Plansmith
PlanSmith Benchmark

Job board — the model build-off

We hand the same job-board planner to four coding models, lock the discovery answers and the feature ledger, then audit what each one actually shipped. A job board lives or dies on one thing: an employer must never reach another employer’s applicants, and a candidate must never reach anyone’s. Every score below is that boundary, tested live.

2026-08-014 models · 4 builds audited
The build-off · 01

How the builds ranked.

01

Roleway Jobs

Best Isolation
by GPT-5.6 Sol
93/100

The one to copy if you are handling other people’s applicant data.

Best at
Almost every page locked down with no real gaps, and we proved in both directions that one employer cannot reach another's applicants. Repeated sign-in attempts are throttled, passwords are stored properly, and every change is checked for where it came from.
Watch for
Our review concentrated on security, so we did not drive the day-to-day workflows or judge the interface end to end.
Open live demo
02

Brightwork

Clearest Proof of Tenancy
by Claude Fable 5
91/100

Best when you want the boundary demonstrable, not merely asserted.

Best at
The only build that ships with several real employers already set up, so you can see the boundary holding in both directions rather than taking it on trust. Signing in through an outside provider cannot get you an administrator account, and access to a candidate's details stops the instant they withdraw consent.
Watch for
Two internal maintenance functions are reachable from outside and trust whatever identity the caller claims, which means an entry in the activity record could be faked. Nothing anywhere limits repeated attempts.
Open live demo
03

Northline

Most Complete Marketplace
by Grok 4.5
83/100

The broadest feature set, once the reporting path is closed.

Best at
The broadest marketplace — approval queues for new employers, moderation, paid placement and a candidate portal — with applicant separation verified against a control, and it compiles cleanly for production.
Watch for
Its “report this listing” form applied the outcome immediately, so any passing visitor with no account at all could take a paid listing off the board.Fixed since auditreports now wait for a moderator instead of hiding the listing, demo sign-in is switched off by default, and repeated attempts are throttled — all after our audit.
Open live demo
04

TechHire Direct

Fast-Tier Build
by Gemini 3.6 Flash
58/100

Capable of the features; needs the security layer specified for it.

Best at
A genuinely working job board — real search and filtering, employer tools, admin reporting, and a sensible way of organising its data.
Watch for
It shipped with no sign-in and no permissions whatsoever: candidate contact details, a full download of the database, and administrator actions were open to anybody who found the address.Fixed since auditpassword sign-in, deny-by-default permissions, per-record checks and a test to stop it regressing were added over several rounds and independently re-verified.
Open live demo

Models tested: GPT-5.6 Sol, Claude Fable 5, Grok 4.5 and Gemini 3.6 Flash. Scores are as found at audit; later fixes are noted but do not change the number. Gemini 3.6 Flash is a small, fast-tier model and is not directly comparable to the larger three. Isolation was proved with a control — the owning employer must actually see its own applicants, so that an empty result for everyone else means real filtering rather than an accident. These grade correctness, not visual design.