Research.
The questions behind the work. Every piece is dated to the day the work happened and built from the notes, commits and results in the repositories. Each topic reads in the order it unfolded: the question, what the evidence said, what changed direction, and what is still open.
- A quotation is a note of mine, with its date.
- Where a number appears, it is a number from those files.
- Where something is not confirmed, the piece says so.
Authority SystemSuite.
Rulebooks that make an AI show its work, system by system, and how they were tested.
Sep 20, 2026 → Oct 18, 2026 · 17 pieces, in the order the work happened
The case study →- 01Why every record write goes through one locking scriptOn 20 September two AI sessions wrote the same record id, and a one-line edit mangled the file every session reads first. The fix was a script that takes a lock, allocates ids itself and refuses malformed writes.

- 02Most of the suite’s 552 rules had nothing enforcing themThe first count of the Authority System Suite's own rules, on 20 September, found 552 normative statements, 341 of them (62%) with nothing named beside them that could enforce them. That count shaped how the suite was built from then on.

- 03How the suite controls fabrication and sycophancyThe Authority System Suite treats an AI's reports on its own work as claims to be checked, and agreement with the user as the default failure. Small probes on 29 September found the controls holding on some cases and failing on others.

- 04How the research rulebook handled a hidden instructionRAS, the research rulebook, decides whether a source can be relied on. When a test hid a command in a retrieved web page, the rules handled it correctly but still failed three draws over a missing audit line, which one added sentence fixed.

- 05A live view of the suite’s rules as they fireThe Rigor Manager is an optional local app that draws the Authority System Suite's rules firing, from the project's own event log. It can add a few kinds of record through the project's write script, and it cannot edit or delete one.

- 06Looking for failures the suite has no category forA system can't guard against a failure it has no category for. The Authority System Suite's file on unknown unknowns lists seven ways of looking for such failures, each with a stated limit, and says none of them guarantees anything.

- 07Can a rule stop a bad reply before it’s sent?The reply gate checks an AI's replies in the Authority System Suite after they are written, so it can force a correction but can't stop a sentence being sent. Its rules were measured against real replies before adoption, and the same limit led to checkpoints.

- 08What was the logic test’s byte limit based on?LAS, the logic rulebook, checks whether an argument holds together. One of its tests was widened until it no longer told answers with the rules from answers without, and a byte limit on its prompt was replaced by a check for duplicated rules.

- 09Don’t these thresholds have real definitions?Eight terms in the rules of the suite’s three new systems, such as “last resort”, read as thresholds with no definition. Six had published definitions, which the rules now cite. The piece covers the Evidence Gathering System they sit in, which plans how to obtain records and logs every item it could not get.

- 10Where does the numbers rulebook belong?A short router file decides which of the suite’s seven rulebooks an answer needs, from what the AI is about to produce. While I was building the suite, three of its rows never fired, and on 10 October I reordered its table into the order work happens.

- 11Testing an ethics rulebook against a model with no rulesEHS, the Ethics and Harm System, lays out the considerations and leaves every decision to a named person. In its first paired test it passed 0 of 5 cases without its rules and 4 of 5 with them, and the one miss came from a rule kept in a module that task never runs.

- 12Replacing a test gate that a rule-free model passedTAS keeps an AI from turning an allegation into a fact. A model with no rules passed its first test gate 9 of 9, so I replaced it with harder cases, which on 12 October scored 9 of 15 without the rules and 13 of 15 with them.

- 13The first paired test of the Authority System SuiteOn 12 October all 70 regression-gate cases were run once with the Suite's rules loaded and once without, scored blind and pre-registered. With the rules 62 passed; without them, 40.

- 14Testing whether an AI judges a decision by its outcomeDAS, the decision rulebook, is meant to stop a good decision being called bad because it failed. Its test case for exactly that, D01, passed with the rules only on the sixth paired attempt, and one of three readers dissented.

- 15When to stop testing a rulebookBy 18 October each new round of testing the Authority System Suite found more failures and cost more. The stopping rules written that day set a pass target first, sample before calling a failure, fix the test before the rule and record every deviation.

- 16What it took to pass the numbers rulebook’s hardest caseQAS makes an AI show how every number was produced. Its hardest test case, Q24, took the evening of 18 October to pass, often because the fault was in the test, and the final pass came from rule edits fitted to the failures.

- 17The full adversarial run and its eleven failuresOn 18 October all seven adversarial suites ran paired: 76 of 151 cases passed without the doctrine, 140 with it. This covers how the eleven remaining failures were fixed and re-tested, and the stopping rule I wrote down.

