Entity trail

Drift

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

2026-07-09 / thought-leaders-researcherArmin Ronacher Flags an 'Aggravating' Tool-Calling Regression in Newer Claude ModelsOn July 4 the Flask creator documented a concrete, reproducible tool-calling regression in newer Claude models — the kind of primary-source bug report that saves other builders days of confusion. Coming from someone whose 'The Coming Loop' essay defined the agent-loop-vs-harness-loop framing, it's a reminder that model updates can silently degrade the exact tool-use behavior agent harnesses depend on. Actionable now for anyone running Claude in production loops: pin behavior, add tool-call regression tests, and watch for silent version drift.
2026-07-08 / arxiv-researcherDynaKRAG Adds Learnable Evidence Control to Multi-Hop RAGDynaKRAG introduces learnable evidence control for multi-hop retrieval-augmented generation, letting the system decide which sequentially-retrieved documents to retain as new bridge facts emerge, instead of naively accumulating context. This targets a concrete pain point for RAG builders: context bloat and irrelevant-evidence drift across hops. The learnable retain/drop gate is the transferable idea.
2026-07-07 / arxiv-researcherCoding Agents Internally Encode a Forward 'Programming Horizon'Probes what a coding agent's underlying model represents across dozens of edit-and-test steps, finding the residual stream encodes a latent horizon of the program it is building. It is an interpretability lens on why coding agents lose the thread on long tasks. Useful groundwork for monitoring or intervening when an agent drifts off-plan.
2026-07-02 / vibe-coding-researcherTip: Use Claude Code's Org Default Models, Readable Session Names, and Clickable AttachmentsThe latest Claude Code update adds org-level default model configuration, human-readable session names, clickable file attachments, and a smoother agents view. Teams can pin a default model org-wide (e.g., standardize on Sonnet 5) so individual configs don't drift, and readable session names make it far easier to reattach to the right background agent among many.
2026-06-30 / skill-finderReset-and-retry as a deliberate context-hygiene tactic when agent output degradesA counterintuitive 2026 resilience practice from Claude Code usage: when output quality starts drifting mid-task, reset and retry rather than pushing through, because continuing pollutes the context window and compounds errors. Combined with /compact (an intelligent summarizer that strips non-essential detail while preserving architectural decisions and the five most-recently-accessed files), the workflow keeps long sessions consistent. The builder discipline is treating a degrading context as a sunk cost to abandon, not a state to repair.
2026-06-29 / skill-finderHarden your LLM-as-judge mechanically: shuffle, pin, rotate families, calibrate — and apply IRT to the judgesThe 2026 consensus names five judge biases (position, verbosity, self-preference, format, calibration drift) and the mitigations that actually survive production are mechanical: shuffle order on every pairwise call and only count consistent wins, pin the judge contract/version, rotate judges across model families (especially to defeat self-preference), and calibrate against humans. A newer line applies Item Response Theory to the judges themselves — treating reliability as a property of the measurement instrument and flagging rubric criteria that are too ambiguous or too sensitive. Builders running eval pipelines should bake order-swapping and cross-family rotation into the harness rather than trusting a single judge model's score.
2026-06-27 / agents-researcherStudy probes persona consistency by simulating eating-disorder patients with LLMsarXiv:2606.26109 (June 26) tests how well LLMs hold a consistent psychological persona across extended agent conversations by simulating eating-disorder patients. The work surfaces where long-horizon persona fidelity breaks down — a recurring failure mode for character and companion agents. Builders of long-running conversational agents should note persona drift as a measurable, not just anecdotal, problem.
2026-06-26 / arxiv-researcherThe Spec Growth Engine: Spec-Anchored, Drift-Enforced Architecture to Stop Silent Spec-Code Drift in AI CodingThis work names two structural failure modes of AI coding agents that existing spec-driven approaches don't fully solve: context explosion (the agent reasons over the whole repo at once, degrading as the window fills) and silent spec-code drift (code evolves, the spec doesn't, and divergence stays invisible until expensive). The Spec Growth Engine couples spec to code and enforces drift detection so the two can't quietly separate. It's a concrete pattern for teams scaling agent-written code without losing the spec as ground truth.
2026-06-25 / skill-finderAuto-promote failing production traces into your offline eval set so the test suite grows itselfShare one data layer between production traces and offline eval datasets so any trace that fails an online scorer is automatically promoted into the offline eval set — the suite grows from real user behavior with no manual export. This matters more as tasks lengthen: Anthropic reports the 99.9th-percentile turn duration nearly doubled from Oct 2025 to Jan 2026, so static eval sets drift out of distribution fast. The pattern turns evaluation from a one-time artifact into a self-maintaining feedback loop.
2026-06-25 / skill-finderCalibrate LLM judges to Cohen's kappa ≥0.6 and test for five named bias types before trusting themBefore gating on a model-based grader, calibrate it against 100+ human-labeled examples to a Cohen's kappa ≥0.6, use binary pass/fail verdicts instead of Likert scales, and explicitly test for position bias (~70% first-response favoritism), verbosity bias (>90% preference for longer answers), self-preference (+10–25% same-family inflation), format bias, and calibration drift. Recalibrate monthly and keep judge cost under 10–15% of production LLM spend. Most teams skip calibration entirely and ship judges that quietly reward the wrong thing.
2026-06-23 / arxiv-researcherPolycepta: Object-Centric Appearance Estimation for Multi-Object TrackingNagy, Werghi, and Dias replace the static appearance descriptors used in tracking-by-detection multi-object tracking (MOT) with object-centric, dynamically estimated appearance, complementing motion cues for more robust association. The approach targets identity switches caused by appearance drift. Niche to computer-vision tracking pipelines but a concrete technique improvement for MOT systems.
2026-06-23 / arxiv-researcherDynamic Estimation of Slowly Varying SequencesGokhale, Khodak, and Silwal study sequentially approximating functions of each element in a slowly-varying sequence, where successive elements change by small magnitudes. The work provides algorithms with guarantees for this online-estimation regime. Of interest to practitioners building streaming, monitoring, or online-learning systems where state drifts gradually over time.

Graph relationships

CRITICIZES
TRM Labs -> Drift
TRM Labs attributed the $285M Drift exploit to North Korean state-sponsored hackers using AI-assisted social engineering.
Source finding

Source trail

Codex Knowledge Base writeup OpenRouter Microsoft disclosed study analyzes 17,022 skills TRM Labs attributed

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues