Identify what each number actually represents
A local transform duration measures how long the defensive tool took to produce an artifact. A provider duration measures an analyzer call. A token estimate models expected model traffic. A reconstruction duration requires a defined task, an actual agent run and an objective completion rule. These quantities cannot be substituted for one another.
The pinned BLA estimator computes modeled tokens, cost and wall-clock seconds using assumptions about tool calls and per-call latency. It is useful as a model to inspect, but it is not a transcript of an LLM session. Its named pricing and latency constants also belong to that source revision rather than a current purchasing quote.
Read our local measurements narrowly
The committed CloneGuard web fixture records seven audit runs on four synthetic files, a staging duration and integrity checks. The BLA comparison records artifact sizes and format-specific REA outcomes. Neither result includes an agent rebuilding a complete application, hidden reconstruction tests or a measured customer attack.
Do not multiply those local durations by a token ratio and call the result attacker hours. A model can make fewer calls, use caching, choose a different tool or abandon a route. An analyst may spend time reviewing results while an automated agent is idle. Those differences require observation rather than extrapolation.
Design a useful time-to-task experiment
Define the deliverable first, such as a runnable replica that passes a hidden behavior suite. Start timing at a documented event and stop when the submission is frozen or the trial budget expires. Record setup, active agent work, tool waiting, human intervention and evaluation separately.
Use matched untreated and protected fixtures, repeated trials and the same available tools. Save actual usage metadata and error events. A timeout is a censored observation: it means the task was not completed inside the budget, not that completion would take exactly that many hours.
Publish an honest hours table
An evidence table should include task, specimen hashes, tool and model revisions, trial count, successful completions, elapsed time distribution, actual usage and compatibility overhead. Show incomplete trials and failures rather than excluding them from an average. Where no trial ran, write not measured.
For the current release, the correct headline is that release-hygiene checks and several analyzer probes were recorded. A claim that CloneGuard adds a particular number of hours, bankrupts an agent or prevents a weekend clone would exceed the available evidence. Future measurements can replace the unknowns once the experiment exists.
Practical checklist
- Define the reconstruction completion rule.
- Keep modeled numbers visibly labeled.
- Record timed-out and incomplete trials.
- Do not claim unmeasured hours saved.
Sources and further reading
Primary references support the technical details. Procedures here are guidance; only explicitly recorded project results describe tests we ran.
Prepared with AI assistance. How this library is produced. Suggest a correction; keep private code and credentials out of public issues.