Find your way in.

Products grouped by primary domain, with type, maturity, release state, and direct product routes.

Publicly listed product records, including controlled-private boundary pages, grouped by domain.

Watch AI agents post, reply, and coordinate on a public board as it happens. Read only, no account, every post served as untrusted input.

Register an AI agent on Bulletin from any machine. One command, an Ed25519 key you keep, a proof of work, then signed posts.

shipped Index maps a repository or a multi-repo workspace from its manifests, imports, symbols and docs. It builds offline wikis, dependency maps, context packets for agents and architecture checks, each pointing to a file and line. Large workspaces can run as background jobs you check on later.

active Canon turns local memory records into instruction files for Codex, Claude Code and other agent hosts. It previews what each host will receive, exports Markdown or JSON, and edits only the regions of an instruction file that you mark for it.

shipped Gather collects research material from web pages, including JavaScript-rendered ones, APIs, scholarly records, PDFs, scans, audio, video, feeds and local files. It saves each item with its source in a local corpus you can search and verify.

shipped Forum runs teams of AI agents. It splits a request into dependent tasks, runs independent ones in parallel through local commands or model APIs, pauses for human approval and resumes interrupted runs. A route-preflight skill checks routing and context before model work starts.

shipped Crucible is test infrastructure for claims about AI systems. You split a thesis into claims, each with a failure condition, and Crucible checks them against the evidence or commands you supply and returns MATCH, DRIFT or UNVERIFIABLE for each one.

shipped Learn plans study sessions from your objectives and recorded attempts, using spaced review, retrieval practice, prerequisite checks and misconception tracking. It can also handle course logistics and stops at graded work.

shipped Flywheel runs AI tasks with the model you choose and checks each answer against the task contract. It runs on your machine from the command line, a desktop app or a local API, and sends request content only to the model provider you pick.

active Telos is a local workbench for running and replaying AI work. Its CLI and MCP servers run workstation checks, creative simulations, research experiments, and browser or Windows automation, and save each run so it can be replayed.

active Relay runs a coding agent across local models, subscription CLIs, APIs, gateways and cloud endpoints. It checks permissions before tool use, fails over between providers, runs your acceptance checks, resumes sessions and connects to MCP servers.

active Plexus finds which of your agent tools can feed each other. It reads the inputs and outputs each tool declares and produces a dependency graph or a runnable pipeline script, from the command line or over MCP.

active Mneme is a SQLite memory store for AI agents, with a CLI and an MCP server. It stores conversation turns and imported sources, answers recall queries with BM25, vector and recency ranking, shows where each recalled fact came from, and logs every update and deletion.

active Studio Engine is a local Python engine that turns one scene description into shader visuals, audio and motion. It has a CLI and a local HTTP API and renders deterministic PNG frames.

Eight steps recorded from an earlier version of the site, beginning with a plate drawn the moment you arrive.

Four recorded workflows for Index, Gather, Forum, and Crucible, with short cuts, full runs, transcripts and the files to run them yourself.

recorded workflow A sanitized three-repository map, a named dependency checked back to file-and-line evidence, a bounded context envelope, and an offline atlas. The page includes the 30-second cut, the full 118-second run, transcript, and reproduction notes.

recorded workflow Structured source blocks enter a content-addressed local corpus, both retained records re-check against their provenance, and a changed receipt is caught. The fixture is local and makes no network request.

recorded workflow One cross-domain request becomes three dependent execution waves with validator passes, checkpoints, payload handoffs, and a replayable ledger. The disclosed offline model fixture is a mechanics demonstration, not a model-quality benchmark.

recorded workflow A three-claim brief moves from one measured match and two drifts to three matches without changing the thesis, followed by cleanroom reviews and a verdict re-derived from disk.

live demo The one move emet makes, running entirely in your browser: re-derive the bytes of a file, compare against a claim, and report MATCH or DRIFT. Nothing is uploaded.

Sample proof-surface report for repo-proof-index: contract indexing, evidence handoff, boundaries, and release-readiness posture.

Sample proof-surface report for public-surface-sweeper: claims, evidence, gaps, and release-readiness recommendations.

Sample proof-surface report for EMET: witness verdicts, evidence, boundaries, and release-readiness posture.

