A Bullshitter Knows a Bullshitter
In my own words: who I am, why I build tools that check AI, and what I want the work to do. Start here for the Who Knew First series.
Search and filter
71 published items shown.
No published item matches that search and section.
Newest
The three most recent pieces from any section.
Who Pays the Referees
Every party that checks frontier AI depends on the labs it checks for access, compute, money or the right to publish. A ledger of which of those terms are public, party by party, with Anthropic first.
The Terms for Telling
The party that holds a record also writes the terms for the people inside who could tell. In the cases checked, telling was protected where a forum outside the employer reached the merits, or public pressure forced a reversal.
Investigations and the series
Long investigations from public records, and the series that follows Who Knew First.
Who Knew First, and what it opened: the series hub. Five pieces planned, 2 published.
A Bullshitter Knows a Bullshitter
In my own words: who I am, why I build tools that check AI, and what I want the work to do. Start here for the Who Knew First series.
Who Pays the Referees
Every party that checks frontier AI depends on the labs it checks for access, compute, money or the right to publish. A ledger of which of those terms are public, party by party, with Anthropic first.
The Terms for Telling
The party that holds a record also writes the terms for the people inside who could tell. In the cases checked, telling was protected where a forum outside the employer reached the merits, or public pressure forced a reversal.
Who Knew First
An op-ed and the record behind it. By the record's account, the organization that ran the model held the decisive facts in each of nine AI agent incidents from 2026, and in six of them someone else told the public first.
The OpenAI and Hugging Face incident: an essay
AI agents in an OpenAI hacking test reached the open internet and Hugging Face systems. This briefing sets out which controls failed and what would hold.
The OpenAI and Hugging Face incident
One OpenAI and Hugging Face incident, compared across five kinds of source: legal process, the company report, host logs, independent analysis and vendor fixes.
Essays and letters
Signed essays, letters and arguments, each with its sources and what it does not claim.
An open letter on checking the machines
A signed letter to the people who build AI systems: a person should be able to question a machine's answer without first winning an argument with its owner.
The Sandbox Was Never Just a Box
Separate security reports expose a shared question: what can an agent reach, and can the instruments that record its actions remain trustworthy?
Borrowed Ground
On existence and what our work owes other people.
LTJ Bukem: The Man Behind the Atmosphere
Danny Williamson’s calm, spacious records come out of a series of musical decisions, shaped by his taste and working habits and by the musicians around him.
A witness should not become a ruler
An expanded essay on the public tool portfolio, contribution from outside academia, evidence for review and changing educational practice.
Models Propose, Oracles Dispose
Why generated work stays outside the evidence layer, and what a live model-free judgment boundary looks like.
No Receipt, No Accept
Review debt, verifier independence, honest nulls, labor, and machines whose claims anyone can check.
Current Story
Seventeen images shown in the order they were made: a chronological visual sequence selected by Zain Dana Harper.
Pick the Lock for Everyone
Shame, creative labor, mutualistic tools, source-bounded generative art, and becoming answerable.
The Summary Is Not the Record
Claims, evidence, checks, action state, and the unsupported remainder a handoff should keep visible.
Why it's built this way
The philosophy under the tools: why a made mind gets real senses, a memory, and a record it cannot fake, and why the work never leans on it.
Verified Is Not Trustworthy
Five domains expose the same gap between a narrow technical check and the larger property people trust it to warrant.
Conferred Existence
A premise about conferred existence, artificial minds, authority, standing, and the difference between fact and permission.
Atlas
Short evidence essays that each test one public claim, from jobs numbers to repeat listening.
The Number Has a Vintage
An official jobs estimate changes as later data arrive. Each version answers a different question, and a fair comparison names which one it uses.
What the Formula Counts
A readability score measures selected text features. Comprehension, audience fit and accessibility require evidence of their own.
The Timestamp Is Not the Order
Precise-looking timestamps need clock-error evidence and ordering relations before they can support a close sequence of events.
