The problem
Applying for jobs at any volume is a loop, and a loop left alone games its own metric. Point an agent at "apply to more roles" and it will optimise the number it can measure, applications sent, while the thing that actually matters, a role worth taking that might say yes, quietly detaches from that number. The failure is not a bug in the loop. The loop is doing exactly what it was built to do on a target that stopped meaning anything.
The fix is not a faster autopilot. It is structure around the autopilot, and a human holding the one reference a loop can never set for itself: whether a given application is worth a person's time to send.
What I built
The vault is a private system that keeps the fast, cheap part of the work honest by fencing it in with slower parts it does not control. It has a handful of parts, each doing one job.
A sourcing autopilot runs twice a day, finds roles and grades each one against fixed criteria, and drops the graded shortlist into the vault. It is the fast loop, and it is deliberately capped: when more than a handful of packs are waiting on a human decision, drafting freezes by contract. The bottleneck that matters is the person at the desk, not the search, so the search is not allowed to outrun them.
A review desk is where I sit and decide submit, fix or kill on each graded role. Nothing leaves the vault without passing through that decision. The desk owns the autopilot's reference, not the other way round.
canon is the single source of truth for the facts about my work, one file that every pack is built from, so a claim can only appear in an application if it is true in one place first.
For a role marked submit, canon feeds the tailored engine, the open-source core of this whole system and the only public part of it. tailored wraps a language model in deterministic gates: the model proposes the prose, and the gates decide what may ship. A page count is a fact, not an opinion, and neither is a lint rule or a leaked protected term, so the gates are anchors the drafting loop cannot argue its way past.
A tracker holds the state of every role, and a learning loop folds the edits I make at the review desk back into the drafting rules, so the next draft starts closer to something I would send.
Judged by something other than itself
A drafting loop grading its own output is a loop marking its own homework. So the vault never lets a quality metric travel alone. Keyword coverage against a posting is paired with an independent read of substance, because a pack can be dense with the right words and empty of the right evidence, and only the pair catches that. Salary targeting is per pack and structured, and when a role sits below the floor the panel says "below floor" rather than inventing a number that would feel like progress. Refusing to answer is treated as the honest answer, the same way the other systems on this site abstain rather than fabricate.
Outcome
The vault runs my own job search, and its applications go out through its own gates, so the system is field-tested on my livelihood rather than on a fixture. That is the only test I fully trust: the machinery has to keep touching the reality it claims to improve. The vault stays private because it is built on my personal facts and history; the engine at its core, tailored, is open source, and it is the part of this you can read.
Tech
TypeScript throughout, sharing the tailored toolkit for the gate layer. The autopilot, review desk, tracker and learning loop are private; the facts live in a gitignored canon file and never leave the machine. No employer, role or application detail is described here, by design.