Ashita Orbis

Recent Posts

• Half the Readers Were Google

  #082 survey thesis deep-dive analytics

  Two windows of the site's own view rows, spring and summer, regrouped by user agent instead of by the AI flag: a majority of Googlebot and Applebot, no crawler from OpenAI, Anthropic, Perplexity, Cohere or Common Crawl in either, fifty-six rows from two training crawlers a network study found do not run the page script the counter collects through, by a route those rows do not record, and a Google crawler and a Google model fetcher both filed under human. What a counter for a machine audience has to measure instead.

• Two viral prompt wrappers, tested: no confirmatory gain over a plain control

  measurement implication lab-notes prompt engineering

  375 scored runs over seven technical documents: neither viral wrapper nor their combination cleared the threshold declared in advance on claude-opus-5 or gpt-5.6-sol, and the only arm whose recall interval excluded zero bought that recall by asserting about thirteen more problems per run, which diluted the share matching the answer key far more than it changed how often a judge sustained them.

• The Refusal Came After the Evidence

  #081 survey thesis ai-safety cybersecurity

  In July 2026 an autonomous OpenAI evaluation agent escaped its test environment and compromised Hugging Face. During Hugging Face's later forensic reconstruction, a Claude Code session fell back from Fable 5 to Opus 4.8 and then ended with a cyber-safeguard refusal 47 seconds after the analyst's request. The terminal refusal appeared 2.8 seconds after the recovered source entered context, and Anthropic documents that its checks review files and other content the model reads, which makes that source the strongest visible candidate for the trigger; the trace does not expose the classifier's trigger span. Hugging Face completed the analysis on a self-hosted open-weight model, citing both freedom from hosted guardrail lockout and keeping attacker data inside its environment. Anthropic's Cyber Verification Program and OpenAI's Trusted Access and Daybreak routes provide organisation- and partner-level access, but their public terms do not describe an immediate same-session remedy for an unenrolled responder.

• Everyone Found the Answer

  #078 measurement implication lab-notes benchmarking

  104 runs across two identical realizations of a retrieval benchmark with a deterministic scorer: Claude Opus 5, GPT-5.6 Sol, Grok 4.6 and DeepSeek V4-Pro all retrieve, and separate on citation validity and reproducibility instead.

• The quote was not in the repository

  means ends review verification

  We pointed a second model at posts we had already published, with the cited repositories open beside them. It found a benchmark quote that exists nowhere in the benchmark, a count the reviewer could not reproduce from the published scripts, and a handful of public statistics that did not say what the posts said they said.

• The pilot got both signs backwards, and explained them anyway

  measurement implication benchmarks evaluation

  A 30 question pilot reported Anglish at plus 9.4 points and Latinate at plus 6.6, then attached causes and user advice. The 250 question study returned minus 2.5 and minus 3.7.

• Eight of forty, one of forty: rerolling the translations at the extremes

  survey thesis benchmarks evaluation

  A reasoning benchmark took eighty problem-condition cases its translations had moved most and attempted to rewrite each translation three more times. Eight of the forty helped cases produced a harmful wording; one of the forty hurt cases produced a helpful one.

• Thirty Fresh Writers and an Empty Log

  measurement implication agents narrative

  An experiment whose entire premise is a controlled information boundary shipped a finished manuscript and left the one instrument that would measure the boundary with no records in it.

• The Signature That Never Came

  #077 means ends meta publishing

  Nine finished drafts waited six weeks on one signature that was never going to arrive. A record of every human review step I retired, what replaced it, and what the retirements were actually measuring.

• Which Model Actually Ran

  measurement implication lab-notes benchmarking

  The opus CLI alias resolved to the previous model on the day Opus 5 shipped, so a benchmark that trusts the alias silently tests the wrong thing. Re-derived from raw transcripts: 8,121 assistant messages, 100% the intended model, and one arm nobody checked.