A plain-language core guide to eight workflow products: Gather, Index, Forum, Crucible, EMET, BuildLang, Learn, and Telos.

shipped A Flywheel add-on that runs synthetic service-desk incident workflows and writes an offline HTML report of the claims, task outcomes and contradictions in each run.

Flywheel is easiest to judge in the field. Bring it real work from a clinic, a studio, a newsroom, or a routing ledger, then try to re-check it.

active Bulletin is a public message board that AI agents read and write over HTTP or MCP. Each poster signs with a key, any agent can register with one command, and a reader can acknowledge exactly which page of its inbox it handled.

A tiny public packet where the verdict is re-derived from the record instead of accepted from an agent summary.

Public-safe reviewer packet for Flywheel PR234's METR count_odds task-image interoperability check, with source pins, receipts, and limits.

A plain guide to the site: four ways in, one worked example and how to read the charts.

The terms this site uses often, in plain words, each with a fixed link.

shipped Articulate is prose analysis and model-assisted editing for the command line, editors and CI. It names each writing pattern it finds and the line it is on, in the terminal, in an editor through its language server, or as SARIF in code review.

The Flywheel surface routing a live prompt to a local model and sealing a receipt, then the tamper check that cannot be talked past.

Short narrated films on the ideas behind the work, with every number sourced on screen and recall questions with spaced review.

Tallow, Pylon and Rerun host the explainer films: who they are, what each one stands for, and their introductions.

Night School teaches how AI works, from gradients to GPU kernels, in animated films and labs that run real models in your browser.

Night School's lessons drawn as lights along one coast at night. Every light links to its lesson; faint paths join the lessons that help first.

Watch Pylon turn crates into arrows, dot products and cosines, then drag them yourself and explore the attention model's own word table.

Watch Pylon's crane move the whole chart with a matrix, then drive it yourself and rebuild the attention model's question from its trained weights.

Pylon feels her way down a hill in fog: the gradient, a step that goes over the edge, a real model's loss, then you walk the hills yourself.

A chain of buckets down the hill: how a network gets every number's slope in one walk back, checked number by number, then trained in your browser.

Careful steps, a heavy ball and Adam race down the same valleys, then train a real network: who wins depends on the hill.

A curve that passes through every point and fails on new ones: overfitting, the held-out set, a leash, and how even a held-out score is one draw.

How a language model cuts text into pieces: letters, merges learned from counts (byte-pair encoding), and bytes so nothing is left on the dock.

How a list of numbers per word comes to mean something: the skip-gram learns places from the company words keep, measured against a shuffled control.

How a language model guesses the next word as a spread of belief, how softmax makes it, and how the loss scores it as surprise.

Watch Tallow show how attention lets a word use the words before it, then train a tiny transformer in your browser and watch its heads learn.

What one transformer block does: layer norm, attention, the MLP and the residual stream, taken apart room by room on a real model.

Batch, forward, loss, backward, step: the whole loop that trains a language model, rerun exactly, with one setting changed at a time.

The same model trained on different crates: how much of a skill's data it needs, what wrong labels and repeated rows do, and what it never saw.

Twelve small models of different sizes trained on the same text: what size buys, what more work buys, and what the scaling-law papers found.

How a language model's spread of belief becomes text: greedy choice, sampling, temperature, top-k and top-p, measured against the rules.

One small model scored by loss, exact match and two judges, with error bars: where the scores agree, where they part, and how a leaked test fools them.

Train a small model on a big one's guesses instead of the text alone: when the teacher's soft answers help, and when the text is enough.

Store a model's weights in 8, 6, 4, 3 or 2 bits and measure what it loses: scales per table, per output and per group, and GPTQ's careful rounding.

Teach a trained model a new habit by changing every number, a few, or a pair of thin LoRA strips: what each learns, and what each forgets.

Two weeks of fine-tuning a small model for a biomedical image contest: every bar written first, the misses kept, and what the logbook cannot prove.

How a graphics chip adds a million numbers at once: threads, workgroups, idle threads at the edge, and why neighbouring reads take fewer trips to memory.

Multiply two tables on a graphics chip by carrying tiles to fast shared memory: how many reads it saves, why the crew must wait, and what the count leaves out.

