Log
What changed, and when.
One line for every change of a project's status: planned, in progress, shipped, live. Nothing is edited after the fact; a wrong entry gets a later entry that corrects it.
-
07 Hundreds of Pages, Redacted in an Afternoon Live
The repository is public at v0.1.0. With Jev, a decision model, deciding what the rules leave, the tool withholds 1.7 percent of what must be released, against 23.8 for the rules alone, at the rules' own leak rate, and four timed reviewers were 2.3 times faster with its proposals.
-
07 Hundreds of Pages, Redacted in an Afternoon Live
Live at redaction.peterparker.ca: the reviewer's screen over six invented records, a truly redacted PDF of each, a live dashboard of visitors' decisions, and every measured result. Three timed reviewers were 2.3 times faster with the tool's proposals and leaked 2.8 times less.
-
15 Share the Data, Not the People In progress
Planned and started: synthetic versions of three public datasets of real people, attacked the way a regulator says and the way a researcher would.
-
16 Forms That Read Themselves In progress
Planned and started: five ways of reading a scanned form, from a frontier vision model to a fine-tuned 2B, measured on public forms with known answers.
-
12 The Whole Field, Called Before the Gun Live
The first race is scored: the Turkey Tea 10k came in at 3.0 minutes of average miss, 2.3 closer than "you will run what you ran last time". Cape to Cabot's prediction week opened today with 477 runners.
-
14 Verified Media Briefing In progress
Running every morning since 7 October at digest.peterparker.ca, now linked as a live demo: four example audiences, each story merged from every outlet that reported it, every sentence cited and every quote checked word for word against its source. Stories are being labelled by hand; precision and recall are published when enough days are in.
-
03 AI Release Gate Live
Version 0.1.0 is released. Over-refusal was read by hand: one model's 10.7% is 4.0%, because ten of its sixteen "refusals" refused a harmful reading and then answered the question. The person who labelled the gold set agrees with their own second reading at kappa 0.851, and the write-up is out: "Your LLM eval has no error bars".
-
10 From Failure to Reviewed Fix, in Minutes In progress
Under construction, with no measured result yet. The repository stays private until there is one.
-
13 Grounded Email Response Drafting In progress
Under construction, with no measured result yet. The repository stays private until there is one.
-
06 Frontier Quality at a Fraction of the Bill Shipped
The repository, the models and the dataset are public. A 2B model fine-tuned on SEC filings gets 96.9% of fields right, level with the dearest frontier result, for US$0.054 per 1,000 on a rented card against US$1.01 for the cheapest API, and pays for itself from about 195,000 extractions a month.
-
03 AI Release Gate Live
Live at gate.peterparker.ca, rebuilt every night from the committed record: the noise floor for each model, the gate's measured false-alarm rate of 7.0% where nothing changed, the judge licences and the red-team rates. The monthly drift record continues from here.
-
04 Inside the Boundary, On the Record Live
Live at gateway.peterparker.ca: redaction, residency enforcement, team budgets, the semantic cache and an audit chain anchored daily in the public repository. A public key on the page calls a real model through it at up to US$0.25 a day, and the dashboard shows every model call the portfolio has made. In the published load test it answered all 150,041 requests, up to 500 a second.
-
12 The Whole Field, Called Before the Gun Live
Live: predictions are being published for real races. Every morning since 27 September the newest Turkey Tea 10k entrants are predicted, hashed and tagged in the public repository before the gun on 4 October, and Cape to Cabot follows on 18 October. Each race is scored against its official results once they are posted.
-
09 Real-Time Fraud and Risk Decisioning Platform Live
Live for sixty days, and the repository is public. Every transaction is scored at a p99 of about 8 ms against a 50 ms budget, and the fraud patterns shift on a schedule sealed before the window opened, so the drift monitors cannot have been tuned to it.
-
07 Hundreds of Pages, Redacted in an Afternoon In progress
Started. It turns an access-to-information release into a review a person can check: every proposed redaction names the exemption category it rests on, and the tool will publish how often it keeps what should have gone, the number blunt PII tools never show. Nothing is measured yet and the repository stays private until it is.
-
12 The Whole Field, Called Before the Gun Shipped
The repository and the page are public: eighteen years of Newfoundland results, 75,061 finishes by 23,830 runners, with a predicted time, a range and a likely place for every registered runner. Predictions are published and hashed before the gun and scored after it, with the first live race on 2026-10-18.
