debugging-ai examples on Programmer.ie: Modern AI programming

Find the first divergence, then move the failure with an override

Mon, 01 Jan 0001 00:00:00 +0000

Debugging is not making a symptom disappear. It is finding the earliest point where execution left intent, and proving that point causes the failure by changing it and nothing else. This program does both halves on the chapter’s billing example: an intent table read in order, and a counterfactual override forcing the discount to the value the spec requires.

Run it

Save this as invoice_divergence.py and run it with Python 3.11 or later. It uses the standard library only.

Stop at the first stage boundary that disagrees with intent

Mon, 01 Jan 0001 00:00:00 +0000

The final number is wrong and the last stage is where it is wrong, so the obvious fix is to patch the last stage. That cannot work: three different causes produce the same wrong total, and only the ordered intermediates separate them. This program checkpoints a pipeline in execution order, stops at the first boundary that disagrees with the spec, then takes the backward slice to name the rows responsible. It composes the chapter’s clean stage and contributing_orders slice over a fixture whose shortfall is €212.40.

A boundary table convicts a rule, not a single input

Mon, 01 Jan 0001 00:00:00 +0000

Forty-seven passing tests and the refunds are still wrong. The suite samples middles, and middles never land on a seam. This program runs the chapter’s six-cell boundary table against a page-count function and reads the shape of the divergences: one failing input is an observation, a pattern of them is evidence about the rule. It composes the chapter’s paginate, its defective chunk_count, and the metamorphic relation the chapter proposes as oracle.

Group the trials by a signature, not by the raw text

Mon, 01 Jan 0001 00:00:00 +0000

One draw from a mixed distribution carries almost no information about the distribution, so the debugging object is the answer-distribution and not the answer. It is read from a fixed projection of each output, and that projection has to be chosen before the trials or the histogram is a Rorschach test. This program composes the chapter’s Signature dataclass and its behavioral_signature projection, then runs a seeded twenty-draw series at temperature 0.7 and a ten-draw probe at temperature 0. The draws come from a simulation, not a model.

Price the design's own numbers before anyone builds it

Mon, 01 Jan 0001 00:00:00 +0000

A design document that claims “meets latency goals” while listing per-hop latencies that sum past the budget does not need a load test. It needs adding up. The chapter’s method is a constraint table where each row is a bound, the design’s own line items, and a verdict, and where one FAIL row rejects the shape and one UNVERIFIABLE row rejects the document. This program prices two candidates on identical rows.

Detect a loop from hashes, then name which kind it is

Mon, 01 Jan 0001 00:00:00 +0000

An agent can burn ninety steps and never move. Step counts measure effort, not progress, so a guard that watches the step budget watches the wrong thing. The chapter’s answer is a novelty metric computed from hashes over a pre-registered window, plus three repetition signatures that name the shape of the non-progress: an exact loop, a thrash, and a retry storm. This program composes the chapter’s window_novelty and repetition_shape detectors over three constructed windows plus a progressing one.