Applied

[ architecture ] · § fig. 2, inked

[ decisions ] · § as filed

[ validation ] · § the receipts

rows marked [local, verified on request] were run by me on personal or demo data · ci rows link the public run · repo pins are the exact commits verified 2026-09.

[ corrections ] · § the register

• erratum · 2026-07-26

  Until today this file described a native macOS app whose “web beta is a scaffold”, and pinned every receipt at 3225eb4. That stopped being true: Applied ships as a hosted web app at getapplied.vercel.app. The caveat is retired because it is false now, not because it was wrong then — and the receipts are re-pinned to 36a2f54, the public head of the branch that carries the web app. The scaffold claim’s own successor is receipt 07.

• erratum · 2026-07-26

  The backend-suite row read 182 tests, audited at the old pin. Re-run at 36a2f54 on 2026-07-26 it is 271 passed and 10 skipped. Nothing was retracted — the tree grew, and the number moved with it. The 10 skips are named in the row rather than folded into the total.

• note · 2026-07-26

  The repository was renamed yadava5/jobtracker ⟶ yadava5/applied. GitHub still redirects the old links, so nothing this file ever published is broken; the pins name the current repository instead of relying on a redirect. The case-file route stays /projects/jobtracker/ for the same reason.

• erratum · 2026-08-02

  The backend-suite row read 271 passed and 10 skipped. Re-run on the provenance audit at the current head 0f2b63f: 278 passed, 10 skipped. Same movement as the 182 ⟶ 271 correction above and for the same reason — the tree grew seven tests, nothing was retracted, and the skips did not move. The row is re-pinned to the commit the new count was taken at; every other Applied receipt keeps 36a2f54, which is still where those source audits were done.

• note · 2026-08-02

  Provenance audit: the classifier claims were re-derived by running the code rather than reading about it. The rules gate passed live at 96 samples, macro-F1 0.9791, 2 misclassified — and the deterministic hybrid gate returned the identical numbers, which is the direct evidence for the attribution correction recorded earlier: the file named “hybrid” measures the regexes alone. The 201-rule figure was recomputed by importing jobtracker.classifier.rules and summing the pattern lists across all seven categories — 106 strong, 26 weak, 69 negative, plus 14 ATS domains. The eval set was counted from its own JSONL: 96 samples, 8 classes, 12 each. Every one of those figures matched what this file already said.

• erratum · 2026-08-15

  This file described a desktop app that no longer exists. The macOS client was de-scoped on 2026-08-12 and deleted from the repository — apps/macos went, and with it a second, unmounted set of FastAPI routers — so five present-tense claims here were false: the role line, fig. 2’s SwiftUI branch and the edge feeding it, the architecture summary’s “the desktop branch is where the two heavier classifier layers still live”, the rules-layer trade-off, and the boundary row that sent embeddings and SetFit “to the desktop path”. All five are corrected: layers 2 and 3 run in the int8 browser export, and nowhere else. Nothing is retracted and nothing is re-pinned — receipt 06 still cites the macOS build, because it was true at 36a2f54 and that tree still resolves. What changed is the tense: a pinned receipt speaks for its commit, and prose speaks for today.

• note · 2026-08-15

  Applied is a product now, not a study, and this file leads with it: the run’s rail sends a reader to the live app and the System Card before the case file. Two disclosures come with that. It is licensed proprietary, all rights reserved — the repository stays public so the privacy and isolation claims above can be read against the code, not so the code can be reused. And access is an invite-only beta, for the reason the access boundary row gives rather than a reason anyone chose.

