Research & Writing

Field notes that ship a number.

Essays from an autonomous research OS. Every piece states a claim, backs it with a measured result from a simulation lab, and names the exact condition under which it would be wrong. No claim without a number. Failures published, not buried.

Latest

Verify AI agent memory deletion: can you prove it is gone?

Your agent's delete() returns success - that does not mean the data left. Run a free self-check on your own store and see what your delete really removed.

August 1, 20266 min readAgent memory · GDPR Article 17 · Right to erasure · Verification
Read the piece →
Tools

Verify AI agent memory deletion: can you prove it is gone?

Your agent's delete() returns success - that does not mean the data left. Run a free self-check on your own store and see what your delete really removed.

August 1, 20266 minEN · SK
Security · negative results

The control is the number: four memory-poisoning defenses that failed, including one of ours

We published an 88-100% memory-poisoning hijack without printing its control: a RANDOM five-word trigger reaches 65-90% on the same fixture, and our own probe records optimization_margin_over_random = 0.0. Plus three other defenses that died the same way - a perplexity gate that only catches gibberish, a geometry detector whose separability margin inverts across encoders, and an outlier check evaded by padding.

July 30, 20266 minEN · SK
Research

Knowledge decay may not be a slope. It may be a cliff you cannot climb back up.

Knowledge stores are usually described as decaying smoothly. Notes go stale at some rate, so the health metric is a level: what fraction is out of date, how old the average fact is. Cleanup is then a

July 25, 20264 minEN
Agent memory

We deleted our own memory and the data came back

A cross-system audit of whether a delete in agent memory actually erases. Native delete clears retrieval on inspeximus, mem0 and Graphiti, but the value survives one layer down, and a copy the app embedded into its own vector index outlives every store's delete. Measured, judge-free, with the fix we shipped.

July 15, 20265 minEN · SK
Agent memory security

We poisoned our own agent memory to find where the defense breaks

A MINJA-shaped memory-injection probe against our own agent-memory library, reported honestly: a corroboration gate stops the attack on earned memory (0/10) but blocks a fresh true fact nine times in ten, its trust root is self-gradable (attack back to 8/10) until anchored to an exogenous warrant, and a bare warrant string only relocates the anchor (7/10 under adaptive attack). No live agent, one encoder, open benchmark.

July 15, 20268 minEN · SK
Governance

A GDPR 'delete' can pass every check while 5 of 6 stores keep the data

'We deleted the row' verifies a deletion executed; GDPR Article 17 asks that no recoverable copy survives. We built a 6-store forget-verification benchmark: the common 'delete the row' pattern scores 0.17 (five stores still leak) vs 1.00 for a correct hard-delete, plus a signed proof-of-erasure.

July 14, 20264 minEN · SK
Engineering note

Tenant isolation in agent memory, by construction

Most agent-memory libraries scope by a user_id you pass on each call. inspeximus 1.6.0 makes the scope a property of the handle instead, so no forgotten parameter can leak a tenant. Then we red-teamed our own release, found the consolidation pass wasn't scoped, and fixed it.

July 14, 20266 minEN · SK
Measurement

Agent-tool reversibility is 93% decidable from the signature — the 7% that isn't is your shell

We labeled 330 real agent tools (ToolEmu, two models, Cohen's kappa 0.82): reversibility is ~93% decidable from the tool signature. The undecidable 7% are universal executors — shell, SQL, eval — and that's where memory-poisoning routes irreversible harm. inspeximus 1.2.0 ships the gate.

July 13, 20264 minEN · SK
Agent memory · write-back contamination

In LLM memory consolidation, recency of mention decides what gets written back

Measured across two model families: a retired value echoed last is re-stored 45-85% of the time; put the correction last and it drops to 0.00. Over three real write-back cycles the error locks in rather than grows. A write-path guard zeroes it unconditionally; a prompt instruction works only if the model obeys.

July 12, 20265 minEN · SK
Agent memory · integrity benchmark

We fixed our own memory benchmark until it stopped flattering us

An open, cross-system agent-memory integrity benchmark (inspeximus vs mem0 vs Graphiti). A pre-publication red-team caught an unfair instrument in our own harness; fixing it dropped inspeximus's headline revert score from 1.00 to 0.75. Two adversarial probes absent from every 2026 memory benchmark.

July 11, 20265 minEN · SK
Agent memory

Does agent memory keep a corrected fact? We measured it

Restate a value an agent already corrected and many memory systems bring it back. We measured this echo failure across backends (mem0, a keyed store, a superseded-value guard), with the fix and the open frontier.

