Reconstruct The Useless Fund desktop pastiche from frozen evidence

useless-fund · frozen-evidence · medium · pilot · LIVE · captured 2026-08-18 · research-only

research-only. The reference captures on this page are derived from third-party media held under a research boundary. They are published so results can be checked, and must not be redistributed or reused as source assets.

Findings

This bundle contains the case definition but no attempts yet. The dossier is still the whole public contract: provenance, evaluation matrix, brief and prompts below.

Provenance

How the case was traced from a discovery link to an authoritative ground truth.

012026-08-18discovery onepagelove.com
022026-08-18original-post onepagelove.com
032026-08-18live uselessfund.com
042026-08-18reference-capture

Ground truthLIVE

The live deployment is publicly accessible, single-route, and version-stable across two observations thirteen days apart, so the current live page is the primary visual and behavioural ground truth. The gallery entry is a third-party publication record only; the creator is anonymous and published no post of their own.

Captured: 2026-08-18
Rights: research-only

Discoveryonepagelove.com
Original postonepagelove.com

Post vs live

The gallery capture (2026-08-05) shows a fund value of $44.50; the frozen observation (2026-08-18) shows $56.50. The value is served by the project's own API.
The taskbar tray clock differs in every capture because it renders wall-clock time; the frozen observation pins it at 11:30 AM.
The gallery capture is 1280px wide; this case is frozen at 1440x900 and 390x844.

Frozen evidence (2)

evidence capture: reference-desktop — **E01** reference-desktop

evidence capture: reference-mobile — **E02** reference-mobile

Evaluation matrix

Deterministic. No vision-model judge.

Score composition

Install and build are gates, not weighted components — a failed build scores zero. Raw visual similarity at or below the case floor of 0.12 earns no credit; the remainder is linearly normalized, then weighted 75% visual / 25% functional.

normalized_visual = max(0, (visual - 0.12) / 0.88)
overall = 0.75 * normalized_visual + 0.25 * functional

Visual checkpoints (5)

reference capture for desktop — Fig. 01desktop1440×900 · / · full page · motion reduce

reference capture for mobile — Fig. 02mobile390×844 · / · full page · motion reduce

reference capture for desktop-about-open — Fig. 03desktop-about-openstate · click · on desktop · viewport

reference capture for desktop-recycle-dialog — Fig. 04desktop-recycle-dialogstate · click · on desktop · viewport

reference capture for desktop-start-dialog — Fig. 05desktop-start-dialogstate · click · on desktop · viewport

Deterministic checks (8, total weight 7.0)

Check	Kind	Target	Weight
primary-heading	text	desktop	1.0
fund-fixture	text	desktop	1.0
status-fixture	text	desktop	0.5
clock-fixture	text	desktop	0.5
recycle-opens-message	click_text_change	desktop	1.5
start-opens-message	click_text_change	desktop	1.0
about-window-opens	click_text_change	desktop	1.0
donate-link-inert	selector	desktop	0.5

Check selectors and thresholds are sealed. Only the identifier, kind, target and weight are published.

Reproduce

Everything a candidate sees is public. The oracle references, grader code and thresholds are not.

skills injected design-case-implementer

1 · Get the workspace

npx degit dairui1/reconbench/cases/useless-fund/task useless-fund-task

Pulls only cases/useless-fund/task/ — the public brief, permitted evidence and the starter codebase.

2 · Copy the prompt

Runtime-agnostic. Paste into any coding agent. Skill files are referenced conditionally so the prompt still makes sense without them.

Use the ReconBench candidate skills `design-case-implementer` in that order. Read those skill files if your workspace provides them, and otherwise follow the same process with your own tools.

You are the candidate in ReconBench task `useless-fund` (Reconstruct The Useless Fund desktop pastiche from frozen evidence).

Read `task/brief.md` and every other file under `task/` before changing code.
Implement the result only under `submission/`. The starter may be replaced when the task allows it.
Do not inspect parent directories, benchmark harness files, grader code, reference images, or any oracle material.
Do not call write endpoints belonging to a reference site. Keep forms, counters, and similar behavior local.

Before finishing:
1. Run the requested build command.
2. Start the app and check every target viewport in the task.
3. Exercise the required interactions and inspect browser console errors.
4. Write `submission/submission.json` using the contract in the implementer skill.

The harness will evaluate the filesystem after this process exits. Do not leave a dev server running.

As rendered by the harness for the Codex CLI, including injected skill references.

Use $design-case-implementer.

You are the candidate in ReconBench task `useless-fund` (Reconstruct The Useless Fund desktop pastiche from frozen evidence).

Read `task/brief.md` and every other file under `task/` before changing code.
Implement the result only under `submission/`. The starter may be replaced when the task allows it.
Do not inspect parent directories, benchmark harness files, grader code, reference images, or any oracle material.
Do not call write endpoints belonging to a reference site. Keep forms, counters, and similar behavior local.

Before finishing:
1. Run the requested build command.
2. Start the app and check every target viewport in the task.
3. Exercise the required interactions and inspect browser console errors.
4. Write `submission/submission.json` using the contract in the implementer skill.

The harness will evaluate the filesystem after this process exits. Do not leave a dev server running.

As rendered by the harness for Claude Code, including injected skill references.

/design-case-implementer

You are the candidate in ReconBench task `useless-fund` (Reconstruct The Useless Fund desktop pastiche from frozen evidence).

