Thesis: filling forms is an I/O-bound scheduling problem with a correctness core. The scheduling half keeps every browser tab busy. The correctness half guarantees that each answer comes from a source the user approved, or the job stops before submitting.

ATS	Fill	Submit and confirm	Special cases
Greenhouse (hosted + embed)	text, react-select comboboxes, radios, checkboxes, file upload, EEO wrappers	"Thank you" page; email security code step	hidden phone-country list filtered out of dropdown options; per-job code files
Ashby	Yes/No buttons, CSS-only required markers, follow-up questions (second pass)	resubmit once on "Missing entry"	bot-check → 24 h cooldown for all Ashby jobs
Lever	location autocomplete limited to your state, EEO selects	confirmation text	hCaptcha → human queue
Workable	text, radios, uploads	confirmation text, 60 s wait	Cloudflare Turnstile → human queue; REGEN_SKIP_ATS=workable when it bot-walls

Each adapter is a generator: it yields at every field (a checkpoint) and at every wait (yield <seconds>), which is what lets the scheduler interleave many forms.

For every field label the resolver tries, in order:

1. Per-job extras (for example an approved arbitration acknowledgement for one employer).
2. Your approved answer bank (profile/answers.json).
3. Ordered rules (~90 regex → preset rules). Order is semantics: acknowledgements before salary, SMS opt-in before email, "(City, State)" before country, generic "N years" before skill questions.
4. Derived answers computed from facts: total years vs "N or more years" (generic questions only), years lived in the US, time zone match, "able to start within N days" from the notice period.
5. Job note for "why us / cover letter", written per job from Fact Bank entries.
6. Drafted answer for open questions (drafts.py): sentences built only from the facts the job's resume chose, logged with fact ids in the event and in drafts_review.md.
7. Yes/no autofill policy: willingness questions → your onsite/relocation stance. "Have you / do you have" → Yes only if a Fact Bank entry backs it, else No. Status questions are never defaulted (sponsorship, authorization, citizenship, export, arbitration, salary, EEO, criminal, government, relatives, prior employment).
8. Nothing → the field is left for you and the job ends NEEDS YOU with the exact labels.

Special tokens: __DECLINE__ (EEO: pick the decline option), __ACK__ (pick an affirmative acknowledgement, never a negative one), __SKILLS__ (multi-select: tick only options your skills name).

Exact → prefix → contains → whole-word containment of a short option → sentence-form "No" ("I have never been employed by…") → preferred-location fallback → aliases (same fact spelled differently: "United States of America", "M.S.", visa-type spellings, careers-page synonyms) → salary buckets (a number picks the range that contains it: $80,000 – $100,000, 120,000+). If nothing matches, the field stays empty and the job goes to you.

Preemptive round-robin over REGEN_TABS tabs. A job runs until its time quantum is used (back of the ready queue) or it yields a wait (sleeps on a heap until its wake time). Quantum per ATS = EWMA mean + 0.84 σ of observed CPU bursts (80th percentile), clamped 2–20 s and persisted. Pick O(1), preempt O(1), sleep/wake O(log S). Measured ×2.05 throughput over sequential on a seeded simulation, ×2.5 with 4 tabs. Batches are interleaved by company so rate limits and bot checks see spread-out traffic.

• Sensitive fields (SSN, bank, date of birth, driver's license, ...) → FLAGGED.
• Unexpected site or redirect → FLAGGED.
• Legal answers that drift from presets → FLAGGED.
• Rate caps: REGEN_MAX_PER_DAY (default 25) and 3 per company per day.
• Kill switch file stops everything.

Every attempt ends with a full-page screenshot in workspace/proof/. A submission counts as verified only if the latest status is SUBMITTED and the proof file exists (engine report).

Related: Email codes · Tailoring · Truthfulness and safety · Algorithms · Configuration