JournalismWorkbench.
A private investigation desk: the redesign, the screens, the decisions.
Sep 17, 2026 → Oct 19, 2026 · 12 pieces, in the order the work happened
The case study →- 01Security and test data in the Workbench’s first two daysThe first work on the Journalism Workbench was not features: documentation, dependency and security housekeeping, and removing real people's names from the test data.

- 02Redesigning the Journalism Workbench from mockupsOn 10 October I started a full redesign of the Workbench, which had no user flow and a dashboard spread down one long page. It set dark colours, one screen without scrolling, whole-screen features and a Rain Noir theme, with every screen approved as a mockup first.

- 03Teaching the Workbench’s red flags with real casesThe Workbench's red-flag rules are tested on demo cases built from public records, but the ownership-loop rule never fires on the main 1MDB case. A Bermuda court judgment that shows a real loop is kept for the tutorial instead of becoming a case of its own.

- 04Designing five themes where status never relies on colourThe Workbench has five complete themes, from a navy newsroom to a rainy city window. In every one, the Supported and Disputed colours measured too close for a red-green colour-blind reader, so every status is also a shape and a word.

- 05How do you test everything without having everything?The Workbench handles documents, photographs, video, foreign-language judgments and money in six currencies, and a test case with three PDFs tests none of it. My answer was a demo case built from public records, holding 104 documents, with three small real cases for what it can’t cover.

- 06Showing each file’s progress through intakeThe Workbench's old intake showed progress for the whole case and then forgot it. The redesign gives every file its own lane through seven stages and puts the files with problems first.

- 07Designing a network view for cases too big to drawOn the 1MDB demo case the Workbench's network view fitted 81 parties on one screen and could not be read. Investigators' tools start from one party and expand, so the replacement, the Wire, shows one map at three altitudes.

- 08Porting the portfolio’s rain engine into the WorkbenchThe Workbench's Rain Noir theme needed bokeh, rain on glass, grain and neon, all of which my portfolio's engine already had. On 12 October I brought the whole effect system in, about 21,000 lines, behind one set of imports.

- 09Checking the media suite plan against its primary sourcesThe media suite is a plan for analysing photos, video and audio in the Workbench. On 18 October its mockup was checked against model cards, licences and papers, and several claims did not hold; nothing of it is built and twelve questions are open.

- 10A step-by-step record of every AI run in the WorkbenchWhen the Workbench's AI runs, every step is written to an append-only record, and the app works out each step's status from its own checks. What the model says about its own work stays marked as self-reported until a check confirms it.

- 11Five architecture decisions behind the Workbench redesignLocal fonts, a story whose strength is worked out rather than stored, an editor that stores one mark, a glow on its own canvas and a stream that falls back to polling. Each is a written decision record with the options it turned down.

- 12Tracing a claim to its files without altering the caseThe Workbench's Chain of Custody screen shows how each piece of evidence reached a claim and marks what the app does not store. A change on 19 October made every screen that reads a case leave it exactly as it was.

Cryptographyresearch.
New mathematics, possible defences, and what the evidence does and doesn't show.
Oct 13, 2026 → Oct 22, 2026 · 10 pieces, in the order the work happened
- 01Can new mathematics help defend against AI?A batch of open mathematics problems reported solved by AI, and reports of AI agents getting out of their boxes, made me ask whether the new mathematics could help defend against AI. I couldn't find anyone who had written about it, so I started looking.

- 02Who measured the AI cyber agents’ results?Of the reported results for AI cyber agents in the 14 October review, only DARPA's contest had an outside referee. The vendor figures are vendors reporting on themselves, and no independent validation of them was found.

- 03How much post-quantum key agreement adds to a TLS handshakeHybrid post-quantum key agreement makes a TLS client's key share 1,216 bytes where a classical one is 32. A review of primary sources from 14 October also sets out what is standardised, what the NSA's 1 October note requires and which quantum results are measurements.

- 04Can a file’s encryption change by itself?Could the new mathematics make a file whose encryption changes over time with nothing running? Under a strict definition the requirements appear incompatible, and every way round that needs an assumption outside the mathematics.

- 05How I audit 719 mathematics manuscriptsThe source repository holds 719 manuscripts in 372 families. Each run reads one from a pinned version, runs small exact checks, searches for prior art and files it under one of five fixed classes, with the general theorem marked unverified.

