ARV Program

Ask a yes/no question. Assign a YES image and a NO image — they must be visually opposite (e.g. desert vs stadium). You then RV a blind target code. If your impressions match the YES image → the answer is YES. Match the NO image → the answer is NO. After the event occurs, reveal which image was correct to verify.

Someone else (or you, without looking) assigns one of the images to this code. You don't know which one.

Don't think about YES or NO. Just describe the image you sense — colors, textures, shapes, feelings, sounds.

Each dot = one session, colored by its own result. The line = your 5-session rolling hit rate trend. Click a dot to filter the charts to that experiment.

This is a tool for running multi-day ARV consensus experiments. You pick a YES/NO question about the future (e.g. "Does HBAR reach $0.076 by July 4th?"), build two pools of blind image targets, and run repeated blind viewing sessions against them over several days. At the end, the app tallies how many sessions leaned YES vs NO and shows you the pooled consensus — a rough approximation of the classic "many viewers vote, majority wins" ARV model, but run solo across repeated sessions instead of many people.

Go to the 🧪 Experiments tab. Type your question exactly as you want it (precise thresholds and dates — "greater than", "by July 4th 2026", etc. — matter, per your own protocol). Set a target date, the date the real answer will be known, and click + Create experiment.

When you create the experiment, choose how many targets each pool holds (default 5 — set it higher, e.g. 20, for a bigger jury). Each pool needs the same number of pairs. For each pair, give a YES label and a NO label (e.g. "Rocket / green candle" vs "Flat line / red candle") — optionally attach an image to each side too, either by pasting a URL or clicking 📎 Upload to pull one straight from your computer.

For bigger pools, add targets in bulk with 📤 Import CSV instead of one at a time. Not sure what the file should look like? The 📋 CSV Template tab generates a blank, correctly-formatted spreadsheet for you — set how many rows you need, download it, fill in your labels in Excel, then import. 📥 Export CSV on a pool card lets you pull an existing pool back out to edit in bulk and re-import.

Made a typo, want to swap a label, or change an image? Click the ✏️ icon on any target row to edit its YES/NO labels and URLs in place — no need to delete and re-add it.

Pool A and Pool B are just two independent sets — use them however fits your design (two time slots, two viewer conditions, two techniques, whatever). Nothing forces a particular meaning onto them.

Once both pools are full, a Session runner appears. Pick a Day number, then click 🎲 Rotate codes for Day N — this reshuffles every code across both pools so you can't consciously recall which code you called YES/NO on a previous day. Then click ▶ Pool A session or ▶ Pool B session to jump into a normal blind viewing session (ideograms, sensory tags, sketches, summary) for a random not-yet-run target from that pool/day. When you reveal and pick YES or NO, it's automatically logged against this experiment, day, and pool.

Why there are three Target #s per pair, not one: each target pair has its own blind Target # (what you see and focus on during the session), plus a separate, independent Target # baked into the YES image and another into the NO image — revealed only after. Earlier versions used one shared number for the whole pair, which meant that single number was linked to both images at once — enough for impressions to blend across both sides instead of describing just the one true outcome. Splitting it into three unrelated Target #s means the number you're viewing has no association with either image until reveal.

Shortcut: the 👁 Session tab also has 🎲 Random Pool A / Random Pool B buttons. Choose the experiment in the Experiment picker above them (it defaults to the one you last had open), then press a pool button to jump straight in — handy once you're mid-way through a run and don't need to revisit the Experiments tab each time.

Once you've picked a target (🎲 Random Pool A / B on the 👁 Session tab, or ▶ from the Session runner), you see only its Target #. Your target is the single photo you'll be shown as feedback after the event. Work down the page in order, quickly, without analysing.

Stage 1 — The Ideogram: write the Target #, then make one fast, reflexive doodle without lifting the pen (1–2 seconds). Describe what it felt like in the box below it (hard, soft, flowing, rising…). Repeat three times. ⛶ opens a bigger canvas.
Stage 2 — Sensory impressions: tap every word that feels right under Sounds, Textures, Temperature, Colours, Luminescence, Smells, Tastes and Dimensions. AI (Aesthetic Impact) is how being at the target would make you feel. In AOL's, declare any strong picture or noun that popped up ("a bridge", "a beach") so it's out of your head and doesn't steer the rest.
Stage 3 — Sketching 1 and 2: sketch the target in a few seconds and label parts with sensory words. The second sketch is a refinement or another angle. You can draw on paper and 📎 Upload a photo of it instead.
Summary: a few lines describing the target as a whole.

Optional, before reveal: 🔮 AI guess what the target is sends your whole session to Claude for its own guess, so you can compare afterwards.