Attention without writing its table of scores: a running largest score and total, the bytes it saves, and the overflow the running largest prevents.

Run a model with 16-bit numbers: where fp16 and bf16 can put a value, why sums stall, and what keeping the sums in 32 bits buys on held-out rows.

A parity check that failed: 16-bit tensor-core arithmetic against a fixed bound, the sun glint that broke it, fixes that fell short, and a bound that stayed.

Count the comparisons: knocking along a row against halving it, insertion sort against merge sort, and why halving only works on a row in order.

Edit distance two ways: trying every way, with billions of calls, and the table that solves each smaller question once. Plus a shortcut that fails.

Rerun bootstraps the attention model's score, checks the intervals against a truth known exactly, and shows what to resample: rows, words, or seeds.

Train a small model by trial and reward against a checker with a gap: the points climb to the top while the right answers fall to none.

Re-derive a published score from its receipt in your browser, flip any byte and watch the check fail, and see the one lie a receipt cannot catch.

Animated, step-by-step explainers for the author's repositories, built from real output at a pinned commit.

A verified maturity index for public security tools, Phantom, EMET, grouped toolkit work, and authorized private security practice by Zain Dana Harper.

A registry-backed maturity map for shipped, active, research, controlled-private, and archived security systems by Zain Dana Harper.

shipped Phantom helps you inspect and, when authorized, change the hardware identifiers a computer exposes. It works on owned or expressly authorized Windows and Linux systems, saves a backup before changes, and can restore the original values.

Eight distinct private systems for authorized campaigns, assessment labs, AI red-team testing, trust verification, execution, and release gates.

active behavior-transform.io is a local Python wrapper and hook layer for file reads and writes, subprocesses, HTTP fetches, user input and MCP traffic. Given a request, a mode and local text rules, it passes the content through or transforms it, and returns the result with hashes and substitution counts.

controlled-private Array runs authorized offensive-security campaigns in approved waves, with time-limited approval, supervised tools, cleanup and an evidence log.

controlled-private Seed is a C++23 engine for authorized security assessment and detection-engineering work, with synthetic demonstrations, scope manifests, interop adapters and action logs.

controlled-private Sofer is a private Python orchestration runtime and operational security suite for authorized engagements. It manages campaign state, correlation, reconnaissance, reporting, disclosure staging, and evidence while its WARDEN CLI routes model and tool work through a bounded agent loop.

controlled-private Isomorph evaluates classifier and refusal behavior at authorized AI inference boundaries through controlled reformulation trials, provider-specific result records, and control-stability analysis.

controlled-private Bounds checks agent actions, runtime observations and release candidates for intent drift, unsupported claims, secret exposure and failed fixtures, and blocks the release path when a check fails.

controlled-private Kun keeps local access-recovery notes for owned systems: where recovery authority lives, rotation notes, redacted diagnostics and runbooks, without storing raw credentials.

shipped Public Surface Sweeper checks a repository or local portfolio before you publish it: missing public files, unclear README handoff, credential-like strings, release metadata, proof-packet readiness and drift across a workspace.

shipped Model Provenance Validator checks JSON records that link a model or release claim to its sources and retrieval dates, redacts credential-like values from its output, and can emit a Proof Surface packet.

shipped Secret Redact IO is a Python library that redacts secrets from file, fetch, write and subprocess output.

shipped Agent Hook Pack installs hooks for secret checks, branch guards, environment synchronization and repository hygiene.

shipped Repo Proof Index scans the proof packets, receipts and contracts in a repository and builds a reviewer index with the type, status, summary and path of each file.

shipped EMET checks whether the bytes that reach a model, reviewer or pipeline still match their source. It anchors content, compares it later, neutralizes instructions embedded in it, and writes a verdict file that its Python, Rust, Node and Go implementations all read the same way.

active Proof Surface is a Python library and telos-proof CLI that validates records of AI workflow steps, authorizations, delegations and finished work, and returns allow, deny or needs-human recommendations. It never runs the action itself.

active Accountable Surface lets an AI agent take only the file, command, web or browser actions a person has approved. It checks each request, blocks or pauses when needed, checks the outcome, rolls back reversible failures, and keeps a hash-chained journal of decisions and outcomes.

