Skip to content

Research

An explanation. Each procoder domain exists because of a belief about how software work goes wrong. This page states those beliefs and, where external evidence exists, cites it. Where it does not, this page says so rather than implying a backing that is not there.

That rule matters more than the citations. A page of references is easy to write and worth nothing if any of them is misapplied, so every entry below is one of two kinds: cited, meaning a real source was read and says what is claimed, or uncited, meaning the premise is procoder's own design judgment and is labelled as such.

Correction policy

If a citation here turns out to be misapplied, misquoted, or to not say what was claimed, it gets corrected in place, with the correction visible — struck through or marked, never silently deleted. A page that quietly drops its mistakes teaches its reader nothing about how carefully the rest was checked.

No corrections have been issued yet. This section is here so the first one has an established place to go, not because the page is assumed correct.

Cited

Agents claim completion they have not reached

Premise behind: the quality controllers, evidence-gated closes, procoder test's "NOT run is never green".

Evidence: SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks — Orlanski, Roy, Yun, Shin, Gu, Ge, Adila, Roberts, Sala, Albarghouthi (arXiv:2603.24755, submitted March 2026, revised May 2026). Across 36 problems, 196 checkpoints and 15 coding agents, the best checkpoint pass rate was 14.8%, and no agent solved any problem end to end. Structural erosion increased in 77% of trajectories and verbosity in 75.5%; agent code measured 2.3× more verbose and 2.0× more eroded than 473 open-source Python repositories.

How it is used here: as evidence that iterative agent work degrades measurably, which is the premise behind refusing to accept a claim of completion in place of evidence of it. Note the version caveat: earlier and later revisions of this paper report different problem, checkpoint and percentage figures. The numbers above are from the revised abstract.

Prose guidance shifts the starting point, not the drift

Premise behind: procoder shipping a binary rather than only an AGENTS.md. See Where procoder sits.

Evidence: the same paper. Explicit quality guidance reduced initial verbosity and erosion by up to one third — without affecting the rate of degradation.

How it is used here: it supports the split procoder is built on, and it cuts both ways honestly. Guidance demonstrably helps; it demonstrably does not stop the drift. That is an argument for keeping both halves, not for dismissing the advisory one.

Developers do admit debt in comments, and much of it stays

Premise behind: procoder debt harvesting debt: markers, and flagging a marker with no revisit condition as rot.

Evidence: An Exploratory Study on Self-Admitted Technical Debt — Potdar and Shihab (Semantic Scholar), studying Eclipse, Chromium OS, Apache HTTP Server and ArgoUML. Self-admitted technical debt appeared in up to 31% of files, and between 26.3% and 63.5% of it was removed after introduction.

How it is used here: the practice procoder relies on — engineers writing down the corner they cut — is empirically real and widespread. The removal range is the more useful half: on the low end, roughly three quarters of admitted debt was still there. A comment alone does not get debt paid, which is why procoder harvests markers into a ledger and treats one without a revisit condition as rot rather than as a record.

Uncited — procoder's own judgment

These premises have no external citation here. Some may have supporting literature nobody has looked for yet; one has literature that does not settle the question. Either way, no claim of research backing is made.

Multi-lens review catches more than a single pass

Premise behind: procoder review's five lenses.

Status: contested, deliberately not claimed as support. The nearest literature compares checklist-based against ad hoc code reading, and it does not agree with itself. A controlled study in a distributed groupware environment found no significant difference in defect detection effectiveness, effort, or false positives between the two (arXiv:0909.4260). Other experiments report checklist-based reading as substantially superior. Reviews of this literature describe the results as inconsistent and conflicting.

Citing only the favourable half would be exactly the failure this page exists to avoid. The five lenses are a design judgment, and the honest statement is that the external evidence is mixed.

Formatting, hygiene and secret drift is costly enough to gate on

Premise behind: the commit gate itself.

Status: uncited. No study has been read that measures the cost of this specific drift in agent-written code. The gate is justified here by mechanism rather than measurement: an unformatted file, a committed secret and a junk artifact are each objectively identifiable, and identifying them at the commit is cheaper than at review.

Specification before implementation reduces rework

Premise behind: the spec → plan → backlog → sprint chain.

Status: uncited. There is a substantial requirements-engineering literature on defect cost escalating with the phase in which a defect is found, and it has not been read and verified for this page. Until it has, the chain stands on procoder's own reasoning.

What procoder has not measured about itself

Nothing on this page is evidence that procoder works. No benchmark of the gate's overhead against defects caught has been run — see Honest limits, and #190, which exists to make that measurable rather than assumed.