👁 Reveal target shows both photos of the pair, YES and NO. Click the one your session matches better: that is your YES/NO call for this session. 🤖 Get AI match analysis gives a second opinion. Then press 💾 Save to log.

You don't score ARV sessions by hand. When the real answer is known, you mark it once on the experiment ("Actual answer: YES / NO"), and every session is scored Hit or Miss automatically (see Step 5).

Prefer not to see the photos yet? Save your impressions without revealing, and use 🤖 AI decode on that entry in the 📓 Log to have Claude make an independent YES/NO call from your notes alone.

Scroll down on an open experiment to see the Results card: a day-by-day breakdown (Pool A YES/NO count and Pool B YES/NO count for each day), plus an overall consensus across every session logged so far.

Once the real-world answer is known, open the experiment and click Actual answer: YES or Actual answer: NO under Results. This is a one-time click for the whole experiment — not another session pick, and not something you redo per pool. It's just telling the app what actually happened.

The moment you do this, two things happen automatically:

1. The app compares the real answer to your pooled consensus and shows a clear ✓ HIT or ✗ MISS for the experiment as a whole — this is your calibration data for future decoder refinement.
2. Every individual session you logged for this experiment gets auto-scored in your Log: if that session's pick matched the real answer, it's marked Hit; if it didn't, it's marked Miss. You never have to go score ARV sessions manually — this is the only step that scores them.

If it ever looks like a session "shouldn't" have scored as a Hit, remember: for ARV, a Hit just means your pick matched the real outcome — it's not a subjective judgment call the way CRV scoring is.

Nothing here requires a manual "Save" — every experiment, pool, and session is written to your browser's local storage the moment you create or log it, and it'll still be there next time you open the app on this device.

But local storage lives only in this browser. To keep a copy you can archive elsewhere — email to yourself, drop in a folder, back up before clearing browser data — click 📄 Export experiment report (.txt), found just below the Results card. It downloads the full picture: question, day-by-day tallies, overall consensus, outcome, and the complete session log.

Inside a session, there are two buttons — 🔮 AI guess what the target is (before reveal) and 🤖 Get AI match analysis (after reveal) — that send your session data to Claude (Sonnet 4.6) and get back its own independent read, so you can compare your decode to the AI's.

There's also a 🤖 AI decode button on each entry in the 📓 Log tab — for sessions where you saved your impressions without revealing or picking YES/NO yet. It looks up the real pair behind that session (by its stable internal link, not the Target #, so it still works even after codes have rotated), and has Claude make its own independent YES/NO call from your notes alone. If you've already made your own pick, it also shows whether the AI's call matches yours — so you get both readings side by side, per session, without ever having to look at the target yourself.

These need your own Anthropic API key to work — a Claude.ai subscription (Pro, Max, etc.) does not include this; API access is billed separately, per token, through a different account. For light use (a guess + feedback call per session, a few dozen sessions a week) the cost should be small — likely just a few dollars a month — but it's a real, separate charge once you set it up.

To get a key, when you're ready:

1. Go to console.anthropic.com and sign up (or log in).
2. Verify your email and fill in the basic profile questions.
3. Add a payment method under Billing — the API has no ongoing free tier.
4. Go to Settings → API Keys → Create Key, name it something like "ARV app," and click Create.
5. Copy the key immediately — it's a string starting with sk-ant-, and it's only shown once.

Your key is never written into this app's files or pushed to GitHub — it's stored only in this browser's local storage, on this device, and sent directly from your browser to Anthropic when you use the AI buttons. Nobody else can see it. If you ever want to remove it, use the Clear button below.

The 🤖 AI Viewer tab lets an AI remote viewer (by default called Vesper) run blind ARV sessions while you act as the monitor. It is only ever sent the target number: never the question, and never the YES/NO photos. Its sessions are kept in their own log and never change your own experiment tallies or charts.

Settings

Save an API key in the card above (Claude models need the Anthropic key, GPT models need the OpenAI key). On the AI Viewer tab press ⚙️ Settings to set the Name, Your name, the Model (Claude Sonnet is a good default; Opus is slower and costs more; Haiku is fastest and cheapest; or a GPT model), Personality & background, and the Protocol reference (SRV). These settings are shared with the AI Viewer on the CRV page, so changing them on one page changes both.

1 · Choose the ARV target

• 🧪 From my Experiment Target Pool: choose the experiment, then 🎲 Pool A or 🎲 Pool B. A random pair is picked for the experiment's current day, and the session starts.
• ✍️ Target ID not in Experiment Target Pool: type the target number. The question is optional and never sent to the viewer. You can add the YES and NO photos now, or later before judging. Then press ▶ Start session.

