Documentation for a tool tends to describe the tool at its best. This page exists to make that harder. It lists what Coldstart claims, what it deliberately words carefully, and what it does not claim at all, so a reader can check any sentence on this site against a published policy rather than a tone.
Publishing the method without this page would mean teaching a discipline about keeping claims attached to evidence, while quietly exempting ourselves.
How to read a claim on this site
Every claim Coldstart makes falls into one of three groups.
- Supported. The repository contains the mechanism, and the claim can be stated plainly.
- Requires careful wording. Something real is there, but the obvious phrasing would overstate it. These claims have a required form and a forbidden form.
- Not supported. No evidence establishes it today. These are not published as claims anywhere on this site, in any phrasing.
A claim moving up a group is a change in evidence, not a change in confidence.
What Coldstart claims without qualification
- Coldstart is a local Claude Code harness and operating discipline.
- It uses a small resume pointer and routed reading lists for multi-session state.
- It separates project knowledge from reusable capability cards using two governed shelf shapes.
- It contains hooks, deterministic checks, a capability-card library, generated indexes, and an intrinsic evaluation harness.
- It records current open work in a bounded queue and uses generated maps for knowledge reachability.
Every item is a statement about what exists and what it is for. None of them is a statement about what it does for you, which is the next section's business.
What is worded carefully, and why
The left column is how this site says it. The right column is the phrasing that would be an overclaim. The difference is not politeness; each pair marks a place where the evidence stops.
| What we say | What we do not say |
|---|---|
| Designed to reduce drift | Eliminates drift |
| Can make mandatory steps less optional | The agent cannot bypass it |
| The repository checks this mechanism | This mechanism always works |
| Current intrinsic results | Benchmark-proven superiority |
| Mostly implemented before cutover | Released |
| Principles may transfer | Platform independent |
| Designed to improve coherence | Improves productivity |
| The current Claude Code implementation | Works with coding agents generally |
Two of these are worth expanding, because the careful wording is doing real work.
"Can make mandatory steps less optional." Coldstart's checks and hooks raise the cost of skipping a step. They are not a wall. The page on rules the agent must not forget describes a coverage gap found in this exact mechanism, and what it took to close the class of gaps around it.
"Current intrinsic results." The repository can show that a generated map matches its sources, that a detector goes red on a represented failure, and that authored queries reach the cards they should. Those are results about structure and reachability. They are not results about your work getting better.
What Coldstart does not claim
None of the following is supported by current evidence, and none of it appears as a claim anywhere on this site.
| Not claimed | Why not |
|---|---|
| Coldstart improves general productivity | No outcome study exists. This is the hardest boundary in the whole project and it has never moved. |
| Coldstart produces better code than a simpler setup | Same. No comparison against a simpler setup has been run. |
| The routing layer ranks better than a simple word-match baseline | Measured. On the authored test set the two tie on recall and ranking. A tie is not superiority. |
| The full harness is green on every platform it names | The hook test rigs fail on one platform for reasons that are about the rigs, not the product, and there is no continuous integration at all, so nothing is verified automatically on any platform. |
| Existing project documentation can be migrated safely | The filing rules work for material created inside the model. Moving an existing corpus in is unproven, the path is missing, and this is the hardest of the open problems. |
| Every subsystem is active in an ordinary session | Some capacity is implemented but dormant. Implemented and exercised are different claims. |
| Public cutover is complete, or the acceptance period is done | Neither is claimed. The verification date at the top of this page is what dates that statement. |
A reader who finds one of these claims made on this site has found a bug in the site, and it is worth reporting.
Where the evidence comes from
Coldstart's sources are private, and will stay private. The product ships as a build rather than as a repository, so nothing on this site links to a file, a line, or a commit that you could open and check for yourself. A public repository under this name exists and holds an earlier version; it is not the source these claims were verified against.
That is a real cost to you and it should be stated rather than disguised. What you get instead is a date: every page says when its claims were last verified against a pinned snapshot of the source. What you do not get is the ability to verify them independently. Nothing about that is going to change, so it is stated here once rather than implied by silence on every page.
Two conventions follow from that, and they apply to every page here.
- Claims are pinned to a reviewed state, never to memory. A number that was true at the last review is not re-quoted later as if it were current. It is re-derived or it is dropped.
- Mechanism and status are separate sentences. How something is designed to work changes slowly. Whether it is finished, accepted, or running on your platform changes fast. When a page mixes them into one sentence, that sentence goes stale invisibly.
When this page goes out of date
A website has no session boundary to force a re-check. It serves whatever it last said until someone changes it. So the claim policy names the conditions that invalidate it. Any of these means the claims above need re-deriving before they are trusted:
- The work on install and command surfaces is finished. (Closed since this policy was written.)
- The project's own knowledge base finishes moving into the model it documents.
- A supported path for adopting an existing project is decided.
- Known open defects are fixed, or reclassified as something other than defects.
- The public cutover starts or completes.
- The test sets behind the routing, memory, skill-firing, or fact-checking results change materially.
Several of these have already moved since the policy was written, which is the point: this page carries a verification date at the top, and the date is the claim.
How current this page is
To ask whether a claim on this page still holds, or to report one that does not, write to [email protected]. The date above is when the page was last checked against the product, not when it was written.