From individual judgments to a scoped recommendation
What did you choose—and what justified it?
A row of green checks does not say where a prompt is safe or useful to use. Keep your recommendation, rationale, limits and next evidence beside the exact project evidence you relied on.
Explicit judgment, not an automatic winner. Choose Prefer A, Prefer B, or No selection / more evidence. You can record a decision with missing or stale evidence, but its coverage is shown before confirmation. Matching a saved snapshot does not prove completeness, quality, approval, independent execution or authority to deploy.
A real use: historical meeting notes
The historical Rust meeting exercise contains twelve authored responses across six source excerpts. Case review found that an apparently valid task overstated a promise to try. The overall recommendation used to live only in the separate guide. This companion stores it with a retained evidence project; the original example files are not rewritten.
All responses and decisions here are agent-authored with shared expectations and self-review. This is an editorial recommendation, not new model execution, an independent evaluation or current Rust assignments. Historical source permission and attribution remain inside the snapshots.
Prefer B—for review-assisted reading, not automatic assignment
Prefer the effort/start-date clarification for this six-excerpt editorial reading task. R01's first pass passed all automatic checks but exaggerated a commitment to try into a promised investigation; the revision preserves effort-only wording. R05 keeps the planned start in task and due:null. The other excerpts distinguish suggestions, completed work and unaccepted invitations. Both authored sets pass structure on 6/6; first-pass expected checks pass 2/6 and self-review accepts 1/6, versus 6/6 for the revision. Those counts describe authored probes and edits, not independently sampled model performance or a causal prompt comparison.
Intended use and limits
Use as a review-assisted historical reading aid under this exercise's disclosed policy. Four historical rows across two excerpts, with four empty cases. Do not assign current work, infer rollout success, or claim Rust endorsement. Soft commitments may need a different policy in another team. Sources are exact licensed historical excerpts; responses, expected values, refinements and judgments share the same agent/context. Six paired cases do not make the evidence independent or representative. No production release, safety certification or model-superiority claim is approved by this recommendation.
Next evidence
Before using this workflow on current meetings, agree on whether effort-only commitments and requests belong in the register; test unseen complete notes with independent responses and a separate recipient review. Check that source context resolves pronouns without guessing responsibility or dates. Reconsider this recommendation if that review finds overstated commitments, misunderstood exclusions or unusable attribution overhead. Source/license metadata currently repeats in prompts; separate metadata remains an unimplemented product hypothesis.
Authored device date (unverified): 2026-11-03T22:42:00.000Z. 6 cases; automatic pairs 6/6; reviewed pairs 6/6. A: 6/6 current, 2/6 automatic passes, 1/6 accepted, 0 stale. B: 6/6 current, 6/6 automatic passes, 6/6 accepted, 0 stale.
Save, change, inspect
- Record and review responses as usual. Inspect exclusions and meaning, not only automatic checks.
- Open Save a project recommendation with its evidence. Supply a rationale and intended use/limits; optionally state the next evidence or conditions for reconsideration. These form fields are unsaved drafts until Save.
- Confirm the displayed coverage. The project stores a new decision with a snapshot. Earlier decisions are not edited or silently replaced.
- Change a source, prompt, check, criterion, latest output, review or recorded model label: the old decision becomes Evidence changed — historical decision only. Its rationale and evidence remain unchanged. Restoring the exact captured evidence can make it match again.
- Download a decision's standalone evidence project to inspect the source set as it was. Use the full project JSON to keep all decisions, drafts and trial history. Reports and HTML handoffs carry the decisions too.
Exactly what is captured?
Current project/case names, case order, source inputs, both variants, applied criteria and case checks, and latest recorded responses with their full saved review contexts. Each decision captures both sides even if it prefers only one. Model labels and device dates are user-supplied, not verified execution metadata.
Excluded: response/model drafts and earlier recordings in trial history. Those remain in the full working project, not this decision snapshot. For clean reopening, snapshot response/model editor fields are initialized from the retained latest records (or blank when no record exists), never from unrecorded drafts. Editing only those excluded fields does not change the evidence match. Current source inputs and prompts can still contain private edits; a snapshot is not automatically redacted. Unapplied criterion-editor text and unsubmitted decision-form drafts are not captured either.
Coverage is recomputed from saved evidence by the current checker, not stored as a historical score or software-version guarantee. This is exact normalized-data comparison, not a cryptographic signature, approval workflow or tamper-proof audit. Files can be edited after download.
Retained evidence also means retained private text
A deleted current case can remain in any earlier decision snapshot. All project JSON, full Markdown reports and HTML handoffs include all saved decisions and their evidence, even if the recommendation no longer matches current work. Inspect every decision before sharing. Deleting current cases is not redaction; explicitly delete a decision to remove its snapshot from this working copy. Downloads and recipients' copies are not erased.
Up to five decisions per project, with 3,000 characters each for rationale, scope and next evidence. The whole project still must fit 10 MB, and browser storage may fill sooner. Save refuses an oversized/full candidate instead of trimming old decisions. Keep downloaded backups. These are provisional product limits, not a reason to treat old decisions as disposable.
Compatibility and portable files
Saving the first project decision opts into project format 6. Existing projects without this feature remain format 4 or 5. Old clients reject format 6 rather than silently drop recommendations. Deleting all decisions does not downgrade the project. Evidence snapshots are ordinary format-5 projects without nested decisions, response/model drafts or earlier recordings; older format-5 clients can inspect those separate snapshots. Version-6 projects still use response worksheet version 2; decisions are not inserted into prompts or response worksheets.
- Full decision-bearing example (format 6)
- Exact decision evidence only (format 5)
- Full report with rationale and saved evidence JSON
- Single-file offline review handoff
- Hosted reading copy (same-origin stylesheet; offline file is separate)
No account, content upload, model call, approval identity or new interaction collection. Browser saving remains opt-in. A separate read-only baseline is not automatically included in a decision: decisions snapshot the current project only. If a recommendation depends on an external baseline, preserve that separate file and explain the limitation; do not claim the snapshot contains evidence it does not.