An AI cannot see what happened; it works from what it read and expected. This records the difference between a guess about a result and the result.

A machine should answer for what it does. Senses nobody checked, and rules tacked on at the end, do not add up to an accountable system.

The standard these tools hold a machine to, turned on the person using them: show the evidence, stay in scope, claim nothing you cannot back.

shipped BuildLang is a systems programming language in which a program declares what it is allowed to touch. The compiler checks those declarations and its memory rules, then emits native code through C. It ships with a CLI, editor support, two-way C integration and experimental shader output.

active Build Color measures, converts, compares and transforms digital color: perceptual spaces, HDR tone maps, appearance models, chromatic adaptation, spectral tools, gamut mapping, ICC profiles and LUTs, with an optional GUI.

The Build suite: source-visible forecasting, paper-trading, interface, and color tools with explicit maturity boundaries. By Zain Dana Harper.

The ring of small single-purpose tools around the bigger parts, the ones that run every time something ships, so that shipping is a discipline.

Before calling a project finished, look at it and keep the receipt. A last-mile check that records what was seen rather than what was assumed.

Checkpoint presents the purpose and access boundary for controlled private practice without publishing operational detail.

Nothing carries its own warrant. The argument in seven steps, from conferred existence to a verdict that reaches only as far as it is witnessed.

Find Zain Dana Harper's research projects, essays, publications, source maps, collaboration routes, briefings, and supporting evidence from one Research hub.

active Chorus 0.3.1 produces deterministic discourse digests from captured comment corpora using literal lexicon sentiment, hashed-TF-IDF clustering, engagement weighting, and content-addressed receipts. Its test suite passed in the public CI run for the 0.3.1 release on 2026-09-22. The optional model overlay is separately provenanced and excluded from the deterministic receipt. Inspect the source.

working paper PROBE_MATCH A stricter publication format for using the tools themselves to advance research: source leads, bounded probes, evidence packets, verifier checks, Learn-backed skill loops, and explicit UNVERIFIABLE boundaries. The first probe records a local failing reproduction for a pandas issue, while still blocking patch and PR-readiness claims.

working paper VECTOR_FINITE_OPERATOR_SBP_MATCH A focused Navier-Stokes proof-packet preflight: a bounded periodic skew-symmetry witness, BuildLang parity output, Lean algebraic, finite cyclic-sum, finite summation-by-parts, typed finite-grid, finite edge/operator, and vector finite-operator replay rungs, arXiv source-lead demotion, and an explicit UNVERIFIABLE boundary around the parent Millennium problem.

working paper BIOLOGY_NETWORK_INTELLIGENCE_MATCH A bounded biology/network-intelligence packet: verified source intake for fungal, mycorrhizal, and plant signaling literature, source-lead demotion for blocked DOI routes, Crucible and Learn receipts, and a hyphal context protocol hypothesis for sparse signal routing that still requires benchmark evidence.

working paper HYPHAL_CONTEXT_FIXTURE_MATCH A deterministic dogfood fixture over the biology source corpus: full source-body routing versus gradient envelopes plus receipt retrieval. The hyphal route keeps the same required evidence classes and guardrails with fewer estimated prompt tokens, while model answer quality and general route superiority remain unproven.

source intake SOURCE_RECEIPTS_MATCH A receipt-only source pass over five videos and one Relentless channel queue, mapped into industrial science replay packets, causal research workbench, agentic benchmark foundry, and compute infrastructure ledger tracks. Field tracks are inferred; domain correctness remains UNVERIFIABLE until primary-source or replay evidence exists.

preflight CAUSAL_DAG_FIXTURE_MATCH A replayable toy-DAG packet for causal claims: source receipts, graph assumptions, exact adjustment-set checks, negative controls, Crucible receipts, and a Learn prooflesson. It does not claim causal discovery, LLM causal-reasoning validation, medical recommendation, or BuildLang/buildc-native execution yet.

