Heat model and evidence boundaries

Reads are lifetime evidence and never decrease. Heat is current attention: each timestamped read contributes 0.5^(age_days / 30), so it halves every 30 days. Folder heat sums current file heat. Cold means inactive now, never obsolete.

The health score combines coverage, router reachability, and search behavior. Its A–F grade is provisional until the dataset is mature (at least 100 reads and seven observed days). Correlation is not causation: a routing change followed by fewer searches does not prove the change caused the improvement.

What can be observed

• Native Read events, explicit Glob/Grep paths, and explicit rg/grep/find documentation targets.
• Expanded Bash reader paths in supported Bash sessions, preserving command behavior.
• Explicit file-like MCP parameters; never remote HTTP targets or response content.
• Subagent attribution and invoked skill names.

Boundaries

• Injected context is invisible to Read telemetry and appears as always loaded.
• Instruction adherence does not infer that injected text was understood. It evaluates only user-confirmed deterministic probes over observable events. followed is supporting evidence; unobserved means evidence was not captured, never violated.
• Agent personas follow the same rule as directives. A definition's name and description sit in the system prompt on each request and its body only when the agent runs, so the two are reported separately rather than charging whole files to every request. A persona never invoked is recurring cost. It is only reported as never-invoked over sessions where agent capture was running; otherwise the status is awaiting-capture and no cost claim is made.
• Always-loaded cost counts every file injected through a CLAUDE.md import chain, not only watched documentation, because injected content is billed regardless of its name.
• Zero probe opportunities is never-triggered, not 0% adherence, and only over sessions where the probe's capture was running; otherwise it is awaiting-capture. A path_avoided rule nobody breached is no-violations-observed, not unused. Capture-disabled and unobservable directives are excluded from rates. Negative evidence that crosses a compaction boundary is degraded and excluded from the primary rate.
• Adherence and heat mature differently. A directive rate is provisional below five usable opportunities; the overall dataset separately remains cold-start, warming, or mature.
• Glob/Grep counts require an explicit path or static directory prefix; scan telemetry undercounts by design.
• A read proves discovery, not comprehension, correctness, or compliance.
• Untouched and dead-path candidates are review prompts, not removal recommendations. Protected, referenced, critical, safety, and template paths remain distinct.
• Hosted tools that bypass local lifecycle hooks are invisible. Other local tools can use the documented ingest entry point.
• The health grade scores evaluable docs — always-loaded context is excluded by design, and the structural inventory follows git's view of the repository (tracked plus untracked-but-not-ignored files).
• Directory symlinks are never followed: a watched surface behind one stays outside the inventory and the score. Such surfaces are named explicitly — unfollowed_surfaces in the stats payload, a warning in tt doctor, and a health driver — so a low score is never mistaken for a verdict on knowledge the measurement cannot see.

See the glossary for canonical definitions. The full probe contract is in instruction adherence.