Graph Climbing

Graph climbing is my name for the loop that works a claim graph to done while the software gets built; I published the protocol under that name. One climb does five things: derive the claim frontier, the open claims whose dependencies are all verified; pick one bounded piece of work from it; build; run the tests against the exact commit that was built; and record what passed. It is a control protocol rather than a search algorithm: a step up means one claim’s test now passes, not a score improving, and the graph being climbed is itself revisable.

A hardening loop runs inside each climb: repeated adversarial review rather than one strong pass. Each pass fixes the exact commit under review and tries one concrete way it could fail: a partial write, a replayed request, two writers interleaving. Each finding gets an exact location and its own verdict: accepted, rebutted, or deferred. A correction changes code, claims, and evidence together, and the next pass reviews the new commit, not the remembered one. The loop stops when no accepted actionable finding remains or a named external blocker applies. A recurring failure can become a permanent regression test, only on repetition, because a single finding may be an accident of one implementation.

The method depends on that revisability, because nothing inside a single climb checks whether the graph cut the work up correctly. The graph, not the loop, carries the correction: every claim names a test that could prove it false, at least one claim states what must not be true, a review finding that checks out can reopen a closed claim or add one nobody wrote, and each climb re-derives the frontier from what is currently true instead of following yesterday’s plan. Human steering enters the same way, by changing the claims, their dependencies, their tests, or which sources count, rather than by naming the next step. A wrong breakdown survives only until a test hits it, which is why a climb ends by checking each result against its claim’s test and recording the outcome.