We audited our own codebase, and published the grade
Most vendors will tell you their code is reviewed. Few will show you the review. In July 2026 we ran a staged audit of the whole LedgerTM repository, one stage per session, with the findings landing in a decision record as they were found so the document was useful at every stopping point. This is what it covered, what it found, and what we did about it.
Five stages. Structure and cross-dependencies: does every feature project follow the sanctioned shape, and does anything reach across the contracts seam. Dependency health: known vulnerabilities, outdated packages, licence risks, pinning. Complexity hotspots and dead code, with the money path read line by line. Error handling and performance: N+1 query shapes, full-table loads, sync-over-async, unbounded growth in aggregates, missing indexes, silent failures. Then a synthesis: themes, what not to fix, a bite plan, and the owner’s decisions after a grill.
The grade was A minus, and the summary line explains why: a disciplined codebase whose defining strength is machine-enforced convention, and whose every serious finding sat precisely in a guard’s blind spot. Twenty of twenty-four feature projects followed the contracts-only shape exactly, and the two that did not were exactly the two the architecture test’s classifier exempted. Zero known vulnerabilities in the .NET packages and exemplary pinning, but one test library had drifted onto a commercial licence, and the two load-bearing frameworks were several minor versions behind. Complexity concentrated where the money is: the accounts module topped churn, size and branch density, with the instalment solver inside a class carrying five responsibilities.
The worst finding was a silent one. Posting a taxable transaction with no tax rate configured zero-rated it without a word, so a configuration drift could under-tax a whole tenant quietly. Beside it, two day-end steps reported success while discarding their failure counts. Silent failure had been driven out of tenancy by earlier work and had moved into return values and fallbacks. That is the pattern an audit exists to catch: the guard rails were good, and the findings were where the rails stopped.
What we did. Seven sessions, in the order the audit set: the safety net first, the architecture-test classifier fixed so the dependency rule is enforced for every project, warnings on the projections’ silent-skip branches, environment-gated messaging providers. Then the quick wins: three missing indexes, a settings cache that no longer opens a database session on a hit, the licence fix, a dead-code sweep. Then correctness: the failure counts threaded through day-end, and the tax throw made visible by them. Then the framework catch-up on its own worktree, the accounts hardening, the solver extracted behind a parity fixture, and a measurement gate against the ten-thousand-accounts-per-tenant target.
What we chose not to fix, and wrote down: read models kept where they are rather than moved into contracts for a wide blast radius and no gain, and a handful of hotspots the team already knew and had tests around.
Why publish it. Two reasons. A buyer who takes the source deserves to know what the last independent read of it said. And an audit that is graded and then acted on, with the fixes traceable bite by bite, is the honest form of the claim every vendor makes. The next one runs when the model has moved enough to be worth reading again, and its grade goes here too.
Written by Jeremy Rich.