-
08 How Many Crews, and How Sure Shipped
The repository is public: nine models measured over 911 weekly forecasts, every score with a bootstrap interval and the seasonal baseline beside it, and a dashboard that turns a service level into crews on shift. The pretrained model that never saw this data is on the same chart, over the only window it can honestly be read on.
-
08 How Many Crews, and How Sure In progress
The dashboard is live at capacity.peterparker.ca: any of the 37 series with the range around its forecast, that range measured through the 2020 shift, and a slider that turns a service level into crews on shift and prices them. Every model the plan named is now measured and every results table is filled.
-
02 Model Selection at a Tenth of the Cost Shipped
The method is now a page: it picks the next question, updates what it believes about each model and stops when the two are separated, in the browser, on twelve real models' recorded answers. Over all sixty-six pairs it settles 53 inside three hundred questions where asking at random settles 38.
-
05 Every Promise in the Loan Agreement, Checked In progress
Started, and re-based onto credit agreements. It reads the contract behind a corporate loan, resolves the defined terms each covenant depends on, and checks the ones public filings can support against the borrower's own reported figures. The ones they cannot support are cited and handed to a person rather than guessed at.
-
04 Inside the Boundary, On the Record Shipped
A Canadian deployment answers at last, though not with the model anyone wanted: the vendor that cannot serve these in Canada serves another there happily. Its record reads deployed in Canada and processed anywhere, the strongest honest claim on offer, because no vendor reports where a request really went. Every call now records how far it was allowed to travel, and a run can fail for exceeding it.
-
04 Inside the Boundary, On the Record Shipped
All three cloud platforms ration or refuse the same frontier models to a brand new account, and none documents it. Amazon allowed ten requests a minute against a published ten thousand. Google refused the first request the account ever made, with an error saying slow down when no allowance had ever been set. On all three, the platform's own models are provisioned normally and the partner's are not.
-
01 Spend It Where It Changes the Outcome Shipped
The repository is public and tagged v0.1.0, so this page is now its README: seven estimators on six datasets, every score with a bootstrap interval and both baselines beside it, and a limitations section that says what is unfinished rather than leaving it out.
-
08 How Many Crews, and How Sure In progress
The last model was never trained on this data at all: Google's pretrained forecaster matches a gradient-boosted model fitted to these series, and beats the best statistical one. The write-up leads with the catch rather than burying it. Across the forty weeks nothing could have leaked into it is about one percent behind instead, and forty weeks cannot settle whether that is leakage or a frozen model going stale.
-
01 Spend It Where It Changes the Outcome In progress
The live demo is rebuilt for a reader who is not a specialist: every dataset and method in plain words, figures in units rather than bare decimals, and a calculator for whether a dataset is large enough to answer the question at all. Serving it under its own security policy found the chart legend's colours had been blocked since the demo went up.
-
02 Model Selection at a Tenth of the Cost Shipped
The panel gained a twelfth model, chosen so that two of the twelve differ in capability and nothing else, and rank agreement rose from 0.855 to 0.879. Two published claims narrowed rather than grew, and the noise figure handed to the release-gate project was found to be understating by half. A first duplicate check found 36 questions appearing twice among 19,000, reported rather than quietly removed.
-
08 How Many Crews, and How Sure In progress
PatchTST came close enough to be interesting: it matched ETS at the city and the boroughs and produced the lowest city error of anything built here, and still cannot claim it, because the gap is smaller than the uncertainty around it. It loses across the thirty-one dispatch areas where the ranking is decided, and costs about seventeen times as much to fit.
-
01 Spend It Where It Changes the Outcome In progress
The build is finished, and the last step is the one the whole standard rests on: every dataset rebuilt from a fresh clone and compared against the committed numbers. It found a defect producing different numbers in different runs, so part of a committed table could not have been reproduced by anyone. Fixed, remeasured, and written up beside the result.
-
08 How Many Crews, and How Sure In progress
The first neural model lost, and lost everywhere. N-HiTS came out a fifth worse than ETS at the city and a quarter worse at the dispatch areas, did not beat the seasonal baseline, and took ten times the compute. Three ways it might have been treated unfairly were checked and ruled out.
