On this page

What a number can promise

We open a repository because something must change. Perhaps a price is wrong, an interface must move, or a service can no longer satisfy its contract. The files are visible, but the task is not contained in their line count. We must discover what the change can disturb and what evidence would justify accepting it.

A counter can tell us how much text is present. A dependency model can tell us which declared assumptions may be invalidated. An evidence model can tell us which declared checks still need support. An economic scenario can attach resource prices to assumed activities. Each statement is useful precisely because it has a different meaning.

CCM keeps these meanings separate. Its name is Code Counter Machine, but its design begins where ordinary counting becomes insufficient. The familiar table remains the entrance to the system. Beyond that table we find a finite Change and Certificate Calculus, an explicit account of accepted evidence, and a research profile for AI development costs. We can inspect the inputs to every stage.

The word confidence in this book means a justified position within a stated model. It does not name a probability that an arbitrary program is correct. An accepted test may exclude some modeled outcomes while leaving many unmodeled behaviors untouched. A passed check can be useful without becoming a proof of everything we care about.

We will follow a small checkout change throughout the book. Pricing and inventory change together. Checkout depends on both, while a payment boundary remains unchanged. This example is deliberately small enough to calculate by hand and large enough to expose shared checks, additional joint checks, invalidated evidence, and a reusable result. The source fixtures and task model live under examples/book. The screenshots are generated by running CCM against those files.

How the argument develops

We begin as readers of a report. We learn the inventory’s units, compare two source snapshots, and declare a task model. Then we descend into the implementation: file selection, bounded workers, lexical state machines, graph closure, obligation activation, and evidence transport. The algorithms become easier to assess once we have a concrete result to explain.

The mathematical chapters state assumptions before drawing conclusions. A closure law follows from a fixed directed graph. A residual-work monotonicity law needs additional conditions on evidence. A cost equation follows from declared resource quantities and prices, while the accuracy of those quantities remains an empirical question. These distinctions matter more than the sophistication of the notation.

The later chapters examine estimation, experiments, architecture, and proposed research. They distinguish mechanisms shipped in the executable from candidates in the design records. Tree-sitter extraction, authenticated evidence, richer code property graphs, persistent topology, and imported work ledgers are research or engineering directions, not hidden capabilities of the current CLI.

Three readers, one account

A developer can follow the worked example and then read the algorithms that explain a particular result. A planning team can begin with the report and economic chapters, where every quantity receives a unit and every default receives a status. A domain expert can follow the propositions, counterexamples, evidence interaction, and research gates. We use the same definitions on all three routes.

We use short names in the interface and precise names in the model. An artifact is a modeled part. An obligation is a required check. A certificate is an evidence record admitted by the declared policy. The technical terms remain available because they help us state exact claims; the report uses ordinary language because a result must be understandable before it can guide a decision.

This edition describes CCM 1.0.0-beta.1. The counting policy is lexical-v5, the preferred task input is model version 2, and the economic profile is the frozen 2026 R&D scenario. Performance measurements identify their own executable and workload. A version number on this book does not retrospectively change an older experiment.

The useful promise is modest and concrete. We can make a measurement reproducible, a finite inference inspectable, and an assumption replaceable. Those are the foundations on which stronger claims can eventually stand.

Vikrant Rathore wrote CCM and this book, with assistance from Ronak Rathore and Kanak Rathore. CCM is free software under the GNU Lesser General Public License, version 3 or any later version.