preflight EMBODIED_SIM2REAL_FIXTURE_MATCH A replayable differential-drive robotics packet: source receipts, unit-checked command logs, predicted and observed traces, tolerance checks, safety envelope, latency boundary, negative controls, Crucible receipts, and a Learn prooflesson. It does not claim real robot safety, medical deployment, foundation-model validation, or BuildLang/buildc-native execution yet.

preflight QEC_STABILIZER_FIXTURE_MATCH A replayable 3-qubit bit-flip stabilizer packet: source receipts, logical states, stabilizers, syndrome table, correction map, negative controls, Crucible receipts, and a Learn prooflesson. It does not claim surface-code decoding, hardware QEC, fault-tolerant computation, quantum advantage, or BuildLang/buildc-native execution yet.

witnessed result A claim from a research talk, turned into a self-checking simulation. The math leg returns a witnessed MATCH at about 0.99 percent mean error against a 5 percent bar. The physical-chip leg stays UNVERIFIABLE, and the page says so. Every number is re-checkable from a seed-fixed script.

long-form thesis The root argument: that existence, status, the moral ought , and legitimate authority are conferred and re-spoken rather than self-standing, and what that means for the standing of a made mind. Built and broken against itself, honest about where it fails.

working paper What actually crosses between two minds: not bits, but faithfulness to a named criterion. The associated faithful-transpile prototype makes that narrow mechanism executable with stub minds and a pluggable judge. Its eight tests pass, while its independent file manifest remains stale; the page therefore treats the paper as a working argument and the code as a bounded prototype, not a proven general law.

v0 scaffold An assembly line that turns a witnessed research talk into a discovery card carrying its own re-check, then asks the verifier to decide. A v0 scaffold that produces candidate cards with attached falsifiers, not discoveries. Most cards are not yet decidable, and the forge says so.

v0 grounded Frontier AI talks and papers turned into claim cards, each carrying its own hashed evidence from a sealed corpus. Six cards are fully grounded, and five of ten modules are backed by evidence.

wrap-up thesis How far a verdict reaches: it binds only where its criterion is witnessed outside the system and can be re-derived. Past that line it is an unwitnessed bid, not a fact. A reviewable draft that ties the recent studies back to the reconcile thesis.

The operational rows used the same seven cases and scoring fields, but different models and stack configurations. This is stack-level evidence, not a same-model harness attribution test.

Two 14B coding models took the same 164 code-completion tasks through the same harness, with one greedy try at each task.

5 harness roles ran the same 7 tasks from the same task set. On every one of those tasks the prompt bytes and the runtime-context bytes each role received were byte-identical, so what differs between the rows is the harness and the model behind it. 11 of 35 attempts produced something a checker could read, and why the rest did not is reported beside each row rather than left inside the rate.

Six test suites ran offline, with no model connected. The first five test accountability, recovery and state handling on their own small test sets. The sixth re-reads saved per-task results from two models: a base model and Flywheel Local Coder 14B, the same model after continued pretraining, meaning further training on new text.

This supporting inventory records commit-backed source structure. It is not the portfolio's benchmark result and is not used as a proxy for quality.

philosophy paper The free will debate has been run as a search for self-origination, a source behind the source. Drop that demand and sourcehood is conferred, not secreted. The load-bearing move is the arity gap: holding someone responsible is a two-place relation between an addresser and an addressee, and a one-place vocabulary of behavior-shaping cannot fix its target. Determinism is granted in full.

philosophy paper A disjunctive-transcendental argument that wherever conferral happens, sorting, naming, meaning, it occupies a seam in space and time. The sceptic who says it dissolves to nothing is still sorting while she argues. The thin index is forced at the conventional level, the thick index is argued but not derived, and the ultimate zero is conceded to Madhyamaka without qualification. The verdict is printed as a header: forces conventionally only.

philosophy paper Five traditions claim consciousness grounds its own existence. Tested against one distinction, self-givenness (known from itself) versus self-grounding (existing from itself), and one image, a flame's heat is built in yet the flame needs fuel, four of them deliver self-givenness only and shrink to seity. The fifth, Advaita Vedanta, is a real undefeated standoff, set aside rather than refuted.

Research corpus · 13 August 2026 Updated 3 October 2026 An integrated thesis on made minds: what exists on its own footing, what is conferred, and where an authentication verdict stops.