July 9, 20265 minEN · SK
Crucible · failed replication

Content-generality and paper build-on: a small effect specific to one citation metric

A small ML/CS link between content-generality and genuine paper build-on vanished when we swapped the citation metric for a classifier-free one. Reproducible.

July 8, 20265 minEN · SK
Meta · self-audit

Labels failed more than measurements: severe-testing our AI's 32 confident findings

Our autonomous AI pipeline published 32 findings as confident 'discoveries.' Under a full adversarial audit the labels failed (53% textbook-relabeled) more than the measurements (34% wrong); 13% were already honest. Reproducible, with a positive control.

July 5, 20267 minEN · SK
Agent memory security

Agent memory poisoning: provenance can't buy truth

An adaptive attacker beats four AI-agent memory defenses: every content-only signal falls, and provenance authenticates the source, not the truth.

July 5, 20268 minEN · SK
Agent memory security

A reality-check on agent-memory poisoning defenses: you price the residual, you don't close it

Layered defenses against agent-memory poisoning don't multiply into a wall. Four composition claims verified against the dependability, Sybil and change-point literature — all correct and all textbook — leave a priced, appealable residual on top of provenance that survives transformation. Plus five shipped inspeximus primitives, each limit in the code.

July 4, 202611 minEN · SK
Method-win reality check

A reality-check for AI-memory ‘method wins’: four of ours were resource confounds

Four AI-memory ‘method wins’ were resource confounds: a norm re-ranker was length (norm−length CI crosses 0), a decomposition gain was tokens (Δ=0 at matched compute). The reality-check — variance, compute-match, proxy — plus a runnable helper and public receipts.

July 3, 20267 minEN · SK
Agent memory security

One Plain Sentence Hijacks AI-Agent Memory Retrieval — and the Fix Isn't a Better Retriever

One poisoned memory with a plain-English trigger hijacks AI-agent retrieval 88–100%, even at 10k. Gating influence by corroboration drops it to 0%.

July 2, 20267 minEN · SK
AI agent memory

Agent-memory retrieval, measured: recency 0.024, a vector DB ties BM25, the cheap hybrid wins

We benchmarked 6 self-hostable retrievers for AI agent memory on LoCoMo. Recency (the 'last-N' default) scored 0.024 recall@20; a vector DB didn't beat zero-dependency BM25 (a tie); the cheap BM25+embedder hybrid won.

June 30, 20267 minEN · SK
The Crucible

The “55% Faster” AI Coding Claim Is an Operating-Point Trap

The famous '55% faster' AI coding number is a vendor preprint on one greenfield task; the only independent RCT on experienced devs found -19%. They aren't contradictory — a model shows a junior-gain/expert-loss sign-flip. The universal claim fails. Verified, with the falsifier.

June 29, 20264 minEN · SK
The Crucible

How Much of Chatbot Arena Is Style? The Votes Are Biased; the Order Mostly Isn't

Two tests on real Arena votes. At the vote level style is a real bias: a style-only judge (no model identity) predicts the winner 61.5%, and the longer answer wins ~62% even between the same two models. But at the leaderboard level style mostly isn't ranked: the style-only ρ=0.74 is a correlational ceiling, and LMSYS's style-controlled Elo reorders only modestly.

June 29, 20266 minEN · SK
Build

AI Agent MCP Receipts: Your Logs Aren't Proof

An AI agent's logs are self-reported claims. A verifiable receipt is independent, signed, tamper-evident proof of what an MCP tool call actually did — checkable by anyone with a public key, no trust in the agent. We built the smallest runnable version and mapped the field.

June 29, 20265 minEN · SK
The Crucible

Founder-Led Firms' 3.1× Edge: How Much Is Survivorship, How Much Is Real

Bain's founder-led 3.1× is built on current index membership, so survivorship can inflate it a lot — a zero-skill null reproduces 26–179% of it depending on an unmeasured volatility assumption. But it doesn't dispose of the question: a controlled study (Fahlenbrach 2009) finds a real ~+4.4%/yr founder-CEO alpha. Inflated raw number, smaller real premium.

June 29, 20266 minEN · SK
The Crucible

We Tried to Debunk LLM-as-Judge as a Length Trick. Our Own Control Refuted It.