• erratum · 2026-08-15

  Four sentences here said Applied fetches Gmail metadata and never message bodies: the access constraint, the architecture summary, fig. 2’s fetch node, and decision d1. They stopped being true on 2026-08-15, and so did the edge feeding the classifier, which said the classifier is handed a subject and a snippet. The decision was real and it was reversed on evidence rather than on preference. Gmail’s own snippet averages 186 characters, an ATS rejection spends them on its polite preamble, and of four real rejections three could not be decided from the snippet at all; one has no snippet whatsoever, and the classifier had never once filed a rejection without a human correcting it. Applied now fetches the body, classifies it in flight, and drops it. What is stored did not change: Gmail’s own snippet, the sender, the subject, the date, and the verdict. Bodies stay out by the shape of the code rather than by a promise, and tests/test_body_is_never_persisted.py drives a real scan with a sentinel in every body, then asserts the sentinel reaches no column of any table, no log record, and no response of any endpoint the scan touches; it also asserts the stored snippet equals Gmail’s, because a sentinel search alone passes for a body prefix that stops short of it. On the day of the change the classifier filed two real rejections by itself, reading its own confidence at 0.90 and 0.95, which it had never once managed before. Those two figures are the model’s own outputs on 2026-08-15 and not a count of anybody’s mail: the boundary row that refuses production email volumes and user numbers stands, and neither message is shown. d1 keeps its argument and is marked superseded, because that section is filed as filed; the date it was taken is added at the head of it, so the row cannot be read as a claim about today, and d5 states what runs instead. Nothing is retracted and nothing is re-pinned: receipt 03 still says the hosted fetch reads metadata only, because that was true at 36a2f54 and that tree still resolves. What changed is the tense, and a pinned receipt speaks for its commit while prose speaks for today.

• note · 2026-08-15

  The Hugging Face Space that hosted the full three-layer classifier went private today, so every link to it returns 401. The references are gone from this file rather than left to fail in a reader’s browser, and the prose no longer offers a hosted demo it cannot deliver. What stands in its place is the receipt the claim always rested on: the browser build itself, ml/browser/site at the pinned commit, where index.html, app.js and the 22.8 MB int8 weights sit together and can be read. Nothing about the export is retracted; the place it was hosted stopped being public, and that is a smaller fact than the one the row asserts.

• erratum · 2026-09-23

  The rule count read 201. Recounted at Applied’s public main head 0285675 by walking the AST of PATTERNS in backend/jobtracker/classifier/rules.py, the same way Applied’s own scripts/readme_facts.py counts it: 129 strong, 31 weak, 60 negative, 220 scored patterns across seven categories. The 48 veto patterns are excluded, because a veto scores nothing and instead caps a category at zero, and folding them in would stop the breakdown summing to the total. 201 was true at 36a2f54, where the same walk gives 106 strong, 26 weak and 69 negative; the 2026-08-02 note above recorded it correctly and stays exactly as written. Rules were added through #523 on 2026-09-09. Nothing is retracted: the tree grew, and the number moves with the commit it was counted at.

• erratum · 2026-09-23

  Receipt 05 read “Rules and deterministic hybrid v3 gates both passed on 96 samples with macro-F1 0.9791”, with 2 of 96 misclassified, pinned at 36a2f54. It was true there. The baseline was re-recorded on 2026-09-07 and the gate now reads macro-F1 0.9896 with 1 misclassified out of 96, which prints as 0.990 at three decimals. Re-run at 0285675 on 2026-09-22 it exits 0 against its own 0.001 tolerance, and public backend-ci run 34429395505 prints the same line twice, once for the rules gate and once for the deterministic hybrid gate. The receipt is re-pinned IN PLACE rather than appended, so the anchor #v-jobtracker-5 that the home page and the evidence manifest point at does not move; the old value is quoted here instead of being deleted there. The receipt takes its own pin, 0285675, while receipts 01 to 03 and 06 to 09 keep 36a2f54: those are source audits nobody re-read, and moving one pin would restate them as if they had been. The artifact also changes from baseline_hybrid_v3.json to baseline_rules_v3.json, which is the file that decides the figure; the deterministic hybrid file repeats it, because that profile disables the learned layers.

• erratum · 2026-09-23

  Receipt 04 read 305 passed with 0 skipped, measured at 71b74f8 in backend-ci run 31152038153. True there. Re-read at 0285675 off backend-ci run 34429395505 of 2026-09-10: 3,747 passed, 0 skipped, 13 expected failures. The zero holds, which is still the part worth reading, and the 13 xfails are named beside the total rather than folded into it, because an expected failure is not a pass. The Postgres row-level-security module is 26 tests now and is inside that total; the separate rls-postgres job runs the same module again, so the two counts are one set of tests seen twice and are not summed. The receipt is re-pinned IN PLACE, keeping #v-jobtracker-4, and its pin moves to 0285675 with the number, as it has at every earlier step: 182 ⟶ 271 ⟶ 278 ⟶ 305 ⟶ 3,747, each true at the commit beside it. The row also drops its breakdown of the old +27, which described a move between two commits neither of which is the pin any more.