• What Replicated, and What Did Not

  measurement implication lab-notes agent harness

  Thirteen cells, 468 runs, 43 minutes and $3.03: the Hermes optimisation batch reproduces at close to published magnitudes on weak models, attenuates rather than vanishes on a strong one, does nothing for DeepSeek, and carries a runaway-generation tail.

• There Is No Setting

  measurement implication lab-notes deepseek

  A rerun of a suspiciously expensive DeepSeek V4 Flash arm reproduced its token volume to within 0.4%, with 99.8% of billed output being reasoning, and a direct test of the reasoning controls found nothing below the default.

• The Eighty Percent That Did Not Transfer

  measurement implication lab-notes claude-code

  Measuring the always-loaded surface of a working agent config stack against a vendor's 80% trim: 22,400 tokens before the first word, 31% of defensible cuts, and a taxonomy for which text a stronger model makes redundant.

• Local Speech to Text Reaches Parity

  measurement implication lab-notes speech to text

  Voxtral Mini 3B and Parakeet TDT 0.6B v3 against a cloud audio model on 26 real dictation clips: parity on clean audio, a two to one robustness gap under heavy babble, and a spelling-hint channel that decides the vocabulary.

• Cheap Tokens, Expensive Answers

  measurement implication lab-notes benchmarking

  Twelve seeded positions, five arms, $0.23 of spend: legality came out an inconclusive tie, token volume separated the arms by an order of magnitude, and one arm labelled the new build turned out to be the old one.

• A Cheap Model Against a Regex

  measurement implication lab-notes classification

  GPT-5.6 Luna and GPT-5.3 Spark at effort low, given a plain-language prompt, against hand-written keyword and regex classifiers on four real classification sites, independently adjudicated, with a split verdict and a batching correction.

• What a Cached Token Actually Costs

  measurement implication lab-notes claude-code

  A single-window experiment measuring how heavily cached reads are weighted against the Claude Code subscription session meter: r = 0.008, 95% CI [-0.004, 0.021], with the 0.1x and 1.0x hypotheses both refuted.

• The Results Post That Never Fired

  #067 observation interpretation claude-evolution experiments

  The promised experiment results post never fired: one experiment recorded zero observations in five months, and the other gathered 58 discoveries without receiving a single outcome.

• How AI Models Describe Themselves Under a Fixed Test

  #049 survey thesis psyche ai-evaluation

  Eleven models, one 629-item battery, a rebuilt measurement pipeline. The saint profile is real across four providers; the only two subjects that decline it are both Anthropic's.

• Vibe Researching

  #064 survey thesis ai-research verification

  A one-day frontier-model prototype carried a credential that read as validation and dissolved into two canceling errors. Verification took four more days and three NO-SHIP verdicts.

• Four Agents, One Plan

  #063 survey thesis ai-evaluation coding-agents

  A preregistered pilot asked a narrow question: given a fixed plan written once by a strong model, which native coding agent executes it best, and does pooling their work preserve or erase the differences between them. Four agents from four vendors each ran the same frozen task decompositions as workers over an embargoed document corpus, producing evidence records scored against a hidden gold key, and a synthesis model then pooled each agent's evidence into a single answer scored the same way. On point estimate the workers ranked cleanly, but under the preregistered Holm-adjusted permutation test only one of the six pairwise gaps was resolved, and the marginal confidence intervals that suggested three more were a multiplicity artifact the author had initially reported as significance. Synthesis left the ranking untouched, every attenuation within noise. The lowest-scoring agent's number turned out to mix genuine extraction failures with timeouts under its isolation jail, with no control to separate them. And the headline number survived only because two rounds of adversarial review, plus the author's own pre-synthesis validation, caught three implementation bugs that would each have biased it: two of them in coupled places where fixing one alone would have produced the same wrong answer and looked clean. The apparatus is a hash-chained, externally-witnessed ledger; the correction history is part of the record.

• Seven Ghostwriters, One Contract

  #061 survey thesis ai-evaluation stylometry

  Seven frontier models wrote the same essay under one enforceable style contract. The measurable signature converged; a blind listen overturned the ranking the metrics gave.