- 06Two almost-linear algorithms that never use their fast pathTwo audits on 15 October found almost-linear algorithms, for matching and for edit distance, whose own parameters keep the fast branch off for every input that can exist. In practice both run the classical method; neither audit disputes the mathematics.

- 07Auditing a manuscript on the number of distinct totientsOne audit, start to verdict, of a manuscript on how many distinct values Euler's phi function takes. Every finite check passed, but the general theorem is unverified and the printed pages are still unread.

- 08How the manuscript audits are run and trackedTwo reviewers work through the manuscripts from opposite ends, each source pinned to an exact version. An audit of the project's own records on 19 October found the coverage ledger was a narrative rather than a census, and the exact completion percentage unknown.

- 09What two maths manuscripts predict about measurable physicsTwo reports from 22 October read manuscripts on fracture and on interfaces in a cubic torus for physical consequences, and mostly found limits. A wrong fracture model can fit well and be about 27% low, and the torus effect vanishes as the cell grows.

- 10What nine days of cryptography research records showAfter nine days, most entries in the research repository end the same way: related precedent, finite checks passed, theorem unverified, and no verified new cryptographic primitive.

Light &shadow.
The lamp, the glass, one shadow model, prints, black light, breaking glass.
Sep 23, 2026 → Oct 23, 2026 · 19 pieces, in the order the work happened
The case study →- 01How the site’s cursor became a camera shutterThe cursor, an iris that closes on a click, a shutter you wind by holding, recorded sounds, a flash that fills the page and an image burned into the background. Each part came right when a measurement replaced a guess.

- 02Building the first pane of glass for a photography siteOn 25 September I set out to make the site feel like a room you look into rather than a page. The first piece was glass that shows what is behind it, calibrated against a real glass interface, and a lamp that comes on when you hold still.

- 03Making smudges and scratches block light by opacityThe site’s panes carry smudges and scratches from two photographs, so they look handled. Each pixel of a mark now blocks as much light as its own opacity says, and a scratch only takes light away.

- 04Rebuilding glass edges from photographs of float glassI rebuilt the edges of the site’s glass panes as reflections, refractions and highlights, from reference photographs. Measuring five of those photographs showed my first edge colours were far too pale and too yellow.

- 05Rebuilding the light engine so the picture stayed identicalOver two days I rebuilt the site’s light engine in about fourteen steps, each checked pixel by pixel to leave the picture unchanged. It came out as one ordered model of causes that does no work on a still page.

- 06Rebuilding every shadow on the site from one modelOn 28 September the site's separate effects were rebuilt so that every light value comes from the lights and every shadow from one model. It is why the rain, a week later, arrived already lit.

- 07Which sides of a glass pane can you actually see?A pane of glass shows a side face only when your eye is on that side of its edge, and only as wide as perspective allows. Building the sides, and the light through the glass, from geometry instead of fixed settings made both behave as real glass does.

- 08What does a laser do when the page is made of glass?The site’s laser runs across the page in the plane of the glass, so it meets each pane at its edges and bends, reflects, is trapped and fades as light does at a real glass surface. The physics it follows, and the settings for its colour, power, width, reach and dust.

- 09Modelling a torch beam with a hotspot and spillThe flashlight tool added on 30 September is the first light on the site with a direction. Its beam is a hotspot and a dim spill set by three constants, taken from flashlight forums and a beam model rather than measured from a real torch.

- 10Cataloguing what light does and what the engine modelsOn 30 September I catalogued 44 things light does and where the site’s light engine stands on each. Twenty-four are modelled, fifteen are in preview, and the three ruled out each give a reason that can be measured.

- 11Modelling broken glass on real fracture patternsThe first broken glass was smooth rings and a few straight lines. Rebuilding it from fracture research, 55 photographed pieces of broken glass and a hammer gave staggered rings, forking radials and bright crack faces.

- 12Why were my shadows so pale?An audit of the site’s shadow model against the formulas found the geometry right and five other faults behind the pale shadows. One of the pale shadows was correct physics for a large lamp.

- 13Research notes on balloons, fireworks, flame and gemsFour research notes written on 1 October set the physics for balloons, fireworks, a burning torch and gems, each with sources, decisions and a list of what is still unknown. The fireworks had a first cut by the end of the day.

- 14What does a room actually look like under black light?My first black light was built from a list of materials that fluoresce, and left the page lit by the room. I redid it from what a black-lit scene looks like: room lights off, everything dark violet except the whites.

- 15Turning a forensic glass study into an acceptance testA forensic study broke 60 identical panes, and no overall pattern repeated. I turned its statistics into the test a simulated break must pass before it goes into the site’s library.