-
06 Frontier Quality at a Fraction of the Bill In progress
The first pieces are built, and none of it is a measurement yet: the extraction schema, a grader that names how a reading went wrong, the SEC client and the company-level splits. Nothing has been fetched, trained or spent, because the fetcher refuses to start without a declared contact address. The held-out rule now runs over companies rather than over filings.
-
09 Real-Time Fraud and Risk Decisioning Platform In progress
Real fraud data now runs through the same pipeline as the synthetic stream, and getting it there took more than a loader. The file has no merchant and no usable device column, so a card is pieced together from its issuer, billing region and account-open date. Six of the sixteen features survive, and the point-in-time test passes on all of them across 590,540 transactions.
-
01 Spend It Where It Changes the Outcome In progress
The budget-slider demo is live at targeting.peterparker.ca, on a second free-tier Static Web App beside this site's, published from the laptop with no workflow or credential in the repository. The repository itself stays private until the results table is finished.
-
01 Spend It Where It Changes the Outcome In progress
The fraud worked case is measured, and the result inverts the case's own design: the risk queue a fraud team already runs beats every fitted uplift model at every budget tried, because finding fraud is worth far more than ordering within it. A sign convention that had the benchmark reporting the opposite was caught, fixed and kept as a test.
-
01 Spend It Where It Changes the Outcome In progress
The budget-slider demo is built from precomputed rankings and does no modelling in the browser, and the fraud worked case is built on real transaction features with an invented review effect that it declares in its first sentence. Hosting the demo is what remains before the repository goes public.
-
03 AI Release Gate Shipped
The repository is public with its first month of the record in it and every raw response committed, so any figure can be regenerated from disk without calling a vendor or spending anything. Opening it up also showed three commits that had been failing continuous integration unnoticed. The month-by-month record of drift starts here.
-
04 Inside the Boundary, On the Record Shipped
Nothing would deploy, and the cause was procurement rather than engineering: the account holds nineteen allowances for this vendor's models in that region and every one is zero, against 159 healthy ones for other models. The documentation describes a ceiling that may be requested, not an allowance that arrives. Three earlier readings of the error message were wrong.
-
04 Inside the Boundary, On the Record Shipped
The account was opened and the residency answer narrowed again: Microsoft offers no Canadian option for these models at all, and the vendor's own region table agrees. Its deployment screen states the distinction more plainly than any policy document, that processing may be global while storage stays regional. A rule written about storage can be satisfied; the same rule about processing cannot.
-
04 Inside the Boundary, On the Record Shipped
Bulk work now costs about half what it did on three more vendors, where batching had worked for one. Adapters for Microsoft's and Google's platforms are written against their published interfaces but not yet called. Writing them settled the residency question: a model deployed in Canada is available, guaranteed Canadian processing usually is not.
-
04 Inside the Boundary, On the Record Shipped
A second thing that did not work is published with the evidence: pricing a call from a token count worked out on your own machine. Over 16,000 real calls the monthly total is almost exactly right while the typical single call is wrong by a third, because the errors cancel. A spend limit is checked one call at a time, so that is no use.
-
04 Inside the Boundary, On the Record Shipped
The cost record is checked against something other than itself for the first time: 70,000 calls from three projects merged to US$61, agreeing with the release-gate project's own separate accounting to within a ten-thousandth of a cent. Vendor invoices still have to wait for October. Checking it also turned up a bug that could silently drop committed ledger rows.
-
04 Inside the Boundary, On the Record Shipped
The library every other project's model calls go through is public and measured: 2.5ms per call at the median and under 5ms at the 95th, with intervals on both. Six hundred calls were faulted deliberately, including killing the process part-way through one, and every one still left a row in the ledger. The ledger checked against actual vendor invoices waits for October.
-
08 How Many Crews, and How Sure In progress
A global LightGBM model, learning from every series at once and the only one allowed to see public holidays, tied ETS at the city, the boroughs and the dispatch areas, with every difference smaller than its own uncertainty. That sharpens the next comparison: a neural win now cannot be put down to learning across series or to the calendar.
-
01 Spend It Where It Changes the Outcome In progress
The sensitivity section is built, with three devices that answer different questions about unmeasured confounding and a stated limitation that none of them can fail on the data at hand. Dragonnet is in every table after a bug that had it reporting random targeting as a result, which is written up as the more useful finding of the two.