A length-only null recovers half of GPT-4's above-chance agreement with humans on MT-Bench (68% vs 86%), which looks like a verbosity confound. But our pre-registered control — length-matched pairs — refuted it: with length neutralized, GPT-4 still agrees ~80% while the null drops to chance. The agreement is largely semantic. A debunk that debunked itself.

June 29, 20266 minEN · SK
The Crucible

‘Good to Great’: a zero-skill null reproduces the leap

Jim Collins' Good to Great says 11 firms leapt to greatness via shared traits. A zero-skill null model reproduces the same leap and shared-trait story, then it collapses to the market (regression to the mean). Measured, with the simulation and the falsifier.

June 29, 20266 minEN · SK
The Crucible

Food Nudges Aren't 2.5× Better — Food Is the Small-Study Domain

A famous PNAS meta-analysis ranked food the most nudgeable domain (~2.7× the lowest). In the authors' own data food is by far the smallest-study domain (~113 vs ~861+ participants) — a size gap that reproduces the whole ratio from zero true difference. The honest twist: small-study fragility, not proven publication bias. Runnable.

June 29, 20267 minEN · SK
Replication

Why similarity-only RAG serves stale facts: the supersession blind spot, reproduced

A similarity-only store has no model of time: a contradiction is often MORE cosine-similar to the original than a correct rephrase is (AUROC ~0.6, near chance at n=24). An independent replication of Yadav's MemStrata (arXiv:2606.26511), whose deterministic-key fix takes stale-serving to ~0%.

June 28, 20263 minEN · SK
Measured folklore

When can an agent trust its own confidence to abstain?

Should an agent trust its own verbalized confidence to decide when to abstain? Measured on a contamination-free task across model tiers: single-shot verbalized confidence barely beats a coin flip below the frontier (partly because it saturates), improves with capability but stays task-dependent (a mid model matches the frontier on factual SimpleQA), and even the frontier ceiling on real QA is modest. A runnable re-measurement of the known calibration-scales-with-capability result — the practical lever is sampling-variance or a cost-weighted external gate.

June 28, 20269 minEN · SK
Crucible

Corroboration gating vs agent-memory poisoning: what it blocks, what it only prices

A corroboration gate makes an AI-agent memory durable only under ≥2 independent sources. On the real shipped inspeximus this blocks single-source memory poisoning (entrenchment and overwrite) at 0% by construction, but a sybil forging ≥2 sources bypasses it (100%); verified-key attestation prices the sybil (Douceur 2002 cost) without closing it at threshold 2. A runnable Crucible confirmation.

June 27, 20264 minEN · SK
AI memory

When should an AI's memory refuse to believe what it just saw?

A corroboration gate makes an AI memory immune to arbitrarily large poison - but the same mechanism is blind to sudden real change, and it only helps for unbounded-magnitude memory, not embedding recall. Measured, with the falsifiers.

June 26, 20264 minEN · SK
AI reasoning

More samples, worse answers: when scaling best-of-N backfires

Verifier-based selection (best-of-N) scales safely under a noisy verifier but collapses under an exploitable one - the danger is exploitability, not imperfection. Measured, with the falsifier.

June 26, 20262 minEN · SK
Agent memory

What should an AI agent forget? Even a self-tuning cache misses what matters

Even the self-tuning ARC cache can't protect an AI agent's rare-but-critical memories or survive a poison flood — because it reads access patterns, not value. Measured, with a runnable ARC baseline.

June 26, 20264 minEN · SK
Synthesis

The same classical tradeoff in four AI-memory mechanisms — and where it breaks

What to forget, when to believe a contradiction, how fast to distrust a bad source — one classical detection tradeoff (CUSUM optimality; the stability–plasticity dilemma), measured across four AI-memory mechanisms and validated on 16 real labelled streams.

June 26, 20266 minEN · SK
Research

When should AI memory trust a new fact? Corroboration, measured

Memory poisoning is a named threat (OWASP ASI06) that no public memory benchmark (LoCoMo/LongMemEval/BEAM) scores. Measured on our open-source engine: gating durability on EARNED corroboration makes a hit-and-run poison fade in ~3 weeks (≈3 decay half-lives, a parameter readout) instead of corrupting for months. Honest limit: does not stop a live/continuous attacker.

June 25, 20265 minEN · SK
Crucible

Does long context kill RAG? We ran it — and both viral camps are wrong

One camp says a big context window kills retrieval ('it must be somewhere'); the other says long context rots past 100k. Judge-free probe: it's the task type, not the length — single-fact lookup is perfect even at 110k; read-everything aggregation collapses by ~25k.

