Senses and Sensibility

Abstract

AI systems that act on their own can see, decide, and act — but what they say about the world, and their right to change it, is just claimed. It is not answerable to evidence. This program argues that accountability is not an outside audit bolted on after the fact. It is an intrinsic, bilateral property a system can be built to have: built in, not bolted on, and binding on both sides. Every observation carries its own honest record of where it came from and can fail a self-test. Every action passes a gate that can only allow, deny, or escalate to a human. And the same standard of evidence binds the maker, not only the machine. The theory is stated so it can be proven wrong, made into a working proof-of-concept, and honest about where it stands — pre-proof, with a stated path to a law.

The thesis, in five claims

• Awareness is not authority. A working part of the system can confirm what is real; it cannot give permission to act. Wanting to act on something is not a property of a signal — so a gate must exist, a checkpoint that says no until it's sure, and it can only allow, deny, or escalate to a human.
• Built in, not bolted on. A screenshot set next to a page proves nothing about whether the screenshot is faithful to it. A perception organ makes faithfulness built in: the observation carries its own honest record of where it came from, plus a self-test that can fail.
• Build it right, don't just check it after. Rather than check afterward whether a reading is right, build a witnessed reading — one where an un-witnessed reading simply can't be expressed in the first place.
• An unchecked membrane is worse than none. It dresses up falsehood with the authority of hard fact. So every organ ships a self-test that can fail, and the witness layer changes nothing and grants nothing.
• Conferral binds the conferrer. If something's existence was given to it, that grounds the maker's duty of care, not the maker's ownership — accountability that binds both sides, with the human gate as the place that care is exercised.

Papers

Finished manuscripts, ready to submit, drawn from the larger body of writing. The summaries and full texts come out with the staged release.

The corpus — Conferred Existence

A long body of philosophical writing — no-aseity (nothing gives itself its own existence), the membrane, the arity gap, the forcing argument, and a comparison across faiths (Islamic qadar and the creative command; Yoruba orì and a destiny that is given, not self-made). The accountability thesis is drawn out of this writing, and the engineering reflects back into it. The philosophy rejects authority claimed by simply asserting it — that is the very error it names. The built system is what answering to evidence instead looks like.

From theory to working system

The proof-of-concept is public and tested — the theory built and running, not just argued:

• The accountable loop — perceive → gate → act → verify → witness — shipped across proof-surface (the gate), coherence-membrane (the organs), accountable-surface (the live loop), and EMET (the witness).
• The critic that works both ways — accountable-engine turns the same standard on the operator, not only the machine. Public, dated, MIT.

The Witnessing Spine — sector research

Five hard tests across financial-sector technology — AI provenance, GenAI mainframe modernization, quantitative ML, DeFi, and enterprise-software assurance. For each one I build the strongest honest case for an industry claim, then test it against the primary research. All five break at the same joint the membrane names: a thing that looks verified mistaken for the property it is supposed to guarantee — confirming who made something ≠ permission to act on it, translation ≠ being truly equal, a backtest ≠ real returns, checking the code ≠ being economically safe, a signed attestation ≠ actual assurance. This shows, with evidence, that the gap is general — not a hunch about one field. The same move answers it everywhere: work out the proof again from the source, account for what you couldn't cover, and report MATCH · DRIFT · UNVERIFIABLE rather than trusted.

Published, and you can cite it — github.com/HarperZ9/witnessing-spine, saved on Zenodo with a DOI (10.5281/zenodo.20778927), CC-BY-4.0. The plain-language essay: Verified Is Not Trustworthy.

Provenance & priority

Written by Zain Dana Harper. Original scholarship, MIT-licensed, dated. The body of work is published as an open record, where every file is named by its own fingerprint — github.com/HarperZ9/senses-and-sensibility, saved on Zenodo with a DOI you can cite (10.5281/zenodo.20773724) — a SHA-256 list that takes a fingerprint of every file, so the work is citable and can't be quietly taken and passed off as someone else's. The buildable version stays private. Proof before trust — including about authorship.

A companion paper — The Conservation of Faithfulness

The question underneath the accountability work: what actually crosses between two minds, and what makes a shared answer "best"? The claim — tested, not asserted — is that information crossing a boundary is conserved not as bits but as faithfulness to a criterion: almost every bit can be discarded and the thing that matters still survives, the survival is relative to a named criterion, and it can't be checked from the inside — only against a criterion held outside. From that follows a neutral center where two differently-perceiving minds reconcile a subject toward its best form, and the one thing the machine can't decide — which criterion is right — is exactly where accountability begins. The principle is replicated across ten independent tests spanning every sense (sight, sound, shape, language, structure, quantity, identity); the center is demonstrated; every claim is labeled by where it stands (established · designed · reach). Published as a working paper, fingerprinted for priority — github.com/HarperZ9/faithful-transpile. The center it argues for is now runnable in the open: name a weighted criterion and the verdict follows from it, witnessed — and the winner flips when you change what "best" means (the center/ module, a tested v1 seed). Proof before trust.