Every claim here is bounded to one pinned source review. A later product state needs new evidence; confident wording cannot substitute for it.
A claim is a sentence about the product. Evidence is the observed source, action, check or bounded case that gives the sentence support. This page places a stop bar between that support and the larger outcome a reader might understandably infer.
For the order app, “a fresh session followed the pointer to the active plan” may be supported. “Coldstart prevents context loss” is not. The second sentence promises a wider outcome than the observation can establish.
The ordinary shortcut is to see a useful mechanism, imagine its likely benefit and publish the benefit as the product claim. That is quick, and for private notes it may be enough. In public documentation it transfers an untested inference to the reader. Coldstart's alternative is slower: name the observed condition, attach the evidence and write down the attractive conclusion that the evidence does not settle. The extra ceremony is useful only when the claim matters enough to need that boundary.
Ask three questions
- What exactly is the claim? Write it so a reader could tell what would make it false.
- What evidence bears on it? Name the inspected mechanism, walked action, planted defect or authored case.
- What does not follow? Name the tempting broader conclusion and stop before it.
| Claim | Evidence | Stop before |
|---|---|---|
| The reviewed pointer carries the active work address | Pinned source and represented pointer tests | No session ever loses relevant context |
| A selected check reports its planted defect | The named defect-injection case goes red | The harness catches every defect |
| The initializer can undo its represented seed | The pinned round-trip cases | Existing documents can be migrated without loss |
Changing the third column into a product promise requires evidence designed for that exact promise. It is not a copy edit.
Mechanisms support mechanism claims
The pinned source contains mechanisms for finding current work, preparing and closing a session, filing project facts, selecting reusable guidance, applying selected boundaries, running named checks and rebuilding generated views.
Those mechanisms can be inspected. A walked smoke action can show that one entry path ran on one coding agent. A defect-injection test can show that one checker detects one represented break. A generated comparison can show disagreement between a source and its represented view.
The nearest limit belongs beside each observation. None of those facts proves that every relevant path ran, every check is strong, every project fact is current or every decision improves because the mechanism appeared.
Evaluation stays inside its cases
The pinned report measures selected routing, memory, skill-firing and claim-grounding cases. On the authored routing set, the Coldstart mechanism and a plain word-matching comparison are tied.
That supports a sentence about that set. It supports neither routing superiority nor inferiority on other tasks. The other results are similarly narrow: they describe committed fixtures and the predicates the scoring code can inspect.
They are internal mechanism measurements, not studies of productivity, code quality, project success or user outcomes.
Release state needs dated evidence
This handbook describes an unreleased source state reviewed on the date above. A local mechanism can work while packaging, acceptance or support remains incomplete. Nothing in a local run proves public availability, a finished release or a support commitment.
Release wording must therefore come from a dated release owner, not from a green mechanism check.
Platform wording names the exercised path
The source declares selected coding-agent entry points and includes launcher, test and smoke evidence. That can support a sentence about the exact path and environment exercised.
It cannot support “works everywhere.” A launcher case is not the complete install-to-close journey. A local pass is not a remote pass. A declared target is not observed behavior, and a skipped case is not a successful case.
Activation is not migration
The project initializer can place a mechanical seed in the order-app repository and record what it owns. It does not interpret the repository's existing plans, instructions, decisions or generated views.
The pinned review contains no verified existing-corpus migration path. It therefore cannot support claims that existing knowledge will be reconciled correctly, moved without loss or fully restored by undo. The seed's reversibility applies only to the represented seed.
Unsupported inferences
The reviewed evidence does not establish that Coldstart:
- improves general productivity or code quality;
- eliminates drift, stale guidance, missed judgment or failed handoffs;
- ranks guidance better than simple word matching beyond the authored set;
- behaves alike across every named coding agent and shell journey;
- safely reconciles an existing project corpus;
- makes every mechanism active in every ordinary session; or
- turns a represented check into a guarantee about the surrounding work.
If a sentence depends on one of these conclusions, remove it or gather evidence that measures that exact question.
The cost of private evidence
The repository behind this handbook is private, so a reader receives a dated review boundary rather than direct source access. That limits independent verification. The honest response is narrower claims, visible dates and a clear route for questions—not a suggestion that private evidence is equivalent to public proof.
Recheck a claim when its mechanism, evidence set, platform path or release state changes—and when a neighboring page changes what the two sentences imply together.
Say the narrow predicate the evidence supports, and stop before the inference it does not.