Every version, and what it changed for you.

Your adoption figure was counting your own edits against Archis. It counts reads now.

Before you upgrade — Nothing to change, and no reindex. The adoption figure on your dashboard will move — usually upward, and further the more you edit — because it is dividing a different and more honest denominator, not because anything about your usage changed.

• NewHow many of your sessions never asked Archis anything. A percentage cannot tell thin-but-even use from heavy use in a few sessions and none in the rest, and the two want opposite fixes — the second means your agent does not know Archis is there until something reminds it, which is what chi agents advice installs. The repository page now says which of the two patterns it is looking at.
• NewWhat one look at your code actually costs, per level of detail. The context bars were drawn against the largest level, so the biggest was always full-width and the one fact worth reading was unrecoverable: a level can be a small share of the lookups and half the tokens. They divide the total now, and each names its average cost per lookup — which is where you can see that reading a file whole costs an order of magnitude more than reading its shape, on your own repository rather than in a benchmark.
• FixedThe dashboard's headline figure — the share of file access going through Archis — divided every recorded access, including every file your agent edited. Archis is what you read your repository with; your harness is what writes to it, so every edit landed in the denominator of a figure meant to answer whether the index is being consulted. On a repository under active development the effect was not marginal: what 0.25.0 reported as 19% is 48% on 0.26.0, from the same stored history and without a reindex. The figure now divides reads, and the edits are shown beside it rather than folded into it, because how much work happened here is a real question — just not that one. The same correction lands in chi doctor, chi dashboard --plain, the terminal view, chi agents and the --json output, so no two of them disagree.
• Fixed"Context avoided" no longer reports a confident 0% where it had measured nothing. The figure compares what Archis kept out of your model's context against what it showed, and it read 0% whenever that particular mechanism had not been used in your sessions — which looks exactly like "Archis saved you nothing" beside a hundred thousand tokens it had served. It reads as not measured now, with the reason, and the saving that *is* measured takes its place in the summary row: how much of what Archis served you was exact source, which is the part bought at full price because the bytes are the bytes.
• FixedA "Model" column whose every row read unknown, and session counts that stopped at five hundred. The first presented a missing reading as a finding — your harness does not tell Archis which model it is running when it reports a read — and it is left blank now. The second silently reported the cap instead of the count on any repository with more history than that, on the dashboard, in the terminal view and in --json.