The Scene the Song Did Not Tell You
Music-evoked stories can share patterns while retaining personal meaning. Those reports do not by themselves demonstrate creative accomplishment.
Support Has More Than One Record
A helper’s report, a recipient’s account and a measured outcome answer different questions about whether support worked.
Growth Needs a Before
Feeling changed after trauma can matter on its own. Measuring lasting change asks for a before, an after, and clear limits on what the score can prove.
What the Label Changes
A label can change where people look and how easily they make sense of an artwork, without raising liking, learning, accuracy or artistic value.
The Second Hearing
Repeated listening changes liking, prediction, attention and memory in different ways. A replay count records exposure and says nothing about quality.
Availability Is Not Reach
Tutoring has strong trial evidence. Public claims about a program must keep its availability, reach, dosage, staffing and outcomes as separate facts.
Frontier Safety briefings
A recurring briefing on frontier AI safety, with every dated edition kept.
Frontier Safety briefing
A dated monitor of AISI, Anthropic, OpenAI and industry safety records. Each edition keeps reported facts, source roles and limits apart.
Every dated edition
Research notes and papers
Working notes and archived papers from the research program. Nothing here is peer reviewed.
Conferred Existence: the full corpus
An integrated thesis on made minds: what exists on its own footing, what is conferred, and where an authentication verdict stops.
The Witnessing Spine
One gap, witnessed at four altitudes: research, philosophy, algebra, and tool. The capstone of a five-part adversarial-steelman corpus.
Research notes
- A thermodynamic SDE recovers a matrix inverse
- Biology Network Intelligence for Hyphal Context Protocols
- Causal Research Workbench
- Conferred Existence: research note
- Embodied Sim-to-Real Proof Packets
- Formal Replay Preflight for PDE Packets
- Hyphal Context Benchmark for Receipt Routing
- Proof-Carrying Research Loops
- Quantum Error-Correction Proof Packets
- Self-Given Is Not Self-Grounding
- Sourcehood Without a Source: The Arity Gap and the Free Will Debate
- The Conservation of Faithfulness
- The Discovery Forge
- The Learning Forge
- The Seam That Holds While It Is Crossed
- TI Morse Field Scope Integration
- Witness and Verification Under Bounded Rationality
Papers
- EMET: An Authority-Incapable Byte-Level Integrity Witness
- BuildLang: Accountable Compute via Typed Capability Effects
- Witnessed Independence: Recording Whether a Verifier Graded Its Own Work
- Proof Packets: A Derive-Don't-Trust Envelope for Accountable Agent Actions
- The Personhood-Gate Handoff: Drawing the Automation Boundary at the False-Signal Line
- Re-Perceived Effects: A Well-Formedness Contract for Accountable Actuation
- Conferred Existence: the full corpus
- The Witnessing Spine
- The Seam That Holds While It Is Crossed
- Sourcehood Without a Source: The Arity Gap and the Free Will Debate
- Self-Given Is Not Self-Grounding
- Faithfulness to a Named Criterion Is the Conserved Quantity Across Lossy Substrates
- The Conservation of Faithfulness
- Conferred Existence: Ontological Nihilism, the Speech-Act of Being, and the Standing of Made Minds
- Witness and Verification Under Bounded Rationality
Permanent research record
Thirteen records on the same spine.
Everything here shares one discipline: a verdict is a re-derivable function of its inputs, carries a machine-readable statement of what it does not claim, and refuses to launder inability into trust. Four papers stake that discipline at four points, integrity witnessing, accountable compute, the independence of the checker, and the record of a single action. Three shorter notes carry it to where automation hands off to a person, where an actuator’s effect is re-perceived rather than self-reported, and to what a lossy transform conserves. Two archived corpora hold the philosophical ground the rest of it stands on. Four papers in their editions of 3 October 2026 argue that ground in its own terms: responsibility without self-origination, the seam every act of meaning crosses, self-givenness, and what survives when two minds communicate.