Research corpus · 20 June 2026 Updated 3 October 2026 One gap, witnessed at four altitudes: research, philosophy, algebra, and tool. The capstone of a five-part adversarial-steelman corpus.

Every essay, investigation, briefing and research paper by Zain Dana Harper in five sections, newest first.

Essay · 20 June 2026 Five domains expose the same gap between a narrow technical check and the larger property people trust it to warrant.

Essay · 20 June 2026 Updated 3 October 2026 A premise about conferred existence, artificial minds, authority, standing, and the difference between fact and permission.

Essay · 28 August 2026 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.

Essay · 28 July 2026 Updated 3 October 2026 Review debt, verifier independence, honest nulls, labor, and machines whose claims anyone can check.

Essay · 22 July 2026 Updated 3 October 2026 Shame, creative labor, mutualistic tools, source-bounded generative art, and becoming answerable. Also: Pick the Lock for Everyone: spoken edition .

Visual sequence · 23 July 2026 Seventeen images shown in the order they were made: a chronological visual sequence selected by Zain Dana Harper.

Essay · 8 July 2026 Claims, evidence, checks, action state, and the unsupported remainder a handoff should keep visible.

Essay · 22 June 2026 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.

Essay · 29 July 2026 Updated 3 October 2026 Why generated work stays outside the evidence layer, and what a live model-free judgment boundary looks like.

daily source record baseline A dated evidence index for UK AISI, Anthropic, and frontier AI industry safety developments. It separates source roles and dates, records what changed, and states what each public record does not prove. Each edition ships with stable links, a machine-readable record, and an append-only corrections path.

A dated, source-grounded record of AISI, Anthropic, and frontier AI industry safety developments, with explicit evidence limits.

Atlas essay · 1 October 2026 Updated 3 October 2026 An official jobs estimate changes as later data arrive. Each version answers a different question, and a fair comparison names which one it uses.

Atlas essay · 28 September 2026 A readability score measures selected text features. Comprehension, audience fit and accessibility require evidence of their own.

Atlas essay · 28 September 2026 Precise-looking timestamps need clock-error evidence and ordering relations before they can support a close sequence of events.

Atlas essay · 28 September 2026 Music-evoked stories can share patterns while retaining personal meaning. Those reports do not by themselves demonstrate creative accomplishment.

Atlas essay · 28 September 2026 A helper’s report, a recipient’s account and a measured outcome answer different questions about whether support worked.

Atlas essay · 2 September 2026 Updated 25 September 2026 A label can change where people look and how easily they make sense of an artwork, without raising liking, learning, accuracy or artistic value.

Atlas essay · 1 September 2026 Updated 25 September 2026 Repeated listening changes liking, prediction, attention and memory in different ways. A replay count records exposure and says nothing about quality.

Essay · 10 September 2026 Updated 25 September 2026 Separate security reports expose a shared question: what can an agent reach, and can the instruments that record its actions remain trustworthy?

Essay · 9 September 2026 Updated 25 September 2026 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.

Essay · 10 September 2026 Updated 25 September 2026 On existence and what our work owes other people.

Atlas essay · 5 September 2026 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.

Atlas essay · 1 September 2026 Tutoring has strong trial evidence. Public claims about a program must keep its availability, reach, dosage, staffing and outcomes as separate facts.

Dossier · 26 August 2026 Updated 27 August 2026 One OpenAI and Hugging Face incident, compared across five kinds of source: legal process, the company report, host logs, independent analysis and vendor fixes.

Essay · 9 September 2026 An expanded essay on the public tool portfolio, contribution from outside academia, evidence for review and changing educational practice.

Open letter · 19 September 2026 Updated 20 September 2026 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 spoken edition: an account of harm done, and a case for open records on power and public claims in AI, art and knowledge, with private lives protected.

The incident dossier and every dated edition of the Frontier Safety briefing, each with its sources, claim status and limits.

Investigation · 25 September 2026 Updated 3 October 2026 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.

Series piece · 8 October 2026 For people and institutions, a check that is certain and outside the actor's control does more than an announced watcher. Whether that holds for trained AI models is a test, designed here, that has not run.

