Lab Notebook
Idea Space
A running list of things I'm mulling, prototyping, or chasing. Some connect to SCBE-AETHERMOORE, some don't. Some will ship, some will archive, some will surprise me.
Nothing here is a promise or a roadmap. It's a notebook. If one of these catches your eye and you want to collaborate, email me.
Status tags:
Concept just an idea
Sketching thinking through it
Prototype building it
Shipping in the product
Archived paused or dead
Kinetic Attestation Primitive
Sketching
Anti-deepfake
A recording device that signs physical effort into the file so the recording is unforgeable by AI after the fact.
The insight
Every AI deepfake needs one thing it cannot synthesize: the correlated physical signal of a real human expending real calories at a real moment. A hand-crank voice recorder that captures audio AND the kinetic waveform of the crank, then signs them together with a timestamp, produces a file that says more than "here is audio." It says a human was present, moving, at this time, producing this sound. That's a claim you cannot retroactively synthesize no matter how good the voice models get.
Why this matters in 2026
Voice cloning is trivial now. Legal, journalism, evidence, and family archival are all trying to figure out how to trust audio recordings. The existing answer is watermarking (easily removed) or cryptographic timestamps (don't prove a human was there). A kinetic signature proves presence and effort in a way that scales with the physical world, not the compute budget of the forger.
What v1 looks like
- Hardware: ESP32-S3 dev board + digital mic + hand-crank generator + supercap + FRAM + small speaker. ~$80-130 BOM.
- Firmware: State machine that captures audio and crank voltage waveform simultaneously, stores both in a single sealed envelope on microSD, commits metadata to FRAM on brownout.
- Signing: HMAC or PQC signature over
audio_bytes || crank_waveform || timestamp. Verification is deterministic and offline.
- Long-form mode: Record indefinitely by cranking while recording — segment files get stitched together, brownout is non-fatal.
Where it fits in SCBE
The GeoSeal CLI already has the envelope primitive. A kinetic attestation device is just GeoSeal with a different AAD: instead of lat/lon/tile, the context is crank_fft/effort/duration. Same sealing logic, same tamper detection, same audit trail. It becomes a physical peripheral for the SCBE provenance layer.
Also interesting as a side angle
- Space mission fallback: FRAM is radiation-tolerant, supercap survives eclipse transitions, brownout-aware state machine is what aerospace firmware wants anyway. A totally independent kinetic-powered crew log for loss-of-main-power scenarios is a legitimate cubesat-adjacent niche — not a product but a reference design you could license.
- Kinetic AI training: Crank rate, rhythm, and pause patterns are a novel annotation signal. Effort correlates with cognitive load on the speaker. "This utterance required thinking" becomes a free label. Nobody collects this.
Live Bit-Signature Visualization in governance-gate Demo
Sketching
Type a prompt, see its Six-Tongues bit signature in real time, watch the classifier score it.
Just published the prompt-injection bit-signatures dataset and got 92% AUC on 31 features. The governance-gate demo already has pattern detection but doesn't show the bit signature. Next step: compute the signature in-browser, render the 16-bin bit histogram as a live bar chart, export the gradient boosting model as JSON so the classifier score updates live as the user types.
Ensemble: pattern detector + bit-signature classifier
Concept
The governance-gate pattern detector catches named attacks with high precision. The bit-signature classifier catches distribution-level weirdness. Combine them and publish the joint AUC.
Both are already built and running. The pattern detector is rule-based and precise but misses novel attacks. The bit classifier is statistical and picks up distributional signals the rules miss. Weighted ensemble should beat either alone. Publish as an update to the evidence page.
Third-party findings report (GPT-4, Claude, Llama 3)
Concept
High leverage
Run the SCBE L6 adversarial suite against the frontier models and publish the results.
The single highest-leverage asset I could create right now. One afternoon of API calls, one PDF, one honest report. Turns "I have a framework" into "I found vulnerabilities in GPT-4 that have real reproduction steps." Requires API credits (I pay my own) and permission-to-publish on the framing. Models mentioned would not be targeted in any misleading way — it's a fair comparison, not a takedown.
Rebrand the pipeline: "Three-Mechanism Combined Defense" instead of "14-Layer"
Concept
The real product is the combined detection, not the layer count.
The evidence page already shows the raw 14-layer pipeline scores 0.054 AUC on subtle attacks — worse than random — and only the three-mechanism combined defense (phase + tonic + drift) hits 99.42%. The "14-layer" branding is load-bearing and technically misleading. Reframe the marketing copy around the combined defense and demote the 14-layer count to an implementation detail.
New ideas drop here first, before they become products.
If you want to suggest one,
email me.