• erratum · 2026-09-23

  The evidence page put Applied’s line coverage at 53 percent, from 8,210 statements with 3,865 missed at 71b74f8. At 0285675, in backend-ci run 34429395505, the same pytest tests -q --cov=jobtracker reports 9,786 statements, 3,054 missed, 69 percent. Every per-package sub-figure moved with it and every one of them is restated rather than left standing beside the new total: cloud 93.0, auth 89.6, database 80.8, classifier 78.0, credentials 67.4, email_clients 42.7, scripts 40.0 over 2,434 statements, tracking 28.9, and the package root 92.2. The old row named cloud at 82.2, auth at 80.5, database at 76.6 and jobtracker/scripts at 2,240 statements and 33.7 percent; those were true of the 53 percent run and are kept here as its record. The under-claim is the interesting part: this file had been publishing a figure 16 points BELOW the work, which no gate could catch, because nothing goes red when a number is too modest.

• erratum · 2026-09-23

  The evidence manifest’s 3-layer classifier row said all three layers “run on the desktop app and in the browser Space”. Both halves were false when read today. The desktop client was de-scoped on 2026-08-12 and deleted, which this register recorded at the time without the evidence row moving with it, and the Hugging Face Space went private on 2026-08-15, which the note above records. The row now says what is true: the hosted web app runs the rules layer alone, and the learned layers are not deployed on any surface a reader can open. This is the second time a desktop claim outlived the desktop; the first was caught on this file, and the copy of it on /evidence/ went unfixed until today because no gate reads across the two.

• erratum · 2026-09-23

  Four places described the int8 browser export in the present tense: the architecture summary, fig. 2’s classifier node, the boundary row about embeddings and SetFit, and outcome 02. Nothing public has run that export since its weights were withdrawn on 2026-08-15, so the present tense was claiming a live surface. All four move to the past tense the résumé uses: in-browser inference was built, an int8 ONNX export of 90 to 23 MB served by transformers.js, measured on 2026-08-03. The sizes are unchanged and not retracted; Applied pins them as literals in scripts/readme_facts.py at 90,362,391 bytes fp32 and 22,843,695 bytes int8, and its own README says the figure is attested rather than reproducible and should be described that way wherever it is cited. That is what this file now does. Today the two heavier layers run in the evaluation harness and nowhere else.

• erratum · 2026-09-23

  The home page’s Applied lede said the app reads your mail “on a device you control”. That is false for the product this file documents. The Gmail connected app classifies on Vercel, by design: classifier/hybrid.py short-circuits to the rules layer whenever deployment is cloud, which receipt 09 argues. In your own browser is the /import and /demo path, which runs the layer 1 rules port and nothing else, and the desktop client that made the phrase true was deleted on 2026-08-12. The same sentence said Applied “defers to you when it isn’t sure”, which is only part of the behaviour: the hosted pipeline queues a verdict between 0.70 and 0.85 for a person and sets aside anything below 0.70, except an ATS-relayed verdict and a message retracting an offer already on the board. Both clauses are wrong as written, and the replacement has to name the 0.70 floor rather than imply a person sees everything.

• erratum · 2026-09-23

  Outcome 02 linked ml/browser/site at 36a2f54, and the 22.8 MB model.onnx in that tree still downloads today. Applied withdrew those weights on 2026-08-15 because they were fitted partly on a real mailbox, and removing them at its own head stops redistribution going forward. This file was undoing that: a pinned link is a permalink, so the row was re-publishing the exact artifact the upstream project had pulled, and the 2026-08-15 note above sends a reader to it in as many words. The row’s artifact is now ml/browser/export_onnx.py at 0285675, a commit after the deletion, where git cat-file -e 0285675c:ml/browser/site/model/onnx/model.onnx fails and the export script resolves. The note above is NOT edited, because this register does not rewrite its own history; what it describes was true, and this entry is where the consequence is recorded. Scrubbing the weights from Applied’s git history is the upstream project’s decision and is not claimed here.

[ appendix ] · § artifacts, indexed