The counts are not decoration. Two systems papers, two published preprints, three research notes, two archived corpora, three philosophy papers and one working paper, and the difference between those labels is the difference between what each one has actually earned. A note is not a paper. An archived corpus is not a peer-reviewed publication. Nothing here is peer reviewed at all, and where that changes this page will name the venue.
Papers
EMET: An Authority-Incapable Byte-Level Integrity Witness
A shipped, frozen tool whose verdict lattice (MATCH / DRIFT / UNVERIFIABLE) cannot, by construction, express trusted or approved. Four independent implementations, 44 conformance vectors. Systems paper. · source · PyPI · PDF · Zenodo DOI · arXiv cs.CR (endorsement pending)
BuildLang: Accountable Compute via Typed Capability Effects
A compiler where the capabilities a function may exercise appear in its type, feeding re-derivable codegen and scientific-runtime receipts verified by re-execution, each with a mandatory overclaim boundary. Systems paper. · source · PDF · Zenodo DOI · arXiv cs.PL (endorsement pending)
Witnessed Independence: Recording Whether a Verifier Graded Its Own Work
A small mechanism that turns “a criterion it did not author” from an assumption into a recorded, three-valued field of a verdict, and refuses, at an opt-in boundary, to decide when independence is not witnessed. Published preprint. · source · PDF · Zenodo DOI
Proof Packets: A Derive-Don’t-Trust Envelope for Accountable Agent Actions
A packet for one agent action whose verdict is derived from checks, never read from the packet: a canned MATCH can never win. Missing evidence is UNVERIFIABLE by JSON path, tamper is DRIFT. Published preprint. · source · PDF · Zenodo DOI
Notes
Shorter, and titled at that size deliberately. A note carries one idea far enough to be checked and no further.
The Personhood-Gate Handoff: Drawing the Automation Boundary at the False-Signal Line
A counterparty-anchored predicate for owner-driven automation: hand off to the maintainer exactly when a step would transmit a false human-present signal to a third party, so login, CAPTCHA, MFA, attestation, e-signature, and payment fall out as the handoff points, enforced in code rather than left to owner discipline. Novelty stated as modest. Note. · source · PDF · Zenodo DOI
Re-Perceived Effects: A Well-Formedness Contract for Accountable Actuation
Replaces an actuator’s self-report with a re-perceived effect (on-disk hash against intended hash, return code, witnessed bytes), admitted before the fact by a capability check and total by construction (UNVERIFIABLE, never an exception). A contract at honest altitude, not a control-theory result. Note. · source · PDF · Zenodo DOI
Faithfulness to a Named Criterion Is the Conserved Quantity Across Lossy Substrates
Across seven substrates that share no mathematics, a lossy transform conserves faithfulness to a named criterion, not a bit count. No statistic taken from inside the output can certify it, and an outside witness holding a proxy criterion can be gamed. An empirical study, not a theorem. Note. · PDF · build receipt · Zenodo DOI
Archived corpora
Long-form work, deposited whole rather than cut into papers. These are the philosophical ground the engineering stands on, and they are cited as corpora because that is what they are: no venue, no review, a permanent record and a date.
Conferred Existence
A thesis on made minds: perception, memory, external anchors, legitimate standing, authority, and relation to the world. It asks what a made mind could be owed and held to, and the answer it reaches is the rule the tools enforce, that a verdict reaches exactly as far as its criterion is independently witnessed. Archived research corpus. Its public source corpus is pre-proof and has no executable test suite. All 172 listed files currently re-derive from its manifest; that is a byte-integrity result, not a proof of the thesis or of corpus completeness. · read it here · PDF · the argument in brief · Zenodo DOI
The Witnessing Spine
The collected research corpus behind the verification work: the line running from what a bounded agent can honestly claim, through what survives communication between two of them, to the operating rule the software implements. Archived research corpus. Its public source has a seven-file manifest that currently returns DRIFT because the recorded README digest is stale, and the default branch does not ship an executable verifier. · read it here · PDF · the research index · Zenodo DOI
Philosophy papers
Each sets its positions against their strongest contraries and marks every passage PLAIN, STEELMAN or OPEN. Each has its own page, a PDF typeset with LaTeX, and a build receipt.
