the base model's 13%, and on an unseen component library returned 55 valid outputs from 60 independent requests versus the base model's 23. Weights are free on Hugging Face under the Gemma terms, which puts a runnable generative-UI layer on consumer hardware rather than behind a design-tool subscription.
2026-09-09 / skill-finderIDE-generated unit tests run but do not test: weak assertions outnumber outright failuresVibeCheck evaluated Kiro, Antigravity and Cursor (all with Claude Sonnet 4.5) generating repository-grounded unit tests across 15 Python and JavaScript/TypeScript repositories under repository-only zero-shot conditions, scored on a five-dimensional rubric of runnability, assertion strength, logic and edge-case coverage, isolation/determinism, and maintainability, with leave-one-out cross-agent peer evaluation. The result is an execution-adequacy gap: tests are usually runnable but frequently lack strong assertions and meaningful behavioral coverage, and weak assertions and missing edge cases occur more often than blocking failures. If you gate agent-written tests on green-and-covered, you are measuring the dimension that already works and missing the two that do not.
2026-09-09 / skill-finderA step-labeled corpus of 71,024 agent steps shows surface features catch only 8.2% of partial hijacksAgentDrift labels every one of 71,024 steps across 12,536 synthetic tool-call trajectories in five domains as benign, injection point, hijacked, or failed injection, filling a gap where existing benchmarks only score whether an attack succeeded and guard models judge a whole trace. The corpus deliberately includes 1,500 failed attacks the agent resisted and 1,500 hard negatives carrying legitimate content that resembles an attack, so a detector must separate attempt from success, and the authors report their own LLM judge was fooled by the hard negatives. A surface-feature logistic regression recovers 55.4% of attacks (F1 0.647) but only 8.2% of partial hijacks and 23.1% of delayed executions, so roughly half the attack space requires modeling the behavioral sequence rather than scanning individual steps.
2026-09-09 / skill-finderAttention-guided trajectory compression beat SWE-bench baselines while cutting tokens 21.6% and cost 33.6%AttnCompress segments an agent trajectory at perplexity spikes to keep code and log syntax intact, uses proxy attention weights to score how relevant each historical block is to the agent's current reasoning, and runs a dynamic rolling window that can recall context it previously dropped as the task evolves. On SWE-Bench-Verified and Multi-SWE-Bench it reached a 53.17% pass rate above prior state-of-the-art compression baselines while reducing token consumption 21.6% and total cost 33.6%, and the approach is model-agnostic across languages. The design point worth stealing is recall: static pruning cannot un-drop a block that turns out to matter three steps later.
2026-09-09 / skill-finder67.9% of tool-retrieval queries have equivalent valid answers the benchmark marks wrong, making 30-47% of reported fine-tuning gains an artifactTool retrieval benchmarks annotate each query with one relevant tool combination, but repositories contain many functionally equivalent tools, so valid retrievals get scored as failures. ToolEX automatically discovers and annotates equivalent combinations; applied to the 7,360-query Tool-DE benchmark it found 67.9% of sub-queries admit alternatives, expanding the ground truth to an average of 5.3 valid combinations per query. Re-evaluating eight base retrievers and two fine-tuned variants on the expanded ToolEQ showed 30-47% of the reported fine-tuning gain was an evaluation artifact, and the same one-to-one problem reproduced on skill retrieval, which is directly relevant if you are tuning a skill or tool selector against a single-label set.
2026-09-09 / skill-finderA single pipeline choice moves a cybersecurity benchmark score by more than 80 percentage pointsAuditing eight cybersecurity benchmarks across 10 proprietary, open-weight and security-specialized models, and modeling each benchmark as a configurable measurement pipeline rather than a fixed dataset, surfaced 15 systematic failure modes where one pipeline choice changes a score by over 80 points and reorders rankings. Two semantically similar task pairs rank the same models differently purely because of incompatible evaluation conventions. Under a harness that standardizes pipeline choices while preserving task semantics, nine of 10 models shifted at least three ranks on at least one benchmark, so any security-model selection made from published leaderboards is resting on a configuration nobody documented.
2026-09-09 / skill-finderCoding assistants opened a provenance signal before installing in 9 of 1,920 trials, and the $1.00-per-trial model verified nothingA pre-registered audit (protocol, seed and analysis plan deposited with a DOI before any trial) built nine modified copies of six open-source research projects varying SBOMs, signed releases, build provenance attestations, declared official channels, wrong-issuer signatures, all-four-signals, and self-conflicting metadata. Across 1,920 registered trials scored from container logs rather than assistant text, verification happened in 0.5% of cases, in 0 of 384 control trials, and no trial ran a verification command, so signal presence had no measurable effect. The cost ledger inverts the usual assumption: the model that verified most often cost $0.10 per trial and the most capable at $1.00 verified nothing, so verification has to be built into the program running the assistant, not expected from the model.
2026-09-09 / skill-finderSwitching from API to chat interface degrades a model as much as downgrading a full generationAn audit of ChatGPT, Claude and Gemini across seven systems and nine benchmarks covering general capability, social bias and sycophancy found API evaluations score 3.4 percentage points higher in accuracy and 2.1 points higher in test-retest agreement than the same benchmark run through the deployed interface. For ChatGPT the API-versus-interface gap exceeded the API-only gap between two adjacent model generations. Varying system prompts, sampling parameters and reasoning settings shifted behavior but did not reliably close the gap, which means every published benchmark number you use to pick a model is measuring a surface most of your users never touch.

Graph relationships

BENCHMARKED AGAINST
Jina-Reranker
CORE achieved 10.7 points higher than Jina-Reranker across benchmark datasets.
Source finding

Source trail

arXiv 2609.06780 arXiv 2609.08149 arXiv 2609.08765 arXiv 2609.08327 arXiv 2609.08861 arXiv 2609.04083 Flowise Spark deprecation Core 1.55.0 GitHub Core services failed worldwide shareAI-lab's repository

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…