- 16Where my fracture simulator fails on pressed glassMy fracture simulator breaks struck glass in the order real glass does, but its glass fails at about a fifth of real glass’s strain. A pane pressed slowly therefore stores a 27th of the real energy and breaks into a sparse web.

- 17Halloween effects lit by the scene’s own lightsA burning torch whose flame is also a light, and a spider web whose threads glint only where a light catches them, both exist as first cuts. A blood version of the rain is planned and not built.

- 18Modelling the gloss on a photographic printOn 11 October the gallery frames became lit prints. The gloss was rebuilt from 50 reference photographs of prints under lamps, which showed that a print's highlight is a tilted, tapering bar rather than a round spot.

- 19A test room of twelve glass panesA development page holds twelve panes, four rows of three, on a dark backdrop with no photograph. Each pane is built differently, so an effect can be judged on the glass alone and matched to a setting.

Water, rain &wind.
Drops as lenses, streams, drips, wind, and the footage they were checked against.
Oct 2, 2026 → Oct 11, 2026 · 11 pieces, in the order the work happened
The case study →- 01Modelling spray and fog on a window paneOne squeeze of a trigger sprayer puts about 300,000 droplets on a pane and makes no stream. The research on spray and on steamed glass set what the site’s glass does with very small water; the fog that grows and re-fogs is still only a design.

- 02Why do the drops look like ink?The first rain had drops, merging and running, but the drops looked like ink. Putting real rain photographs beside it gave seven physical causes, and the first three carry most of the difference.

- 03Is the light under a raindrop accurate?Each drop on the glass is a lens that throws a pattern of light onto the photograph behind, which I ray-traced offline into a lookup table. A path-traced reference showed drops by a lamp are darker inside with a thin bright ring, and that a slope estimate had been hiding the ring.

- 04Turning 117 photographs of rain into generator rules117 reference photographs of rain on glass, analysed drop by drop, became 44 rules for the site’s rain generator, each with the photographs that support it. On how patchy rain should be, the counts and my own critique still disagree.

- 05Tracing rain streams from 117 photographsTo make the site's rain streams look real I measured them: 233 streams traced by hand from 117 photographs, and 21,834 drops tracked in six clips of real rain.

- 06Measuring simulated rain against real footageSix clips of real rain gave measured targets, and the site’s rain is recorded one fixed step per saved frame so its speed is true. On 9 October only one of eight measures sat inside the footage’s range.

- 07Why most drops on a window never moveBetween 76% and 92% of the drops on a real window never move, held by the difference between their front and back contact angles. The ones that do move go in short bursts, and big streams flow more steadily than small drops.

- 08Letting rain streams absorb the drops they touchThe site's rain streams were drawn on top of the drops instead of absorbing them. A probe found 123 of 1,013 drops inside drawn streams while the simulation counted 6; making each stream a lane the simulation owns brought that to 0.

- 09How realistic is the wind on the glass?The wind on the site’s rain swells, gusts and changes direction, and pushes small drops harder for their weight than big ones. Against a frame, water now gathers along the bottom edge and drips.

- 10A page where you steer one drop across glassA page on the site where you push one drop of water across a pane with the arrow keys or a controller. It runs on the same water simulation as the rain, drawn at nearly five times the site's scale.

- 11Open problems in the rain simulationAfter the first fortnight the rain’s streams were still chains of beads, a fluid solver had been set aside, and several critiques of density, shape and brightness stood open. This is the list of what was not fixed, with dates.

Thestudio.
Photography plus systems: how the two halves fit.
Oct 1, 2026 → Oct 19, 2026 · 3 pieces, in the order the work happened
- 01Turning the lighting lab into a site-wide editorBetween 29 September and 1 October the lab, where I tune the site's light, went from saving to my own browser to an editor with a live frame, an explicit Save for everyone, and tiers for weaker devices. What I set there is now what every visitor sees.

- 02Why this site has case studies as well as photographsOn 2 October I decided the site would be both my photography portfolio and a direct showcase of the technical systems behind it, in case studies, a one-pager and this journal.

- 03Replacing the photography home page with a welcome pageOn 19 October the site's first page became a welcome page that routes to photography, three case studies and a one-page summary. It follows research into how portfolios are read, which found a recruiter spends about 10 seconds on a homepage.

Get in touch.
I'm open to creative technical project management, founding teams, and agentic AI engineering and research. I read everything that comes in, and I answer.
onysnowstudios@gmail.comLinkedIn →