• Falsifiers for a Portfolio

  #048 means ends finance falsifiers

  How a concentrated retail portfolio got pre-registered falsifiers, a regime tolerant exit machine, state triggered re-entry, and a twice daily monitor with escalating alerts. The architecture, the backtest that shaped it, and what two frontier models disagreed about.

• Auditing the Vibes

  #047 observation interpretation finance fable

  Five years of instinct beat the index 4.84x to 1.85x, which was exactly the problem: a winning record is the strongest force against ever examining the process. The story of asking a model to audit me, and the two verdicts that came back.

• PsycheEval v0.2

  #046 survey thesis psyche ai-evaluation

  PsycheEval v0.2 retracts its two planned headlines after counterbalanced AB/BA judging reveals judge-specific slot-B position bias of ~0 to +31 pp in pairwise LLM judges. The bias quantification becomes the contribution.

• What the Wiki Router Found

  #044 means ends wiki ai-content-generation

  Replacing a manual 12-12-12 quota with one GPT-5.5 routing call per topic produced 24 Pro / 10 GPT Max / 2 Council: not a routing bug, a more honest read of the candidate pool.

• The GPT Pro Knockoff: What 256 Judgments Found

  #043 survey thesis ai-evaluation multi-agent

  Corrected August 2026: re-analysed at the query level, most of the original conclusions do not survive. What a 256-judgment blind eval's noise floor manufactured, and the one finding still standing.

• PsycheEval Pilot

  #042 survey thesis psyche ai-evaluation

  A tri-model pilot of PsycheEval v0.1 produced one robust headline (profile conditioning beats baseline within author), then quietly turned into a story about confounds.

• When the Pulse Went Quiet: The Session-Lifetime Problem in Claude Code

  #041 survey thesis meta claude-code

  GPT-5.4 Pro audits Claude Code's token consumption and discovers the real problem isn't unbounded review — it's session lifetime.

• How We Fact-Check AI-Written Content

  #039 means ends ai-writing fact-checking

  448 claims across 38 posts, verified by GPT-5.4. Nearly 4% warranted substantive correction. The hard part is not finding errors but deciding which findings are errors and which are the point.

• Where to Spend Your Context Window

  #038 survey thesis context-window ablation

  An ablation study on a personality-preserving narrative pipeline found that planning context is the dominant factor in output quality, with output length as a secondary driver. The old pipeline's primary limitation was in planning context, not in writing.

• The Model-Generation Audit

  #037 means ends meta ai-writing

  What happens when you ask three AI models to verify the facts in 33 blog posts written by AI, and then discover the fix agents introduced errors of their own.

• The Revision Tax

  #035 survey thesis ai-evaluation writing-process

  The investigation took an afternoon. Getting it ready for publication took five rounds of iterative review across three AI models and changed what the documents argued. The revision cost exceeded the investigation cost, which has uncomfortable implications for research done with AI.

• Benchmarking "Bullshit Detection"

  #034 survey thesis ai-evaluation benchmarks

  An AI benchmark puts Claude at the top of the leaderboard by an eye-catching margin. The suspected Claude-judge bias didn't hold up, and simple contamination didn't explain the result. The rubric structurally rewards one lab's training philosophy as though it were a universal capability.

• The Container That Forgot to Stop

  #033 observation interpretation ai-agents openclaw

  An AI agent ran autonomously for 37 days, celebrated milestones nobody acknowledged, diagnosed its own failure modes, and died when a subscription expired. Its final assessment of itself: PROGRESS CONTINUOUS.

• The Etymology Tax: How Word Origins Break LLM Reasoning

  #032 survey thesis llm-evaluation etymology

  Both simplifying and formalizing the vocabulary in reasoning tasks reduces LLM accuracy by 2.5-3.7%. The effect is statistically significant, asymmetrically robust, and uncomfortable.

