percept-video 0.1.0

Percept — neuro-engagement scoring client for the TRIBE v2 endpoint. Score any video 0–100 against a brain-response model.

Cerebra Neuro-Eval — a Pika plugin

The measurement layer Pika is missing. Every Pika skill generates a video. This plugin tells you whether it will actually land — and then makes it better on its own.

Cerebra Neuro-Eval scores any video with Meta/FAIR's TRIBE v2 brain model (a population-average cortical-response predictor) and turns that into a decision-ready 0–100 neuro-engagement score, a per-second attention curve, four cognitive channel traces, and the single weakest beat to fix. A second skill wraps that score in an automatic generate → score → fix → re-score loop, so a Pika video improves itself with a neural reward signal and no human in the inner loop.

It's a drop-in: download the folder, point it at a TRIBE v2 endpoint, and the two skills appear to the Pika agent.

---

Why this exists (the gap)

Pika's skill library is entirely generative — ugc-ads, app-sizzle, founder-product-video, viral-hook, … all end at "here's your video." There is no skill that evaluates or optimizes a video (searching Pika's own skill registry for "score / evaluate / will this perform / brain response" returns nothing). So creators ship blind: no signal on what works, no loop to improve it.

This plugin closes that loop.

	Before	With Neuro-Eval
After generating	"Looks good?" 🤷	61/100, weak at 0:11–0:14, LANG dip
Picking a cut	gut feel	A/B scored, winner by margin
Improving	re-prompt and hope	auto loop, score climbs each round

---

What's in the box

pika-plugin/
├── skills/
│   ├── neuro-eval/SKILL.md       # score a video → engagement report
│   └── neuro-optimize/SKILL.md   # auto loop: generate → score → fix → re-score
├── server/
│   ├── engagement.py             # TRIBE v2 cortical tensor → 0–100 engagement
│   ├── server.py                 # FastAPI /score + /health
│   └── requirements.txt
├── colab/
│   ├── cerebra_tribev2_server.ipynb   # host TRIBE v2 on a free T4 + public tunnel
│   └── _build_notebook.py             # regenerates the notebook from server/
├── client/
│   └── cerebra_eval.py           # stdlib CLI: score / A-B a video locally
└── .env.example

---

Setup (≈5 min)

1. Host the brain model (Colab, free T4)

TRIBE v2 needs a GPU, so the weights live on Colab — not on the Pika host.

1. Open colab/cerebra_tribev2_server.ipynb in Google Colab.
2. Runtime → Change runtime type → T4 GPU.
3. Optional: add a Hugging Face read token to Colab Secrets as HF_TOKEN after accepting access to meta-llama/Llama-3.2-3B. Without one, the server uses TRIBE's ungated video pathway instead of failing. The notebook downloads and verifies V-JEPA2 before accepting requests, bypassing the hf-xet stall observed during real T4 testing.
4. Run all cells. The last cell prints:

   NEURO_API_URL = https://<random>.trycloudflare.com
   NEURO_API_KEY = <generated secret>
   

   Keep that cell running — it is your live endpoint.

2. Point the plugin at it

cp .env.example .env     # then paste in NEURO_API_URL + NEURO_API_KEY
export NEURO_API_URL=https://<random>.trycloudflare.com
export NEURO_API_KEY=<secret>
curl -s "$NEURO_API_URL/health"      # {"ready": true, "cuda": true, ...}

TRIBEV2_TEXT_MODE controls the gated language path:

• auto (default): use text when authorized; retry with the configured ungated modalities on failure.
• off: reliable ungated scoring for any user.
• required: require WhisperX + LLaMA and return a clear diagnostic if unavailable.

TRIBEV2_MODALITIES controls the ungated base path:

• video (Colab default): one public 4.14 GB V-JEPA2 encoder; most reliable on T4.
• video,audio: also download/use the 2.32 GB Wav2Vec-BERT encoder.

The notebook pins the tested TRIBE v2 Git revision, logs the API to /content/neuro-server.log, and exposes authenticated /diagnostics. A failed score prints the exact stage, hint, diagnostics, and log tail instead of only HTTPError: 500.

Debugged Colab failure

The original HTTP 500/hang had four independent causes:

1. Colab retained a Torchaudio build with a different Torch ABI.
2. dependency resolution replaced TRIBE's exact NumPy 2.2.6 pin.
3. /health reported ready before the lazily loaded V-JEPA2 encoder was present.
4. hf-xet transferred roughly 3 GB, then deadlocked after a CDN request failed, leaving /score blocked until its 30-minute client timeout.

The notebook now pins the compatible Torch/Torchaudio/NumPy stack, downloads V-JEPA2 before server startup via resumable curl, verifies its exact byte size and SHA-256, and defaults to video-only TRIBE inference so Wav2Vec-BERT is not a hidden 2.32 GB dependency. A real T4 run completed /score and a three-candidate plateau loop using one reused clip.

3. Install the skills into Pika

Drop skills/neuro-eval and skills/neuro-optimize into your Pika plugins source (the same place the official Pika-Labs/Pika-Plugins skills load from). The Pika agent will then discover them via search_skill ("score this video", "make this video better", …).

---

Use it

Score (quick check, no Pika needed):

python client/cerebra_eval.py myclip.mp4
python client/cerebra_eval.py cutA.mp4 cutB.mp4    # A/B → names a winner

In Pika (natural language):

• "Score this video" → runs neuro-eval, returns the report.
• "Make this ad more engaging" → runs neuro-optimize, iterates until the score plateaus, returns the winning clip + the score history.

---

The score, briefly

TRIBE v2 predicts average cortical response on the fsaverage5 surface. We read four engagement families off the Glasser/HCP-MMP atlas — auditory, language, attention, visual — and measure their activation relative to the full cortex at each timestep. The canonical adScore is a transparent weighted sum: 55% TRIBE cortical features, 30% OpenCV production features, and 15% YOLO semantic/composition features. Every response includes raw values, normalized subscores, weights, and contributions. Every upload is normalized to at most three seconds. The weakest weighted features and explicit generator instruction become the reward feedback for a maximum-five-iteration Pika optimization loop.

It is a prediction of population-average response, not a measurement of any individual viewer and not a sales guarantee — a strong, fast, directional signal for the inner loop, with an optional human panel to confirm the final pick.

Provenance

The scoring math is ported from Cerebra's production TRIBE v2 worker; this plugin packages it as a standalone, Pika-installable evaluation + optimization layer. Model: facebook/tribev2.

License note: the plugin code and bundled YOLO/OpenCV path are separate from the TRIBE v2 weights. Meta publishes TRIBE v2 under CC-BY-NC-4.0. Commercial ad-scoring use requires permission compatible with that license.