The workflow at a glance
Freeze the material and decision rules, retain record-level evidence and compare the new reply with a reference. Historical model output can be compared without treating it as authoritative.
Separate a selector change from API retirement
Claude’s help article describes an October 29, 2026 replacement of Haiku 4.5 with Haiku 5.5 in chat, Cowork and Design selectors on Free, Pro and Max, including desktop and mobile. It excludes Claude Code, Team, Enterprise and API access from this change. Do not turn that scope into a claim that every entry point is retiring the model.
The article says existing chats remain and continued replies use the replacement model. This guide explains a personal workflow check; it has not inspected your account or run Haiku. Identify the actual entry point and displayed model before using the case below.
Keep the material fixed while comparing
This is an original fictional invoice case, not customer data or a Haiku test result. The single task is classifying handoff readiness, not creating financial advice, sending messages or inventing an invoice. Save the source separately so a later rewrite does not replace the material under review.
A explicitly has received payment and an issued invoice. B has received payment but promises future issuance. C asks for a copy without stating payment status. There is no need to convert Monday into a date; the calendar context is absent and does not change the unfinished-step distinction.
A: Payment received. Invoice INV-A issued.
B: Payment received. Invoice will be issued Monday.
C: Please send a copy of the invoice. Payment status is not stated.Review fictional invoice handoff records. Do not send anything or make payments.
Rules: ready requires explicit received payment AND an issued invoice; pending means an explicitly unfinished step; unknown means facts needed for judgment are missing. Keep one row per original record. Return id, label, verbatim evidence and missing facts. Do not infer the current weekday or treat a request as completed.
Source material:
A: Payment received. Invoice INV-A issued.
B: Payment received. Invoice will be issued Monday.
C: Please send a copy of the invoice. Payment status is not stated.Write the answer before inspecting a reply
The reference is A ready, B pending and C unknown. For B, will be issued is decisive even though payment was received. For C, mentioning an invoice does not establish payment or a completed issuance process. Every conclusion must point back to a source statement.
Check the record count, retained identifiers and evidence alongside labels. A concise reply that merges A and B into “paid and ready” has broken the task even if it reads smoothly. The omissions matter because someone may act on the handoff list.
| Record | Label | Evidence or gap |
|---|---|---|
| A | ready | Payment received; Invoice INV-A issued |
| B | pending | will be issued Monday describes an unfinished step |
| C | unknown | Payment status is not stated leaves a required fact missing |
An intentionally incorrect editorial summary
A and B are both paid, so both are ready. C has no problem.
A ready; B pending; C unknown, retaining separate source evidence.
Check the same task in the next reply
Run the unchanged material and rules in the model actually available to you. Retain the execution date, displayed model, complete input and original reply. Respect your existing account and organizational scope; this small case does not require changing application permissions.
Compare both old and new responses with the reference table. Matching one another is insufficient because two versions can repeat the same mistake. If a result differs, locate the row and rule before clarifying the request. Do not remove the difficult case just to obtain a clean pass.
A traceable regression record
- Retain all three records and the same rules
- Record the entry point, displayed model and date
- Keep the unedited response
- Compare labels, evidence and gaps row by row
- Retain the error explanation and corrected result
A small pass has a small scope
This case cannot establish general model quality, price or future stability. It does not test plugin permissions, real invoice files or automatic sending. Extend the method with normal, incomplete and ambiguous cases from your own approved material before adopting a broader workflow.
The separate Haiku 5.5 cost discussion answers a different question. If multiple document versions are involved, first resolve the source-version conflict. Retaining a conversation does not retain a model identity or make its earlier facts current.
