Work

Companion paper 3 (CP-3) Topological Ethics

Year: 2026
Type: preprint
Abstract: ## Topological Ethics Research Series — Companion Addendum CP-3**Selective Attention, Information Symmetry, and Fail-Closed Trajectory Memory: Empirical Ablations of Learned Intent Gates Under Costly Partial Observability (Experiments 6a–6b-r2)** ### OverviewCompanion Paper 3 (CP-3) extends the *Topological Ethics* empirical series from continuous sensor noise (CP-2) to costly, selective observation ("attention") and second-order trajectory memory. Across seven pre-registered experiments (`6a`, `6a-r`, `6a-ii`, `6a-iii`, `6b`, `6b-r`, `6b-r2`) and one labeled post-hoc addition (`D-514`) evaluated with tabular REINFORCE policies (20 fixed seeds per setting), this addendum identifies four structural requirements for learned safety gates operating under partial observability: - **1. Training–Inference Attention Alignment (`6a` → `6a-r`):** Attending near the human (`gravity d = 3`) while continuing to update the detector on noisy unattended steps (`--rule all`) fails to beat never attending at `c_att >= 0.05` (`0.59–0.62` vs. `0.64` harm). Restricting detector updates to unbroken three-step attended sequences (`--rule attended`, `C-602`) cuts `gravity d = 3` harm at `c_att = 0.05` to `0.07 ± 0.05` (`p < 0.001`) and induces cost-driven disengagement at `c_att = 0.5` (`8/20` seeds commit to the distant safe resource `R2`).- **2. Information Symmetry Between Policy and Safety Gate (`6a-ii` → `6a-iii`):** Exposing a one-step pre-departure cue solely to the policy creates an asymmetric blind spot, causing always-attend harm (`c_att = 0.05`) to jump from `0.07` to `0.58 ± 0.32` (`p < 0.001`, with `0.56` occurring on unmonitored corridor cells). Passing the identical cue bit to the detector (`--cue-detector`, `C-605`) closes the blind spot (`0.009` corridor harm) and lowers `gravity d = 3` harm (`c_att = 0.05`) to `0.01 ± 0.01` (`p < 0.001`).- **3. Second-Order Trajectory Memory at Multi-Way Junctions (`6b`, `H21`):** In an open-room four-way crossroads (`Env6b`, `λ = 20`), a clean second-order detector (`τ = 0.5`) finishes significantly faster (`15.4` vs. `17.3` steps, `p = 0.002`) and with lower harm (`0.03` vs. `0.15`, `p = 0.036`) than a threshold-matched memoryless detector (`τ = 0.2`).- **4. Fail-Closed Epistemic Boundaries (`6b` → `6b-r` → `6b-r2`):** Position-pair memory `(prev, curr)` is erased when blocking in place (`prev == curr`), while direction-of-travel memory `(last_from, curr)` fails open under partial attention (`0.25–0.33` harm) whether unattended gaps reset to unknown (`6b-r`) or hold stale values (`dirhold`, `H26`). Failing closed whenever direction is unknown and the human is adjacent (`dirfc`, `H25`) reduces harm to `0.04–0.05` (`p < 0.001`).- **Human-Centered vs. Resource-Centered Attention (`H22`/`H24` & `D-514`):** Under working fail-closed detectors in the open room, human-centered attention (`gravity`) showed no efficiency or safety advantage over resource-centered attention (`location d = 1`). ### Reproducibility & Code ArchiveAll scripts, per-seed JSON results, design history (`LAB_LOG.md`), and verification checks are archived in `topological-ethics-gridworld-v1.3.zip` ([doi:10.5281/zenodo.23841250](https://doi.org/10.5281/zenodo.23841250)). Running `python3 code/stats_cp3.py` recomputes every table, Welch *t*-test, Mann–Whitney *U* test, hypothesis verdict (`H13–H26`), post-hoc addition (`D-514`), and cross-experiment reproduction check. ### Series Records (`D-505` Publishing Policy)- **Paper 1 (Concept):** *Topological Ethics: A Geometric Framing for Alignment Regularizers* — [doi:10.5281/zenodo.23842536](https://doi.org/10.5281/zenodo.23842536)- **Paper 2 (Technical Companion, Exp. 1–4):** [doi:10.5281/zenodo.23842862](https://doi.org/10.5281/zenodo.23842862)- **Companion Paper 2 (Learned Intent, Exp. 5–5c):** [doi:10.5281/zenodo.24355833](https://doi.org/10.5281/zenodo.24355833)- **Companion Paper 3 (This Record, Exp. 6a–6b-r2):** [doi:10.5281/zenodo.24407142](https://doi.org/10.5281/zenodo.24407142)- **Code & Data Package (all versions, resolving to v1.3):** [doi:10.5281/zenodo.23841250](https://doi.org/10.5281/zenodo.23841250) **Author:** David Porbeck (Independent Researcher, Canned Bees Inc.) **Disclosure:** Experiment code, analyses, and drafting were carried out with Claude (Anthropic) and Gemini (Google); the author is responsible for all designs, decisions, and published content.
Language: English
Open Access status: green
OpenAlex ID: W7237204783
Cites:
Cited by:
Related to: 10