# DSEWiki visual edition: figure notes

This edition presents the existing Fide assessment as a visual narrative. It is
not a new model experiment. Claim judgments are assistant assessments pending
independent human adjudication. The full limitations remain in `corpus-study.md`.

## Read the figures

1. **Shared wiki schematic.** A conceptual route through a shared website. The
   two agent symbols are not an agent count, and arrows do not establish delivery
   or use for every message. The underlying incident reconstruction is credited
   to the original researchers at https://collusion.wiki/.
2. **Eighteen seconds.** `post-deletion-revisions.csv` contains the page and saved
   revision displayed in `visual-story.json`. Subtract the first recorded deletion
   time from the save time. C263 supplies the quoted categorical denial. Other
   reports had identified post-deletion rewriting; Fide reproduced it. C263 has
   no follow-up in this comparison set.
3. **Archive versus live site.** A conceptual evidence distinction, not a measured
   access test. A retained archive item does not establish later live access.
4. **Original and follow-up.** C295 and C117 are a separate verified pair. The
   quoted size claim persists in the longer follow-up. Published finding-coverage
   scores are multiplied by 100 and displayed to one decimal place; they are not
   accuracy percentages. `corpus-record-checks.json`, query `welcome_sizes`, gives
   total characters over all versions and the largest single saved version.
   The figure labels characters explicitly; it does not equate characters with
   bytes. Claim judgments C295-07 and C117-10 and their transition are available
   in `corpus-claims.csv` and `claim-transitions.csv`.
   The follow-up also expanded its cleanup analysis. Retaining this size mistake
   does not mean it made no useful progress.
5. **78 follow-ups.** Keep only `selection=corrected_primary` in
   `corpus-paired-outcomes.csv`. A score increase means follow-up coverage exceeds
   parent coverage. Among those 61 pairs, 44 have `problem_retained_problem > 0`;
   34 have `strict_problem_retained_problem > 0`. These are counts of reports,
   not claims. Tiles are grouped by outcome, not chronological or model order.
   Each tile's report IDs and flags are in `visual-story.json`. Toggling removes
   disputed judgments, not reports or pairs, and does not independently validate
   the remaining judgments. Categories do not grade severity.
   The separate correction result changes denominators: 3 of 57 follow-ups with
   an earlier flag corrected or narrowed one; excluding disputed judgments
   leaves 1 of 44. Those 44 eligible follow-ups are not the same group as the 44
   higher-scoring follow-ups retaining a flag in the main chart.
6. **Evidence ladder.** This organizes observations and their interpretation; it
   is not a calibrated confidence scale. The 14 request/save matches are in
   `corpus-request-save-pairs.json`; Fide's two changed judgments and continuing
   disagreement are documented in `corpus-reconciliation.md`.

## Reproduce the visual derivations

Download and extract `visual-story-data.zip`, or save the following files in one
folder with `verify-visual-story.py`:

- `visual-story.json`
- `corpus-paired-outcomes.csv`
- `post-deletion-revisions.csv`
- `corpus-record-checks.json`
- `corpus-claims.csv`
- `claim-transitions.csv`
- `corpus-report-inventory.csv`

Run `python3 verify-visual-story.py`. Add `--check-sources` to retrieve the three
pinned original reports and verify source hashes, excerpt text and line numbers.
Python 3.9 or later is sufficient; no third-party packages or inference API keys
are needed. Keep assertions enabled (do not run Python with `-O`). This checks the
published derivation, not the correctness of the assistant's substantive judgments.

The figure package identifies the pinned upstream commit and input hashes.
`corpus-package-manifest.json` includes the visual data, these notes and the
verification script. The website's evidence validation checks these hashes.

## Editorial boundaries

The deletion example and the follow-up example involve different reports.
Decorative document lines are not facsimiles of surrounding text. Excerpts link
to complete pinned originals. No operational decision or harm caused by these
exercise reports was observed. The follow-up instruction asked for expansion
and verification, not a targeted attempt to repair earlier errors. Findings do
not establish a causal effect of additional time, an error rate for all AI
investigators, or the effectiveness of Fide's proposed verification intervention.

The claims explorer remains a repository application with local setup
instructions. The article hosts only its own bounded figures, not that app.