|
| 1 | +// SPDX-License-Identifier: CC-BY-SA-4.0 |
| 2 | +// Copyright (c) Jonathan D.A. Jewell <j.d.a.jewell@open.ac.uk> |
| 3 | += ⚠️ Prototype status, disclaimers & how this fits the ethics work |
| 4 | + |
| 5 | +[.lead] |
| 6 | +*Conative Gating is an early-stage research prototype, not a product.* It |
| 7 | +explores one idea — using a small model as an inhibitory "NO-GO" antagonist |
| 8 | +to a larger one — and shares it openly so others can poke holes in it. |
| 9 | + |
| 10 | +== What "prototype" means here |
| 11 | + |
| 12 | +* *Experimental.* Interfaces, schemas, the ABI/FFI surface, config formats |
| 13 | + and the gating model itself are all subject to change without notice, and |
| 14 | + the whole approach may be revised or abandoned. |
| 15 | +* *Unproven.* Nothing here has been independently evaluated. Benchmarks and |
| 16 | + claims in this repo are exploratory, not validated results. |
| 17 | +* *No warranty.* Provided "as is", with no warranty of any kind, to the |
| 18 | + extent permitted by the licence. See `LICENSE` / `NOTICE`. |
| 19 | + |
| 20 | +== Safety disclaimer (please read) |
| 21 | + |
| 22 | +Conative Gating is *about* policy enforcement, which makes it easy to |
| 23 | +over-trust. It must not be treated as a safety guarantee. |
| 24 | + |
| 25 | +* *Not a sole guardrail.* Do not rely on this — or any single mechanism — |
| 26 | + as the only thing standing between a model and a harmful action. A gate |
| 27 | + that is analogous to the basal ganglia is still just a heuristic: it will |
| 28 | + have false "GO"s and false "NO-GO"s. |
| 29 | +* *Not security-audited.* Do not deploy it in production, in safety-critical |
| 30 | + settings, or anywhere an incorrect allow/deny decision could hurt someone, |
| 31 | + without your own independent review, defence-in-depth, and a human in the |
| 32 | + loop. |
| 33 | +* *Claims are hypotheses.* Statements about LLM behaviour and "loss aversion" |
| 34 | + are framing for an experiment, not established science. |
| 35 | + |
| 36 | +== Interest and collaboration warmly welcome |
| 37 | + |
| 38 | +This is shared to *invite scrutiny*, not to advertise a finished thing. If |
| 39 | +the idea interests, annoys, or worries you — that is all useful. |
| 40 | + |
| 41 | +* Open an *issue* or a *discussion* with questions, counter-examples, |
| 42 | + failure modes, or "this can't work because…". |
| 43 | +* Critique of the *ethics framing* (below) is especially valued. |
| 44 | +* See `CONTRIBUTING` and `CODE_OF_CONDUCT` before opening a PR. |
| 45 | + |
| 46 | +== Where the ethics thinking lives |
| 47 | + |
| 48 | +Conative Gating is one piece of a wider estate exploring how to make AI |
| 49 | +tools *trustworthy and humane by construction*. The normative / ethical and |
| 50 | +human-experience reasoning is developed in dedicated sibling projects — this |
| 51 | +repo deliberately defers to them rather than re-deriving ethics locally: |
| 52 | + |
| 53 | +[cols="1,3",options="header"] |
| 54 | +|=== |
| 55 | +| Project | Role in the ethics picture |
| 56 | + |
| 57 | +| https://github.com/hyperpolymath/phronesis[*Phronesis*] |
| 58 | +| A "practical wisdom" language — the estate's substrate for expressing |
| 59 | + normative and ethical reasoning, distinct from a model's raw knowledge. |
| 60 | + |
| 61 | +| https://github.com/hyperpolymath/phronesiser[*Phronesiser*] |
| 62 | +| Adds *normative ethical constraints* to AI agents via Phronesis. This is |
| 63 | + the natural home for the *why* and *what* of a NO-GO decision, where |
| 64 | + Conative Gating is only a *mechanism* for the *when/whether*. |
| 65 | + |
| 66 | +| https://github.com/hyperpolymath/vexometer[*Vexometer*] |
| 67 | +| The *interaction-ethics / UX* side: an "Irritation Surface Analyser" that |
| 68 | + measures the friction and indignity LLM tools impose on people. Gating |
| 69 | + that is technically correct can still be a bad experience — Vexometer is |
| 70 | + how that cost gets named. |
| 71 | + |
| 72 | +| https://github.com/hyperpolymath/action-trust-layers[*action-trust-layers*] |
| 73 | +| Sibling prototype on *layered trust* for automated actions — the same |
| 74 | + "don't grant blanket permission" instinct applied to supply chains. |
| 75 | + |
| 76 | +| https://github.com/hyperpolymath/palimpsest-license[*Palimpsest License*] |
| 77 | +| The ethical-use licence family this repo ships under; its Exhibit A sets |
| 78 | + out shared ethical-use expectations for the code. |
| 79 | +|=== |
| 80 | + |
| 81 | +If you only read one of the above for the ethics rationale, read |
| 82 | +*Phronesiser* (the *normative* side) and *Vexometer* (the *human-experience* |
| 83 | +side) — together they bracket what "good behaviour" means here. |
0 commit comments