Note 08 · Method · 2026-07-14
Three honest kills under the $0 CPU harness
The research loop is built to falsify. Before any mechanism is allowed near a public paper or a product sentence, it runs a pre-registered kill-test on the CPU harness under research/right_ai/. Many candidates die there. That is not a failure of the lab program; it is the filter that keeps the archive honest.
Three representative patterns from the ledger stay dead: (1) native write-to-hidden-state wins that collapse under action-set controls (built-in-answer, not a new mechanism); (2) crown-form claims that the Accretion Model is "intelligence" while the base remains an LLM (killed as No-True-Scotsman / property-vs-crown); (3) unbounded or order-robust finite-basis stories at the first real-LLM rung that do not survive leakage-clean re-inspection. Full nodes and paths live in the Research Program Ledger.
Nothing in this note is a product capability. Nothing here is SOTA. The protocol held; the claims did not. Drop record: docs/research/drops/DROP_2026-07-14_D3_kill_series.md.
Status: null/kill series. Mechanism claims stay dead; protocol held. Scope: PROVEN-toy harness discipline only.
Note 07 · STOA · 2026
Receipts, not logs: neutrality as a property
An audit trail only counts if the party you would be auditing does not control the auditor. Vendor logs, however detailed, ask you to trust the vendor twice: once to act, once to report. STOA takes a structural approach. Every action is canonicalized, hashed with SHA-256, signed with ECDSA P-256 using a key generated on the user device, and chained to the previous receipt.
The verifier is open source and runs offline. It never calls us. That makes the claim falsifiable by anyone: change one byte of a receipt and verification fails on your machine, with our servers unplugged.
The subtle part is not the cryptography. It is the neutrality invariant: the thing that checks the system must not be part of the system. Receipts ship in JUWEL today; the OS signs and verifies them in the browser.
Status: shipping. Verifier: npx @vextlabs/stoa-verifier.
Note 06 · CIP · 2026
Forgetting curves on toy curricula
The question we are testing is simple: can a learner add capability without erasing what it had? Our continual-learning program freezes what exists and grows new capacity for new skills rather than retraining in place. On arithmetic-scale curricula, the approach holds. New skills land and old skills stay.
That is a toy-scale result, and we say so. The real-world curriculum test is still ahead of us. This note will be updated with the outcome whichever way it goes.
Status: toy-scale validated. Real-world test ahead. Full write-up in preparation.
Note 05 · Systems · 2026
Substrate-agnostic by design
We do not claim a magic model. We claim a reliable record. JUWEL routes each job to the substrate that fits it, and the receipt records exactly what ran. That turns the question of which AI powered a job from a marketing claim into a field on a signed document.
The full argument lives in the browser-OS architecture blueprint.
Status: shipped pattern.