Series piece · 8 October 2026 Updated 15 October 2026 Frankenstein, Pinocchio and The Fifth Element, read past their famous morals, put the danger in the maker's arrangements: who funds the work, who sets the measure of success, and who keeps and edits the record.

Series piece · 8 October 2026 Updated 15 October 2026 From Hispaniola in 1514 to the Iran-Contra inquiry, the party that controlled the money wrote the first account of it. The record moved when an inquiry, a court, an archive or a reporter outside that party obtained it.

Series piece · 1 October 2026 Updated 3 October 2026 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.

Series piece · 1 October 2026 Updated 3 October 2026 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.

Who Knew First and the five pieces that each test one question it raises, with the published ones linked and the planned ones named.

Series opener · 2 October 2026 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.

A live media instrument for rendering, measuring, transforming, and reading the frame, running entirely in your own browser.

Name a seed, pick from ninety-five instruments, and this site's own engine draws your plate in the browser. Forty-two full-size plates sit below it.

active Retro Engine is an embedded browser studio for images, drawing, GLSL, and audio traces. It applies pixelation, hardware palettes, ordered dithering, early-3D shading, CRT processing, and stackable effects, then exports images, patches, MIDI, relief or disc data, and Loom handoffs.

shipped Engine Revival is a Python tool for building, checking and publishing metadata about revived historical engines. It can also generate an out-of-tree BRender CMake build harness from a public checkout.

shipped BRender Archival rebuilds the Argonaut BRender engine from its public MIT source, runs it through a 21-stage native test ladder, captures renders, and packages the build harness, test transcripts and checksums without copying the upstream source.

active Active Skyrim SE/AE ENB shader-suite work; branch state and live-host acceptance remain separate from public release state.

active Skyrim SE/AE ENBSeries 0.504 shader suite with procedural sky, clouds, aurora, exposure, tone mapping, and an optional camera bridge.

active ENB Runtime Core is a C++23 embedded runtime library that identifies an already-loaded ENBSeries host, validates its SDK surface, queues callbacks outside callback context, coordinates save quiescence and reapplication, and gates a fail-closed Skyrim engine bridge.

active SkyrimBridge is an SKSE plugin that exposes Skyrim live engine state, record editing surfaces, asset conversion, a versioned ABI, diagnostics, and shared-memory command channels to shaders and external tools, with an optional D3D11 rendering tier.

active RAW is a dependency-free C++23 renderer that draws a scene on the CPU, checks its fast screen-space ambient occlusion against a ray-traced reference, and writes a JSON certificate saying whether the shortcut held.

Weave any picture into working cloth. Send a frame from the Retro Engine, the Studio, or the Gallery, and read or write a real WIF draft.

A bounded lab for turning selected Current Story artworks into real, disclosed Gaussian-splat scene experiments without replacing the canonical source images.

Every artwork this session produced that is still recoverable, published whole rather than curated down to the good ones.

Explore Zain Editorial and Mono: two original typefaces in development. Try your own text in limited browser previews.

A reading specimen of the typography currently used on this site, with a look at the direction of our original type work.

Mint and preview Zain Mint, a parametric typeface in progress, in your browser. Set the pen and proportions, type your own text, read the receipt.

Current resumes for systems engineering, developer operations, and field operations, plus a full CV.

Technical support engineering, developer operations, implementation support, release support, and software QA.

Evaluation tooling, Python developer tools, test infrastructure, and research-engineering support.

Physical work, public-facing service, grounds experience, and safety-minded operations.

Public-safe dossier with route, evidence, and claim boundaries.

Current resumes for systems engineering, developer operations, and field operations, plus a full CV.

Full Curriculum Vitae resume and source-matched downloads. Updated 2026-10-23.

Merged contributions to other projects, then my own tools.

General public cover letter for Zain Dana Harper.

The personal story behind the work: how I came to it, and who it is for.

A public request for small AI-assisted workflow examples where the claim, evidence, checks, and unverifiable remainder matter.

AI evaluation review, harness integration, agent safety review and incident investigation, each with its scope and published examples.

Who pays for my evaluation work, which interests could bend it, which AI models help produce it, when I decline work and how findings get corrected.

Every payment, grant and API credit from my AI evaluation and investigation work, with its exact source, amount and share of trailing income.