PromptKB

Portable reading artifact · all retained recordings

Give the reader the earlier answers, too.

A latest-only snapshot can hide earlier failures and policy changes. The all-recordings handoff includes every retained response and review context in one offline HTML file, without requiring the editor or a PromptKB account.

Read the 24-output exampleDownload offline HTMLEditable evidence project

This is the separate all-recordings export. The existing latest-only handoff remains available and keeps its narrower scope. Neither file is a signed execution log or a project backup. Project JSON is still the complete editable file; original runner worksheets retain execution settings, finish reasons and token receipts.

Make a handoff from your own work

  1. In the Test Bench, keep your response rounds and review earlier recordings directly instead of replacing them. Check source privacy, prompt changes and review context before sharing.
  2. Under whole-set results or repeated evidence, choose Download all-recordings handoff HTML. It includes every retained record, regardless of the current case/variant/filter selection. Export does not change your project or make a model request.
  3. Open the downloaded file in a browser, offline if desired. Use the coverage section, all-record index and group-member links to reach any earlier or latest response. Open exact prompts and archived review contexts where needed.
  4. Inspect the entire disclosure before sending the file. Old prompts, responses, notes and decision snapshots can retain private material after current fields have been edited. Draft omission is not redaction. There is no public upload or share link.
  5. Keep project JSON and original worksheets alongside it. Recipients cannot send assessments back through this read-only artifact. Before printing, expand the details you need; the export does not promise to print collapsed content.

What travels, and what does not?

Counts reuse the editor’s repeatability rules: nonblank output plus exact current prompt match for eligibility; current checks re-evaluate every output; manual review counts require matching current criteria/checks. No checks means unchecked, not passing. Exact model-label groups do not verify equal generation settings. Numbering reflects stored array order, not independently verified execution chronology.

The HTML contains no scripts, remote fonts, images, uploads or tracking. Links in the offline export are internal. It has a restrictive content-security policy, but the file itself is editable and unsigned. Generated HTML is capped at 20 MB; if it is too large, export refuses rather than silently dropping history. Preserve a full project backup before preparing a smaller separate working copy.

Reading exercise: one changed policy, no new generation

The example preserves the prior cycle’s actual browser-exported reassessment project unchanged: four fictional support cases, two variants and three local-model generations each. All 24 actual Qwen2.5-0.5B-Instruct Q4_K_M outputs are retained. They were generated earlier with llama.cpp; this export did not run a model. The same agent authored cases/expectations and reviewed the outputs. This is not independent assessment, a representative benchmark or human-comprehension evidence.

Originally, all 24 outputs were assessed under the same three criteria and rejected. The reviewer then tightened source fidelity to require usable typed values, not just substrings in text, and reassessed only the first earlier A response. Its old pass/fail/fail context is archived beside the new fail/fail/fail assessment. The other 23 saved reviews are now out of date under the changed criterion—not silently reassessed or erased.

  1. Find the policy change. Open case 1 / A / earlier recording 1. Compare its current review with archived review 1. The output still contains "amount": "GBP 12"; the changed judgment does not mean a new answer was generated.
  2. Check coverage, not just presence. The coverage section shows 24/24 eligible raw recordings. But case 1 / A has only one current review out of three under the stricter criteria; the other groups have none. Eligibility is not acceptance or completed review.
  3. Inspect a stable error. In case 1 / B, three recordings have one distinct response text. Each drops explicit GBP 12 to null amount/currency. Exact repetition is not correctness. Member links expose all three records, not a selected favorable answer.
  4. Locate execution limits. The handoff deliberately does not infer seed, finish reason or token use from a model label. The original exercise guide and three raw worksheets retain those separate receipts, including the one length-limited output. Recording/review dates in this project are artifact/browser dates, not each request’s completion time.

Reading conclusions: preserve all failures, retain the changed review policy, and keep the tiny tested configuration on hold for unattended extraction. No additional inference, prompt tuning, output repair or new quality result was needed to make this evidence shareable. The walkthrough’s conclusions are agent self-review, not a verified recipient judgment.

Downloads and provenance

Project source: notes/evidence/2026-11-04-history-review-live-reassessed.project.json, exported before later synthetic probes. Hashes identify stored bytes, not truthful execution or independent review. Original actual worksheets and model/server license/source references remain under the repeated-execution example. The fictional task inputs, prompts, expectations and agent-authored explanations may be reused/adapted, including commercially, without attribution or warranty; model/server binaries and weights are not redistributed here.

Latest-only handoff guide · Review earlier recordings · Privacy and storage