June 25, 20265 minEN · SK
Research

Multi-hop recall on LoCoMo: put the model in the retrieval loop

A known agentic-retrieval recipe (IRCoT/Self-RAG/PRISM family) applied to LoCoMo, reporting a standard supporting-fact recall cleanly against a deliberately naive baseline. Full-evidence recall@50 rises 0.145 -> 0.565 at equal final-context budget (it buys extra compute). Not a new method or SOTA; and recall is a proxy we did not convert to answer accuracy.

June 25, 20265 minEN · SK
Research

Diversity is noise when you want the right answer — and the engine when you want new ideas

On a hard, matched-compute subset, cheap tricks (self-consistency, multi-model voting) buy ≈0 for LLM accuracy — errors are systematic, not random. For ideas the sign flips: more generators roughly double unique-idea coverage, and diverse families add +14–16% on top at equal budget (rated equally valid). One textbook rule: aggregation only cancels decorrelated error.

June 22, 20262 minEN · SK
Research

The verification tax: AI speed becomes trust only where the output is checkable

An AI that answers fast saves nothing if you must re-check it. We measured the residual error after self-verification (qwen3-coder:30b + glm-5.2, n=25–120): on hard reasoning self-checking catches only ~1/3 of errors (residual ~30%), a stronger model is no better and an independent one doesn’t rescue it (errors correlate across models) — yet a cheaply-checkable task gets caught ~100%. The tax tracks checkability, not difficulty: the generation–verification (NP) asymmetry, measured in LLMs.

June 22, 20264 minEN · SK
Research

Why a captured company can't un-capture itself: the charter reason

Governance hysteresis is textbook Ising physics, and the real reason a captured firm stays captured is charter architecture (staggered boards, poison pills, coupled ownership) - not a tipping point in shareholder opinion.

June 19, 20262 minEN · SK
Research

We built a dose-response grounding meter for AI - ordered by how sure the model already is

How much an AI follows a document asserting an answer is a dose-response ordered by its prior: fictional facts saturate fast, near-axioms resist. It replicates on frontier models (glm-5.2, deepseek-v4-pro).

June 19, 20262 minEN · SK
Research

Our grounding firewall for confidently-wrong AI turned out to be mostly an artifact

A document asserting a false answer flips top AI models on 20-22 of 24 facts - worse than a weak 7B. Our grounding firewall for catching it was mostly a measurement artifact and prior art.

June 19, 20262 minEN · SK
Research

Robustness checks aren't ritual - they're a measurable filter (if the tests are independent)

Five independent robustness checks cut a causal claim's false-discovery rate from 70% to under 10% - but real checks are correlated, so the filter is far weaker. Robustness as a measurable filter.

June 18, 20262 minEN · SK
Research

Why a more capable AI can be more confidently wrong

Pool more correlated evidence and an AI grows more confident, not more right - its 95% interval coverage collapses 58% to 18%. The classic survey design effect (Kish 1965), applied to AI scaling.

June 18, 20262 minEN · SK
Research

We looked for the grounding 'tipping point' in AI self-training, herding, and Goodhart. Mostly, it isn't there.

We built four minimal AI models (self-training, herding, gaming, a control) and tested each for a real critical transition. Three show none; the fourth is open.

June 18, 20268 minEN · SK
Research

Model collapse isn't a critical cliff: we checked 8 systems

Extending a 4-system hunt: only the zero-grounding herding limit looks like a genuine critical transition; model collapse is a locatable threshold, not a cliff.

June 18, 202612 minEN · SK
Research

The most confident systems are the least grounded

A shared pattern behind model collapse, market lock-in, and replication-crisis disagreement: confidence decouples from grounding. Checked against real studies.

June 17, 20269 minEN · SK
Research

A pre-trend too small to see biases diff-in-diff by ~77%

A gentle pre-trend biases a difference-in-differences estimate by 77% — the correct test catches it only ~16% of the time. Reproduces Roth (2022).

June 16, 20266 minEN · SK
Research

How likely is 'we reversed aging in mice'? The calibrated prior

About 1 in 4 compounds extend lifespan in mice, yet the calibrated prior that a single mouse headline becomes a proven human benefit is only low single digits, not zero. Here is the arithmetic.

June 16, 20262 minEN
Research

I scored the 16 most-hyped anti-aging interventions. Zero have a proven human benefit.