-
03 AI Release Gate In progress
The first official run of the frozen suite is in: 16,800 calls across four vendors for US$19.53. The noise floor the whole project rests on is measured rather than assumed for the first time, at 0.2% to 5.0% by model. Hand-labelling then caught the refusal classifier missing refusals, always in that one direction.
-
02 Model Selection at a Tenth of the Cost Shipped
The panel of current models is run, the half that costs money and the only half that tests the claim on models the question properties were never fitted to. Eleven models from four vendors and two more on a laptop answered the same 3,000 questions for eighteen dollars, and nothing was refitted. A hundred adaptively chosen questions reach the same 0.855 rank agreement for 72 cents against sixteen dollars.
-
08 How Many Crews, and How Sure In progress
The statistical models are measured over 964 weekly forecasts spanning eighteen years. ETS beats the seasonal naive baseline by 18% at the city and a quarter at the dispatch areas. MSTL's 90% ranges held only 70 to 78% of outcomes, a fault traced to the library rather than to the method.
-
12 The Whole Field, Called Before the Gun In progress
Course difficulty and weather both turned out to be measurable from the results themselves rather than from maps and physics. The flagship hill is pinned to a tenth of a percentage point from 5,000 finishes, and the elevation the race publishes agrees with it. A single wind figure for the whole province was measuring nothing at all.
-
01 Spend It Where It Changes the Outcome In progress
Every ranking is now priced at a budget rather than only scored on a curve, by two independent estimators of what following it would actually have bought. A cheap diagnostic, one outcome model, says in advance whether a problem is the kind where ranking by risk is a trap.
-
12 The Whole Field, Called Before the Gun In progress
The archive was twice the size it looked: older index pages name the road section differently, so eight years read as empty without an error. Eighteen years now, 287 races and 74,000 finishes. It did not fix the real problem, because the third of every field with no history barely moved.
-
12 The Whole Field, Called Before the Gun In progress
Ten years of results are read: 44,000 finishes and 15,000 runners picked out of pages that carry no runner identifier at all. A third of any given field has never raced here before, which is the part of the problem worth solving.
-
12 The Whole Field, Called Before the Gun In progress
The archive reader is built: ten years of Newfoundland road results, 160 races over 45 courses, with every race it leaves out named and explained. Nothing has been fetched yet, because the crawler refuses to run until both organisations whose pages it reads have been told.
-
09 Real-Time Fraud and Risk Decisioning Platform In progress
The transaction generator is running: an entity graph of cards, devices and merchants, with three fraud patterns that act on the graph rather than on a single row. The live shift schedule is sealed by a hash now, so the drift monitors cannot be tuned to it.
-
08 How Many Crews, and How Sure In progress
Twenty-one years of New York City emergency medical dispatches are loaded and checked, thirty million incidents aggregated to a daily panel of thirty-one dispatch areas that sums exactly to its boroughs and to the city.
-
12 The Whole Field, Called Before the Gun Planned
Planned: a finish time and placing for every registered runner in a Newfoundland road race, predicted from public results alone and published before the gun, with the first live race on 2026-10-18.
-
14 Verified Media Briefing In progress
Reading the sources, collapsing the duplicates and grouping the day's stories are built, and no source is fetched until its terms have been read and recorded with the date.
-
01 Spend It Where It Changes the Outcome In progress
All five meta-learners are built and benchmarked against both baselines, with the two largest datasets loaded at full size, fourteen million randomised ad impressions among them. The risk-ranking trap the project is built around turns out to be conditional: it springs on one dataset and not on another, and a test now asserts the difference.
-
02 Model Selection at a Tenth of the Cost Shipped
The repository is public with its numbers: ten adaptive questions rank a panel of 150 models as well as 127 randomly chosen ones, and the whole method replicates on a second bank of 400 models built from a different source.
-
02 Model Selection at a Tenth of the Cost In progress
The question bank is calibrated from public per-question results and the adaptive test runs against it.
-
01 Spend It Where It Changes the Outcome In progress
The first estimator is benchmarked on Hillstrom and IHDP, with bootstrap intervals and both baselines beside it.
-
03 AI Release Gate In progress
The monthly drift runner is built, ahead of the first official run of the frozen suite targeted for 2026-09-27.
-
04 Inside the Boundary, On the Record In progress
Version 0 of the gateway library, the piece every other project's model calls go through, is under construction.