# The Cache Debt Score

> Score any cache, model store, transcript folder, or vector index from 0 to 15 across volume, staleness, ownership, recovery cost, and blast radius, then decide what to clean first.

Canonical page: https://www.cacheddebt.com/score/

Five questions, scored 0 to 3, decide what to review first. Size alone is a poor guide.

## How it works

Score each retained item on five dimensions. Add the scores for a total from 0 to 15. The total sets a review band. A Blast radius of 3 means act now, whatever the total.

| Total | Band | Action |
| --- | --- | --- |
| 0 to 4 | Log it | Record it in the inventory and check again next quarter. |
| 5 to 9 | Schedule it | Assign an owner and a cleanup date within 30 days. |
| 10 to 15 | Act now | Review this week. |

## The five dimensions

| Dimension | 0 | 1 | 2 | 3 |
| --- | --- | --- | --- | --- |
| **Volume**: How much space does it take? | Under 100 MB | 100 MB to 1 GB | 1 GB to 10 GB | Over 10 GB |
| **Staleness**: When did a current workflow last use it? | This week | This month | 30 to 90 days ago | Over 90 days, or no current consumer |
| **Ownership**: Who decides whether it stays? | Named owner with a written retention rule | Known owner, no written rule | Shared or unclear | Nobody |
| **Recovery cost**: What does deleting it by mistake cost? | Rebuilt by a documented command | Rebuilt with a large download | Rebuilt by hand | Cannot be rebuilt |
| **Blast radius**: What does it affect while it stays? | Disk only | Build or test speed | Output quality, such as retrieval or context | Security: credentials or personal data |

## Score an item

The HTML page has an interactive calculator. Agents: score each dimension 0 to 3 from the table above, sum them, and apply the bands and override rule, or read the same rubric as data from https://www.cacheddebt.com/api/score.json.

## Worked example

Measured on October 22, 2026, with the audit script, in the cloud development container used to build this site. One machine, not a benchmark. Run [the audit script](https://www.cacheddebt.com/checklist/#measure) to score your own.

| Item | Size | Volume | Staleness | Ownership | Recovery cost | Blast radius | Total | Band |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| **Playwright Chromium build 1194**: The project pins Playwright 1.61, which expects build 1228. No test run uses this copy. | 920 MB | 1 | 3 | 2 | 1 | 0 | 7 | Schedule it |
| **Claude Code transcripts**: Two sessions, stored in plaintext, including every file the agent read. | 39 MB | 0 | 1 | 1 | 3 | 3 | 8 | Act now |
| **uv cache**: This project has no Python dependencies. The cache came from the environment image. | 424 MB | 1 | 3 | 2 | 0 | 0 | 6 | Schedule it |
| **npm cache**: Used by every install of this project. Roughly the size of node\_modules. | 321 MB | 1 | 0 | 1 | 0 | 0 | 2 | Log it |

The largest item, Playwright at 920 MB, is not the most urgent. The 39 MB of agent transcripts lands in "Act now" because plaintext transcripts can hold credentials.

## Limits

-   The thresholds are this site's judgment, not an industry standard. Adjust them to your disk sizes and policies, and write down what you changed.
-   Staleness from file dates marks review candidates. A file read daily but never rewritten looks old.
-   The score ranks cleanup work. It does not prove anything is safe to delete. Check the consumer first.
