01
Paid work
Teleox may be paid for creator archive setup, private memory work, proof runs, or registry access. The commercial relationship is disclosed.
Chris Royse / leapable.ai
Governance - scope guards, disclosures, operating boundaries
Strong claims need boundaries.
Every strong claim should say what was measured, what is pending, and who may have paid.
02
Operations
Teleox.ai is the public site for Chris Royse's research and contact path. leapable.ai hosts expert context, private databases, transcript search, registry writes, verification calls, and billing.
03
Scope
Case 3, the 91-signal ceiling, the Shumailov scope argument, and G_tau all carry public guard language. G_tau is a cosine-threshold predicate and a different property class than scalar reward.
Why every claim starts as unverified.
Watch on YouTube· 5:02
Verification workflow: guilty until proven innocent
This belongs on governance, not the home page. Chris explains the verification posture behind agentic engineering: assume the output is wrong until a separate investigator can trace evidence, failure modes, and confidence. That is the same posture teleox.ai applies to the Case 3 measurement, the 91 constructive ceiling, Shumailov as a scope argument, not a refutation, and G_tau as a different property class.
Watch videoOpen the full video on YouTubeScope guards
Case 3 — voice cloning result
0.961 mean WavLM-SECS (max 0.975, N=10, one speaker, English, within-WavLM-family). ECAPA-TDNN EXP-1 cross-encoder validation pending. Not a parity claim against any published TTS system.
Counting identity upper bound
up to 91 supervisory signals per input, from n · (N + C(N,2)) at N=13, as a constructive upper bound. Pairwise-MI (EXP-2) audit pending. Realized floor will be published alongside the ceiling.
Shumailov recursion regime (scope argument)
Structurally outside the recursion regime assumed by Shumailov et al. (2024). Scope argument — not a refutation of the collapse result. One constructive path for deriving labels from fixed real data.
G_τ — predicate alignment class
G_τ is a per-output cosine-threshold predicate over embedder outputs. This is a different property class than scalar-reward alignment (e.g., RLHF), not a benchmark contest with it.
Issuer-pays disclosure
Teleox may be paid for cooperated evidence rows, verification API access, proof runs, or methodology review. Cooperation permits accuracy review of the evidence trail; it does not alter rubric assignment.
Operational disclosure
Teleox.ai publishes the research, public proof trail, Standard, and contact intake. leapable.ai hosts registry writes, transcript search, verification API calls, and billing.
Have one context problem?
Send the audience, data type, target task, proof bar, and sharing limits. The first pass should stay bounded.