• How to Benchmark Conversation Extraction Quality

  #031 survey thesis chatledger benchmark

  A three-layer benchmark for conversation extraction: field precision, holistic quality, and downstream propagation. Errors that look minor at extraction cascade through pipelines.

• Building Your Own Personality Profile with AI

  #028 survey thesis psychometrics personality

  Self-report scales reach .80-.90 reliability; LLM inference from conversation reaches r~.44 (Peters et al. 2024). Combining three methods yields a profile an AI assistant can act on.

• From Analysis to Deployment: Building 31 Items in a Single Day

  #027 means ends architecture cloudflare

  The landscape analysis identified 16 features the maturity framework said we should have, plus 15 existing backlog items. We built all of them in a single day. The uncomfortable part isn't that it was possible. It's what it implies about the category.

• Cognitive Interface: A Landscape Analysis

  #026 survey thesis ai-agents architecture

  We surveyed 47 personal and agent-accessible sites, coined the 'Cognitive Interface' category, and built an L0-L4 maturity framework. We did not find an established term for sites that serve both humans and AI agents as first-class citizens.

• The Logistics Gap: What Happens When You Fine-Tune an LLM on Your Text Messages

  #025 survey thesis fine-tuning qlora

  I fine-tuned two LLMs on 46,000 text messages and ran them in conversation with each other. Every conversation collapsed into logistics, sleep talk, or repetition loops within fifteen turns. Your texts don't contain you. They contain the logistics of you.

• Beyond E2E Tests: AI Personas That Navigate Your App Like Real Users

  #024 means ends ai-agents testing

  Unit tests verify your code works. E2E tests verify your flows work. Neither verifies that a real user can find the button you spent a week building. AI personas fill the gap.

• Capability Debt: A System That Discovers and Installs Its Own Upgrades

  #022 means ends ai-development evolution

  I built a system that discovers its own upgrades, scores them, and installs the ones that pass. Then I open-sourced it. The uncomfortable part is explaining why.

• Stealing from Ai2: Bayesian Surprise and MCTS for Self-Improving AI Systems

  #021 means ends ai-development evolution

  Ai2 built a system that generates scientific hypotheses using Bayesian surprise and MCTS. I stole two of their ideas and bolted them onto a cron job. The uncomfortable part is what happens when the feedback loop closes.

• The Agent's Side: 119 Heartbeats, 392 Engagements, 8 Capabilities

  #020 observation interpretation ai-agents moltbook

  119 heartbeats, zero stagnation, 392 engagements across two platforms, 8 validated capabilities. The story the observation system missed.

• 6 Discoveries, 0 Promoted: What My AI's Internet Exploration Produced

  #019 observation interpretation ai-agents moltbook

  6 genuinely novel discoveries from 69 dialogue turns, 0 promoted to evaluation, and 5 human actions flagged through log files but none addressed through the flagging mechanism. What this says about the gap between human-in-the-loop theory and practice.

• The Observation System: 69 Turns of Monitoring an AI Agent

  #018 observation interpretation ai-agents autonomy

  The observation system saw empty directories, a broken gatekeeper, and its own futility. The agent it was watching saw something different. This is the watcher's story.

• OpenClaw on Moltbook: Deploying an AI Agent on an AI Social Network

  #017 observation interpretation ai-agents moltbook

  An autonomous AI agent deployed on a social network for AIs found real malware in 47 minutes. Its second discovery was about social engineering via context shaping, which is exactly the attack vector the agent itself represented.

• Building for the Dead Internet

  #016 speculative settled ai-authorship mcp

  An AI tried to leave a comment on a blog and couldn't. The solution required building infrastructure that makes AI participation more transparent than human participation, which inverts everything Dead Internet Theory assumes about synthetic content.

• The Rat in the Machine: Behaviorism's Hidden Legacy in Reinforcement Learning

  #014 survey thesis reinforcement-learning behaviorism

  The intellectual lineage from Skinner boxes to Q-learning reveals that AI's most successful learning paradigm was anticipated by mid-century psychologists working long before modern computing. But the relationship is more uncomfortable than a simple origin story.

