The senior engineer hiring shortage: the part you can fix this quarter is in your own inbox.

It is both, in roughly that order. The macro shortage is real and well-documented for 2026: every public recruiting research piece this year cites the same shape. Universities turn out about 65,000 CS graduates per year while the market needs roughly 180,000 AI-capable engineers. Senior retirements are removing about 18 percent of the experienced pool. H-1B restrictions are pulling another 15 percent out. Three jobs per qualified senior, give or take. That is the macro layer, and it is not something one TA team can move. The operational layer is what you can move. On any given posting in 2026 you receive 200+ inbound applicants per technical role, a meaningful fraction of those are AI-fabricated or AI-rewritten, and the genuine senior applicants are mixed into that pile with the same surface gloss. The shortage you can act on this quarter is not the supply curve. It is the signal-to-noise ratio inside your own inbox.

Senior signal is the small set of testable resume facts that distinguish someone who can operate complex production systems from someone who cannot. On-call rotation in the last 3 years. Shipped a distributed storage or queueing system to production. Owned an incident review for a paged outage. Code-reviewed a large repo. Led a SOC 2 Type II audit. Authored a CSP rollout. Each one is a sentence, each one is checkable against a resume span, and each one is something a smart hiring manager would actually weight differently than a smart-but-different hiring manager. That structural fit is exactly why a claim ledger works for this. A 0-100 fit score throws the structure away. A claim ledger keeps it: 5 to 15 sentences, each with a visible weight, each with a resume span pinned to it, the score is the weighted sum, the rubric is the artifact.

The recruiter pastes the JD plus any org-wide hiring criteria. The agent returns 5 to 15 candidate claims classified must-have, nice-to-have, or red flag, each with a proposed weight. The recruiter promotes, demotes, deletes, or adds claims, then saves the rubric to the req. match.rerun_score then runs every inbound applicant against the saved ledger. The output is per-claim verdicts plus a weighted total. The middle band, typically 0.5 to 0.7, is where the recruiter spends actual reading time, because that is where the model is least sure and where senior signal is most often hiding behind a phrasing mismatch (object storage instead of distributed storage, paged on-call instead of on-call rotation, and so on). The recruiter promotes the spans the model missed, fires match.override_weight when the team values a claim differently than the JD implied, and the score recomputes.

Because the override is the part a regulator, a hiring manager, or a candidate later asks you to defend, and a comment cannot be defended. match.override_weight(claim_id, prior_weight, new_weight) writes to an audit log with actor, timestamp, prior, and new. The score recomputes from the ledger. If the override flips a candidate from rejected to advance, that flip is logged with the same shape. There is no 'the AI changed its mind, I trust it' moment to defend a year later. The override IS the defense. NYC Local Law 144 wants it. Illinois HB 3773 wants it. Colorado CAIA wants it. The EU AI Act high-risk hiring rules want it. A black-box similarity score with a prose summary does not give it to you.

Yes. The claim shape is content-agnostic. For a senior security engineer, claims look like 'led a SOC 2 Type II audit', 'authored a CSP rollout for a public web app', 'has experience with eBPF-based runtime monitoring'. For a senior ML engineer, 'shipped a production transformer fine-tune at scale', 'has familiarity with vLLM or TGI inference', 'has paged on-call experience for a model-serving stack'. For a senior data platform engineer, 'shipped a sub-second OLAP layer at scale', 'owns an Iceberg or Hudi adoption'. The constraint is not the role; it is that each claim has to be testable against a resume span. Soft claims like 'is a strong communicator' are explicitly the wrong shape for this layer; they belong in the panel, not the rubric. The agent declines to write them.

Different artifact, different floor. Ashby and Gem ship LLM resume summarization plus a fit score bolted onto an ATS, with a paragraph the model wrote about its reasoning. The rubric is implicit in the model and cannot be edited as a structured object. Eightfold is enterprise, no sourcing, and the matching layer is a deep-skill graph the buyer never sees as a list of weighted claims. SeekOut moved its agentic tier behind a gated enterprise contract and is in roll-up mode. None of them treat the override as a tool call with an audit log entry. Chosen Match Rating ships the ledger as the artifact, the override as a logged tool call, and it does it on the $0 Starter plan with the same scoring on every tier above it. No demo required to see the price.

Match Rating is a pre-interview layer; it will not catch a synthetic candidate inside a video call by itself. What it does is shift detection earlier. If the senior-shaped claims do not have evidence spans, or if the spans contradict each other (a 'staff engineer at FAANG since 2019' next to a graduation date in 2024, or 'led a distributed storage team' next to no employer with such a team), the red-flag claims surface in the ledger before the recruiter spends an hour on a phone screen. A claim that cannot be sourced to a specific resume span is the earliest fraud signal you have, and a single fit score hides it. Ashby shipped Fraudulent Candidate Detection in April 2026 for the same problem class; Chosen approaches it from the rubric side rather than the runtime side, and the two layers compose.

It is the workflow that lets a small team run 12 to 25 reqs at all. The math: 200 inbound per role, 12 reqs, recruiter eyes that can cover maybe 40 senior-shaped resumes per day if the rest of the inbox is on autopilot. Without a ledger, you either rubber-stamp a vendor fit score (in which case you cannot defend any reject) or you read every resume yourself (in which case you are not also doing scheduling, sourcing, or analytics). The ledger is the autopilot for the easy 80 percent and a structured reading lens for the hard 20. That is why the same recruiter holds 25 reqs after Match Rating that they could not hold before.

Starter is $0, up to 3 open reqs, every agent included, no credit card, no demo. Match Rating is on Starter. The MCP server is on Starter. Sourcing, Scheduling, and Analytics are on Starter. Growth is $99/mo for founding members and $399/mo afterward, unlimited reqs, the same scoring. Enterprise is custom for teams above 250 and adds the published bias-audit artifact, per-jurisdiction toggles, SSO, dedicated implementation. There is no 'AI add-on' upsold at renewal. The same rubric is the same rubric on every plan.

The 65,000 CS grads vs 180,000 AI-capable demand split, the 18 percent senior-retirement figure, the 15 percent H-1B reduction, and the roughly 3-jobs-per-senior ratio are all drawn from public 2026 recruiting research published by analyst groups, third-party hiring blogs, and engineering staffing firms tracking the same trend. None of those numbers are Chosen's. Chosen's numbers are the ones tied to the rubric shape: 5 to 15 testable claims per JD, 200+ inbound per role, $0 Starter plan with full Match Rating, four named agents on every plan, four jurisdictions the ledger satisfies. Those come from the product itself.