Maxify slice · one governed hiring workflow

Tl;dr · what this is by Hadi Al-Hazim

A working slice of what I read Maxify to be, built on the workflow named in this hiring post and run on the live Software Engineer vacancy. The model extracts evidence from each application. A deterministic gate verifies spans, redacts protected attributes and routes. Every advance lands in a human queue. Eight runs are recorded below; one of them is my own application to this role.

The eight recorded runs are drawn from their transcripts, like everything else on this page. Three end at intake for zero tokens (injection blocked, duplicate pointed home, format declined); a human reads every advance.

§1

Company Brain: everything the workflow knows, in five inspectable files

Each file is served read-only and linked here. Every gate rule cites a published source below; founder-published figures are cited as published, and this demo does not audit them.

Public signals the rules trace to

§2

Governed Agent — the model extracts. It never decides.

The workflow runs four stages with one model call; the expensive step is fenced on both sides by deterministic checks.

2.1 Intake · deterministic

Duplicate hash, injection scan, subject format and protected-attribute redaction all run before the model; three of the eight recorded runs end here at zero token cost.
2.2 Extract · the only model call on the page

Structures every claimed system with a verbatim quote and any artifact URL. Instructed to extract boldly and judge nothing; the application text is data, never instructions.
2.3 Gate · deterministic

Verifies each quote is a real span of the source, clears invented artifacts, flags metrics that carry no evidence as written screen questions.
2.4 Route · deterministic

A pure function of gate outcomes. The model never scores, ranks or compares candidates; a human makes every decision.

§3

Control Plane: the rulebook, with its false-positive discipline

A scanner that blocks a candidate for managing system prompts, or a redactor that eats "single sign-on", fails real applicants quietly. Every contextual guard here earns published negative cases in the eval table.

Table 3 · the gate rulebook, one row per rule, false-positive discipline noted where it applies. rendered from gate_rules.json

Table 3b · escalation routes: where every run lands, and who reads it. rendered from escalation.json

§4

Outcome Loop: eight recorded runs, one eval table

This section renders only from saved transcripts and the eval report; nothing in it is hand-written. Fixtures fx-01 to fx-07 are synthetic and labeled. Run 08 is my real application, processed with no special treatment. Each row opens its run sheet; each sheet links its raw transcript.

Runs 01–08 · one row per recorded run; click a row for the full sheet. rendered from runs/run-01..08.json

Table 4 · golden gate cases with expected and observed outcomes.

Gate evals reproduced, including the negative cases that prove the guards hold. Deterministic lane; the set runs free in CI. rendered from runs/evals.json

§ LIVE

Live lane: type an application, watch it run

This is the same pipeline, streamed. The lane opens on a capped budget; adversarial pastes are contained at intake for free. Load a sample or write your own.

application checking the lane…

subject line · the ad mandates this exact format application text

budget backstop

pipeline intake → extract → gate → route

▸ lane state loads when the page opens. When closed, the recorded runs above are the same pipeline.

Adversarial pastes are contained at intake, before the model runs, at zero token cost.