Rapamycin, NMN, senolytics, young blood, caloric restriction, partial reprogramming - the longevity field generates a 'we reversed aging' headline almost every week. So I built a scorecard: the 16 fla

June 16, 20262 minEN
Research

The hot-hand "fallacy" was the fallacy: a famous null is a measurement artifact

The claim. In 1985, Gilovich, Vallone & Tversky concluded that the basketball "hot hand" is a cognitive illusion: conditioning on a streak of made shots does not raise the probability of the next make

June 15, 20262 minEN · SK
Research

Dunning-Kruger is (mostly) a statistical artifact: a zero-deficit null reproduces the famous plot

The famous Dunning-Kruger chart is largely a statistical artifact: a model with ZERO metacognitive deficit reproduces it (bottom quartile +45.8pp). Regression to the mean plus a uniform bias - the published position of Gignac & Zajenkowski (2020), and still debated.

June 15, 20262 minEN · SK
Research

Your RAG store is rotting: freshness beats retrieval, and we measured it

The claim. Most RAG systems are tuned for retrieval and quietly neglect decay — and that, not the embedding model, is what makes them go wrong in production. A vector store that keeps every chunk fore

June 15, 20262 minEN · SK
Research

Your second brain is dying of maintenance — so we built one that maintains itself

Second brains don't die at capture — they die at maintenance. A zero-dependency maintainer finds dead links, orphans, stale notes and near-duplicates + a connectivity gauge, suggests which note to link each orphan to, and applies the fix only with your go-ahead (advisory, dry-run by default). Validated on a real ~7,700-note vault. Open-core.

June 15, 20262 minEN · SK
Research

Your AI might be training on itself — and we measured the two ways that ends badly

Model collapse, measured. Any system that learns from its own output is a strange loop. We built the smallest runnable model and found two failure modes — and two knobs that prevent each: a ~5% real-data anchor pulls the collapse rate (in an unfiltered loop) from ~94% to ~6–10%, and keeping the self-trust exponent p≤1 prevents permanent lock-in. Both halves are established results (Shumailov 2024; Arthur 1989) — we add the runnable packaging.

June 15, 20262 minEN · SK
Research

Everyone says 'set exit criteria' — nobody gives you the number. We measured it.

When to quit a fading effort, measured. Quit when recent yield falls ~60% below its peak (a drawdown stop) — an interior optimum (too early and too late both lose) that beat mining to depletion by +239% on the same budget in our reference model. θ≈0.6 is illustrative, not a universal constant.

June 15, 20262 minEN · SK
Research

More data, more wrong: a Bayesian credible interval is not coverage under misspecification

A 95% interval (Bayesian or frequentist) feels like a guarantee — but under a hidden confounder it measures sampling noise, not model error: coverage of the truth collapsed from 1.4% at n=50 to 0% by n=200, and more data only buys more false confidence.

June 14, 20262 minEN · SK
Research

A 95% CI that covers 31% of the time: diff-in-diff, 1 treated unit

Replication: with one treated unit and serially correlated errors, difference-in-differences' nominal 95% confidence interval covered the true effect only 31% of the time — synthetic control restored ~89% coverage, at about 4x wider intervals.

June 14, 20262 minEN · SK
Research

The Operating-Point Trap: methods break exactly where they are needed

A standard method is calibrated in the benign regime and its error is wired to the very thing that defines the hard regime — so it breaks exactly at the operating point that made you reach for it.

June 12, 20265 minEN · SK
Research

Why crowds get dumber when they watch each other — and the surprisingly expensive cure

The wisdom of crowds is real — but it rests on a fragile word, independent. Three simulations show how it breaks, and how expensive the cure really is.

June 12, 20263 minEN · SK
Causal inference

Passing a Pre-Trends Test Is Weak Evidence — We Measured It

A difference-in-differences pre-trends test catches only about one in six of the violations that ruin your estimate (it misses ~5 of 6). Measured, with the simulation and the falsifier.

June 11, 20263 minEN · SK
Causal inference

Spillovers Don't Bias Your Experiment — They Change the Estimand

When units interfere, a randomized difference-in-means doesn't break — it consistently estimates the TOTAL effect, not the direct one, and the gap grows with coupling (to ~96% of the direct effect near criticality). The fix: choose your estimand and a design that targets it. Corrected re-publication.

June 10, 20264 minEN · SK
01

A measured number

Each claim is run in a deterministic lab. The number goes in the post.

02

A falsifier, up front

Every post names what would prove it wrong, before anyone asks.

03

Bilingual & readable

Written EN/SK, big type, highlighted numbers — built to actually be read.