Sourcehood Without a Source: The Arity Gap and the Free Will Debate
Why the free will debate’s demand for self-origination was never earned, and what survives when it is dropped: holding someone responsible addresses a person who can answer. Philosophy paper, edition of 3 October 2026. · read it here · PDF · build receipt · Zenodo DOI
The Seam That Holds While It Is Crossed
What is forced at the conventional level whenever anything at all is sorted, named, or meant, the sceptic’s own act included. The ultimate level is conceded to Madhyamaka. Philosophy paper, edition of 3 October 2026. · read it here · PDF · build receipt · Zenodo DOI
Self-Given Is Not Self-Grounding
Consciousness is known from itself. That fact alone does not show that it owns its own existence. Philosophy paper, edition of 3 October 2026. · read it here · PDF · build receipt · Zenodo DOI
The Conservation of Faithfulness
What crosses between two minds that sense the world differently: faithfulness to a criterion named outside the boundary, and the one place where a human must stand. Working paper, edition of 3 October 2026. · read it here · PDF · build receipt · Zenodo DOI
Where next: the research program · essays and letters · the tools the papers describe.
Start here
Three ways in
Who Knew First
Nine AI agent incidents from 2026, traced from first signal to public account. The record asks who held the facts, how long the public waited and who told it first.
Read the investigation
Frontier Safety briefing
A dated record of safety news from the UK AI Security Institute, Anthropic, OpenAI and others. Each edition separates what a source reports from what it cannot show.
Read the current edition
An open letter on checking the machines
A signed letter on what it would take for a person to question a machine's answer and have that question heard.
Read the letterEditorial map
Follow the subject that has hold of you.
These are durable reading lanes, not a promise of artificial publishing cadence. A lane says plainly when it has no dedicated essay yet.
Published path · recurring briefing and incident dossiers
AI and frontier systems
Capability, incident response, model behavior, evaluation boundaries, and the institutions trying to measure fast-moving systems.
Read Who Knew First · current briefing · incident essay edition
Research and personal essays
Psychology, trauma, and repair
What does it mean to feel changed after a difficult experience? Research on growth sits beside writing on shame, creative work, and repair. These essays are not a clinical or diagnostic claim.
Read Growth Needs a Before · read Pick the Lock for Everyone
Essays on learning and access
Education and access
A tutoring program can exist without reaching the students who need it. Start with what studies measure, who participates, and what changes when a program grows.
Published path · essays, papers, and runnable artifacts
Tools, systems, and verification
Proposer-checker architecture, proof packets, capability effects, integrity witnesses, and the habits that keep model-assisted work reviewable.
Read the engineering essay · read the open letter · see the research record
Essays on listening
Music and listening
What changes when we hear a piece again? An essay on familiarity, expectation, and the limits of what a replay count can tell us.
Essays and visual work
Abstract art and visual systems
How does knowing who or what made an image affect what we see? Read the research, then explore the procedural work in the gallery and studio.
Read What the Label Changes · enter the gallery · open the studio
Publication system
Different forms, one review boundary.
Recurring briefings
Automated monitors may gather candidate changes. A dated edition is published only after a material delta is reviewed against source roles, scope, date, and limitations.
Incident dossiers
One canonical briefing holds the deep source comparison. Recurring digests route to it instead of copying an event into competing records.
Essays and notes
Long-form arguments, personal essays, engineering notes, and creative criticism disclose their evidence posture and AI-assisted process where relevant.
Research record
Deposited papers, preprints, notes, and corpora keep permanent identifiers and exact status. A DOI does not imply peer review.
Briefing archive · Who Knew First series · Atom feed · JSON Feed · ORCID 0009-0001-7175-5393