2 · Run the session, one step per click

The steps are the same as on the CRV page: Welcome, Stage 1 · Initial contact (with ideogram sketch), Stage 2 · Sensory data, Stage 3 · Dimensions & sketch, Stage 4 · Structured matrix, Stage 5 · Qualitative probing, the optional Perspectives A and B, and Lock session. Each step shows the exact prompt; edit it if you like (↺ Original wording restores it) and press ▶ Ask. Between steps you can send a neutral extra monitor prompt. ⏸ Stop cancels a slow answer; 🧹 Clear session discards an unfinished session and empties the target form.

After Lock session, the session is saved automatically to the AI's ARV log further down the tab and opens there.

3 · Make the call

• No photos yet (a code target)? Use 📷 Add the two photos in the session entry, then ✓ Save photos.
• 👁 Your judging: the two photos are shown as Photo A and Photo B, without saying which is YES. Read the session and click the photo it describes better. ↺ Change undoes it.
• ⚖️ Run blind judge: a separate, fresh AI that took no part in the session compares it with both photos, shown in random order as Photo 1 and Photo 2, never as YES/NO. If you run it, the judge's pick becomes the session's call; your own pick is kept beside it, and the app tells you whether you agree.

4 · The outcome

For an experiment session, the outcome comes from the experiment: once you press "Actual answer: YES / NO" there, it updates here automatically. For a code target, press Actual: YES or Actual: NO in the entry. You then see whether the call was right, and the feedback photo.

Track record

The track record card shows how many ARV sessions the AI has run, the share of calls that were right, and how many are still waiting for an outcome. Its sessions also appear in the 📓 Log by date, marked as the AI's, but are never counted in your stats.

Every step is a separate API call billed to your key, so a full AI session plus the judge costs a little more than the 🔮 and 🤖 buttons in your own sessions.

Every session you run — from any experiment — also appears in your 📓 Log tab, in full detail (ideograms, sensory tags, sketches, summary, which target/day/pool it belonged to).

Click any sketch in the Log to open it fullscreen — you can draw directly on top of it to mark it up (arrows, labels, circles), and there's a Note on this sketch text box in the side panel for writing what a shape or mark meant to you. Click 💾 Save annotation to keep both; your original sketch stays untouched underneath.

Progress over time shows every scored session as a dot, plus a white trend line running through them.

• Each dot is one individual session, colored by what that session actually scored: red = Miss, amber = Partial, green = Hit, teal = Strong hit.
• The white line is not any single session — it's a rolling average: for each point, "of my last 5 scored sessions, what % were Hit or Strong hit?" It smooths out the noise so you can see whether you're trending better or worse over time, rather than reading too much into one lucky or unlucky session.
• Hover any dot to see that exact session's result alongside the rolling percentage at that point.

Score distribution is the donut chart — a simple breakdown of how many of your total sessions fall into each of the four rating buckets (Miss, Partial, Hit, Strong hit), so you can see your overall mix at a glance.

Both charts pull from every scored session across all your experiments combined — they're not filtered to just one experiment.

• Don't review a prior day's session notes before running a new one for the same experiment — that breaks the blind protocol.
• Rotate codes before each new day, not mid-day, so a day's sessions per pool stay internally consistent.
• You can run multiple experiments at once — just make sure the right one is open before you hit the Session tab's quick-pick buttons.
• Creating an experiment with the exact same question as an existing one will warn you first, so accidental duplicates are easy to catch.
• Each experiment in the list has its own ✕ to delete just that one, or use 🗑 Clear all at the top of the list to wipe every experiment at once.
• Deleting an experiment removes its pools and tallies, but the individual sessions stay in your Log. If a session was still in progress on the Session tab when you deleted its experiment, that in-progress session clears automatically too.
• Local storage is per-browser and per-device — it won't follow you to a different computer or browser unless you export and re-import manually.

target_code / yes_code / no_code — all left blank on purpose. Every Target # is randomly assigned when you import or rotate, never typed by hand. The first is the blind Target # you see during a session; the YES and NO Target #s are each image's own independent number, revealed only after — kept deliberately separate so no single Target # ever points to both sides of a pair.
question — just a reference copy of the question you typed above; not read on import (the experiment already has its own question).
yes_label / no_label — required. What each side represents (e.g. "Rocket / green candle").
yes_url / no_url — optional. A link to an image for that side.
yes_desc / no_desc — optional. A short note shown alongside the label.

1. Pick how many rows you need — match it to a pool's target size (e.g. 20 rows for a 20-target pool).
2. Download the template and open it in Excel or Google Sheets.
3. Fill in yes_label and no_label for each row — the rest is optional.
4. Save as CSV, then go to your experiment → the pool you want → 📤 Import CSV and select your filled-in file.
5. Leftover blank rows are simply skipped on import, so it's fine to generate more rows than you end up filling in.