Carry evidence, not assumed approval

Keep the recording with your receipt ledger.

Bring a batch of recorded answers into the receipt editor without copying each source and response by hand. Keep the exact original context beside later edits—and make a separate decision before an expense reaches the CSV.

This follow-up uses the unchanged fictional, agent-authored project from the earlier receipt exercise. It does not create new model calls, reviewer judgments or real expenses. Original recordings, device dates, reviews and the earlier published files are preserved. Attachments and hashes are unsigned; they do not authenticate a model, author, receipt or payment.

1. Preview a variant, then choose the rows

Download your current Test Bench project. In the receipt editor, open Bring recorded answers from Test Bench, choose that local file, select A or B, and preview.
Inspect the available current nonblank latest recordings and all omission reasons. Missing, blank, stale or unrenderable-prompt cases do not become rows. Rejected, unreviewed, mechanically invalid and clarification answers remain visible; they are not silently filtered into a success count.
Select the recordings you want. Selection starts empty. Append keeps existing rows and approvals; replace discards the old batch after confirmation. Both require the entire candidate to fit50 rows and3MB without trimming. An untouched empty starter row is replaced on append.
Every imported row starts without expense approval, even if its Test Bench review says accept. Preview the receipt against its source and approve only an actual supported paid purchase.

In the browser exercise, a manual approved row survived canceled intake and confirmed append. Replacing with six B recordings then cleared all old approvals. Four explicit receipt approvals reproduced the exact earlier CSV: EUR14.40, JPY1250, KWD4.125 and USD10.00 separately. Two correctly accepted clarification responses stayed excluded. A separate A preview retained six rejected recordings, including invalid receipt shapes.

Original reviewed project · Six imported rows with origin attachments · Actual four-row CSV

2. Inspect what changed without losing the original

Each imported row has an Attached Test Bench recording panel: project/case/variant labels, original source, exact recorded prompt/answer, model/authorship and device-date labels, recording-time checks, current checks/criteria and latest saved review context. Drafts, earlier recordings, archived reviews and saved decisions are omitted. Keep the original project for that wider history. The panel can download the original recorded prompt and answer even after ledger edits; the ordinary extraction-prompt controls instead use the ledger's fixed template and current source, which may differ.

Changing a ledger label or note does not change whether source and answer match. Editing either source or answer—including whitespace—marks that field different and leaves the original attached. The first row of the edited work example has an additional explicitly fictional transcription note and a trailing answer newline. Reopening preserves those edits and the unchanged origin, but no approvals.

This is exact equality, not a permanent audit or verified lineage: restoring the original text matches again. An externally edited file can forge an internally valid origin. Use Detach recording origin to remove old/private evidence from the working row while retaining current fields and notes; that also clears its approval. Saved/shared files are unchanged. Removing a row or clearing the batch also removes its attachments from this tab.

3. Choose spreadsheet types deliberately

The prior exercise stopped at Python CSV and SQLite TEXT parsing. This follow-up actually imported that same browser CSV into LibreOffice Calc26.2.6.3 on Linux, headlessly, using two fresh isolated profiles with explicit UTF-8/comma/quote filters. Network access was disabled for the conversion processes. The resulting ODS files were inspected for stored cell types/text, not through an interactive graphical session.

All five columns as text: all25 header/data cells matched the CSV exactly, including ISO date strings, EUR 14.40, KWD 4.125, USD 10.00 and the merchant's literal leading apostrophe. No formula cells. Text amounts preserve representation but require deliberate numeric conversion before spreadsheet arithmetic.
Only total as a standard numeric column: the four amount cells became numbers. Stored cell text changed 14.40 → 14.4 and 10.00 → 10; 1250 and 4.125 retained their text. Other columns stayed text and the merchant apostrophe remained literal. No formula cells. Do not sum different currencies together.

Actual text-column Calc workbook · Actual numeric-total Calc workbook · Settings, inspected cells, hashes and execution limits

Reproduce the explicit text-column import locally

With LibreOffice installed, place the downloaded CSV in your own working directory, create an output directory, and choose a new absolute local profile path. Replace the placeholder path below. This command is a standalone local conversion, not something the website runs for you. Inspect any file before opening it in an office application.

libreoffice -env:UserInstallation=file:///ABSOLUTE/NEW/PROFILE \
  --headless \
  --infilter='Text - txt - csv (StarCalc):44,34,76,1,1/2/2/2/3/2/4/2/5/2,1033,true,false,true,false,false,0,false,false,false' \
  --convert-to ods --outdir output expenses.csv

The explicit filter uses UTF-8, comma separation, double-quote delimiters, all five columns as Text, quoted fields as text, no special/scientific-number detection, no space trimming and no formula evaluation. Language identifier1033 is explicit. In the graphical text-import dialog choose corresponding options and Text for every column, but that graphical flow was not tested here. LibreOffice26.2 filter reference.

The comparison filter keeps columns1–4 as Text and uses Standard for total, with quoted-fields-as-text disabled: Text - txt - csv (StarCalc):44,34,76,1,1/2/2/2/3/2/4/2/5/1,1033,false,false,true,false,false,0,false,false,false. Prefer exact text preservation unless numeric interpretation is an intentional downstream choice.

This establishes behavior only for the installed version and these settings—not default CSV opening, Excel, Google Sheets, arbitrary formulas, every locale, visual formatting, tax treatment or a bank/accounting import. The apostrophe may be visible. Never infer payment or authentic source provenance from an ODS file or a matching quote.

Portable work format and privacy

Manual-only work still saves as version1. Any attached origin opts into promptkb-receipt-work version2. The newer editor reads both; older editors reject version2 rather than silently discard context. Opening requires fresh receipt previews/approvals. No account, content upload, model call or browser storage is introduced.

Version2 field contract

The envelope has format, version, rows. Each row retains version1's label, source, response, note and optional origin. Origin has exactly format: "promptkb-recording-origin", version: 1, projectName, caseNumber, caseName, variant, variantName, input, criteria, rules, record. Record has the original latest prompt, output, model, note, verdict, recordedAt, rules, review, without reviewHistory. Its rules/date may be null when genuinely unavailable. Current checks/criteria at intake are separate from recording-time and review-time snapshots.

Source and recorded output must copy without normalization; incompatible selected records refuse with instructions to open/export a compatible copy in Test Bench. Intake supports exported Test Bench versions4–6. Imported project metadata follows Test Bench validation. Origin validation requires its exact supported fields and text; it is consistency checking, not signature verification. Extra top-level/row work fields are discarded, including any manually supplied approval; an origin in version1 refuses rather than silently promoting it.

Existing limits remain50 rows and3MB UTF-8 for the full work file, including retained originals. Row limits remain label120 characters/single line, source20,000, raw answer30,000, note4,000; LF line endings. Edited or detached rows are not mapped back into the Test Bench project automatically. There is no fuzzy matching, transaction ledger, source verification or model execution receipt.

Work downloads include retained original private text even when current fields were edited. Review attachments before sharing. They are unredacted, unencrypted and unsigned. Artifact sources and SHA-256 hashes identify these fixed browser and Calc outputs, not authentic authorship or payment. Fictional examples may be reused commercially without attribution or warranty; no rights in receipts you supply are granted.