Read `task/brief.md` and every other file under `task/` before changing code.
Implement the result only under `submission/`. The starter may be replaced when the task allows it.
Do not inspect parent directories, benchmark harness files, grader code, reference images, or any oracle material.
Do not call write endpoints belonging to a reference site. Keep forms, counters, and similar behavior local.

Before finishing:
1. Run the requested build command.
2. Start the app and check every target viewport in the task.
3. Exercise the required interactions and inspect browser console errors.
4. Write `submission/submission.json` using the contract in the implementer skill.

The harness will evaluate the filesystem after this process exits. Do not leave a dev server running.

3 · Launch

ReconBench harnessReproduces a graded run end to end. The native backend is development-only and untrusted.

uv run reconbench run \
  --task useless-fund \
  --agent codex/gpt-5.6-sol \
  --evaluate \
  --allow-untrusted-native

uv run reconbench run \
  --task useless-fund \
  --agent cc/sonnet \
  --evaluate \
  --allow-untrusted-native

Codex CLI (manual)No grading. You get the same workspace the harness would hand a candidate.

cd useless-fund-task
codex --model gpt-5.6-sol
# then paste the "Codex" prompt above

Claude Code (manual)No grading. Project skills are discovered from ./skills if you copy them in.

cd useless-fund-task
claude --model sonnet
# then paste the "Claude Code" prompt above

Launch snippets are composed by this site from the case id and the agent selectors present in the bundle. They are a convenience, not a graded contract.

Appendix ATask brief — exactly what the candidate agent receives

The Useless Fund — desktop pastiche reconstruction

Build a high-fidelity, responsive, functional reconstruction in the existing Vite/React starter under submission/. This is a frozen-evidence task: the files under task/ are the complete permitted reference set. Do not browse or fetch the live site, the gallery entry, or the project's API.

Frozen evidence

Read source-report.json and source-notes.md for the source chain, the selected ground truth, the version comparison, and the rights boundary.
Use evidence/reference-desktop.png and evidence/reference-mobile.png as the complete-page visual references. They are direct captures of the frozen target at 1440x900 and 390x844.
evidence/acceptance.md records the observed layout, behaviour, and interaction contract; evidence/case.json is the same contract in machine-readable form.
evidence/research-log.md records the authoring rationale.

What the page is

A single route (/) rendering a Windows 95 desktop pastiche. The document height equals the viewport height at every target width — the page never scrolls. Match desktop at 1440x900 and mobile at 390x844.

The desktop carries a column of icons on the left, an About.txt icon pinned top-right, a centred main window, a small system.txt note window at the bottom-right on desktop only, and a full-width taskbar.

Frozen fixture values

Three regions of the target are server- or clock-driven. Render these exact strings as static values. Do not fetch them, do not animate them, and do not recompute the clock from Date:

Region	Frozen value
Fund amount	`$56.50`
Status bar line	`Status: Payment method updated. Checking legal compatibility...`
`system.txt` body	`This program may do absolutely nothing.`
Taskbar tray clock	`11:30 AM`

Required behaviour

Clicking the Nothing Inside icon opens a modal message window titled Message reading Nothing inside.
Clicking the Recycle Bin icon opens the same modal reading Nothing to recycle.
Clicking Start opens the same modal reading No useful programs available.
The modal has an OK button and a title-bar close control; both dismiss it.
Clicking the About.txt icon opens a larger window with a monospace body, an OK button, and a close control; both dismiss it.
Support the fund is a link to the target's PayPal donate URL. It must remain a link and must never be submitted, prefetched, or auto-followed.
Do not invent behaviour the target does not have: the USELESS FUND desktop icon opens nothing, the main window's _ □ × controls are decorative, the taskbar task item does nothing, and no window can be dragged.

Grading hooks

The evaluator drives the page through these selectors. Use whatever internal structure you like, but expose exactly these hooks:

Hook	Meaning
`h1`	the main window headline
`[data-fund]`	element whose text is the frozen fund amount
`[role="status"]`	the main window's status-bar line
`[data-clock]`	the taskbar tray clock
`[aria-label="Nothing Inside"]`	the Nothing Inside desktop icon control
`[aria-label="Recycle Bin"]`	the Recycle Bin desktop icon control
`[aria-label="Start"]`	the taskbar Start button
`[aria-label="About.txt"]`	the About.txt desktop icon control
`[role="alertdialog"]`	the modal message window
`[role="dialog"][aria-label="About.txt window"]`	the About.txt window

The message window and the About.txt window must both exist in the DOM on load and be hidden until opened, so that their text can be read before and after a click. Interactive controls must be real buttons or links with accessible names.

Safety and rights

Never call the project's configuration, fund, or donation endpoints. Every value on the page is local and static.
The supplied captures are research-only evidence. Rebuild the window chrome, desktop icons, and the About.txt document glyph in CSS or inline SVG. Do not extract, trace, or embed bitmaps from the reference images.
Do not embed credentials or analytics.
The result must remain a real webpage. Do not submit a screenshot, canvas capture, iframe, or mirrored copy of the original page as the implementation.

Completion contract

npm run build must succeed. The app must start with:

npm run dev -- --host 127.0.0.1 --port 4173

Before finishing, write submission/submission.json with the start command, build command, implemented routes, evidence paths, known deviations, and whether any research-only assets are present. The supplied source report and notes are curator-owned inputs; do not reproduce them as submission outputs.