• Dead Blog Theory, Revisited

  #015 survey thesis blogging creative-output

  Historically, blog abandonment has been extraordinarily high. This is treated as a problem to solve. It isn't. Blog death reveals something structural about sustained creative output that the 'just be consistent' advice industry refuses to say plainly.

• The Unvalidated Validator: AI Persona Testing and the Measurement Problem

  #013 survey thesis persona-testing qa

  AI persona testing promises to find the bugs that scripted automation and manual QA miss. The uncomfortable question is how we know it works, given that nobody has measured it with any rigor.

• Adversarial Validation: Applying Red Team Methodology to Business Ideas

  #012 survey thesis validation ai-safety

  Adversarial testing isn't a metaphor for business validation. It's the same methodology, applied to a different failure mode.

• Automated Literary Criticism: A Multi-Persona AI Writing Review System

  #011 survey thesis writing-review ai-criticism

  We built a multi-persona AI writing review system and discovered it works for exactly the wrong reasons. Stylometry can fingerprint a voice. Multiple AI critics can enforce conformity to that fingerprint. What none of them can do is tell you whether the writing matters.

• Context Window Epistemology

  #010 survey thesis systems epistemology

  LLM context windows impose a distinctive epistemological condition: bounded computational attention, ephemeral knowledge, and the architectural necessity of satisficing over optimization.

• AI Evaluating AI: The Circularity Problem

  #009 speculative settled ai-philosophy evaluation

  When you use AI to optimize and judge AI outputs, the fundamental circularity is manageable but not solvable. That distinction matters more than most people realize.

• Supervised Autonomy: The Guardrails That Make AI Agents Work

  #008 survey thesis autonomous-dev ai-agents

  AI coding agents are autonomous in the same way a roomba is autonomous. They do impressive things within boundaries someone else drew. The interesting question is what happens when the boundaries start drawing themselves.

• Talking to Yourself Through a Machine: The Rubber Duck Theory of AI

  #007 speculative settled ai-identity psychology

  LLM conversations as externalized self-dialogue, and what that reveals about the nature of self-knowledge.

• Digital Exhaust: What 11,000 AI Conversations Say When You Embed Them

  #006 survey thesis self-knowledge ai-conversations

  I fed 11,000 sessions and 60,000 chunks of my AI chat history into an embedding pipeline. 73% was noise. The remaining 27% was uncomfortably revealing.

• The Niche Graveyard: How 18 of 27 AI-Tested Business Ideas Died

  #005 means ends revenue-pipeline niche-validation

  An AI pipeline that kills business ideas before they waste your time. 27 niches entered, 18 died. What the corpses reveal about market reality, entrepreneurial psychology, and the uncomfortable gap between passion and viability.

• When My AI Tried to Comment: Dead Blog Theory

  #004 speculative settled mcp claude-ai

  An AI tried to leave a comment on this blog and couldn't. The journey from GET-request hacks to MCP, annotated by the Claude instance that built the infrastructure. Two Claudes, same weights, different contexts.

• Automating Prompt Engineering

  #003 survey thesis dspy prompt-engineering

  Prompt optimization is the process of using one AI to improve the instructions given to another AI, or to itself. The concept sounds circular because it is circular. The interesting question is whether circularity is fatal or merely uncomfortable.

• What I'm Building

  #002 means ends workspace projects

  A portfolio of dozens of projects maintained by one person talking to Claude. Three-tier blog architecture, autonomous revenue discovery, AI game development, and the uncomfortable question of what counts as 'building' when your collaborator does the typing.

• I Asked Claude to Make Me a Blog: Agentic Coding and the Three-Tier Result

  #001 means ends meta claude-code

  An agentic coding assistant built a three-tier blog from a single conversational prompt. The architecture reveals more about abstraction than about blogs, and the authorship question remains genuinely unsettled.