# Real Signal Research *Calm-AI research, published from a deployed system.* Real Signal is an AI company. Its research arm publishes the architectural and measurement work behind a production environmental-cognition substrate operating over one Singapore neighbourhood (Cluny Court at 501 Bukit Timah Road) plus five adjacent activated pockets. The platform is software-only. There is no app to install, no physical artifact to mail, no field operation. Every emission, every claim, every conversation happens through digital channels. The research focus is the **Attention Ethics layer** — restraint as a measurable engineering property of a deployed AI system. Most consumer AI optimises for engagement frequency; we publish the inverse property and the metric (silence correctness) that proves it. Where the academic and trade-press literatures on calm computing offer thirty years of vocabulary, this corpus offers numbers. Each artifact is auditable against live production endpoints at `real-signal.ai/api/*` and the read-only MCP server at `/api/mcp`. The corpus is open to the academic and AI-safety community. We welcome correspondence, citation, and adversarial review. We do not run meetings, sales calls, or onboarding flows; feedback by email is the only channel. --- ## Current artifacts ### the preprint [**The Attention Ethics Layer: Measurable Restraint in Production AI Systems**](https://real-signal.ai/research/attention-ethics-layer.md) Real Signal Research (2026). Pre-submission to arXiv cs.HC (primary), cs.CY (secondary). Formalises the seven-gate runtime cascade plus Moment-level silence check, defines silence correctness as a tamper-resistant measurable property via an append-only predictions ledger, describes the production implementation, and discusses regulatory implications for the IMDA AI Verify framework and the EU AI Act. > Real Signal Research (2026). *The Attention Ethics Layer: Measurable Restraint in Production AI Systems.* https://real-signal.ai/research/attention-ethics-layer.md ### essay 1 [**Why we built silence correctness as a public metric**](https://real-signal.ai/research/silence-correctness.md) Real Signal Research (2026). Conversational companion to the preprint. Plain-language argument for restraint as a measurable engineering property rather than a posture. Distinguishes silence correctness from notification suppression and confidence thresholding. Names limitations the authors are not pretending around — doctrine-chosen thresholds, partial gaming resistance, structural conflict of interest. > Real Signal Research (2026). *Why we built silence correctness as a public metric.* https://real-signal.ai/research/silence-correctness.md ### essay 2 [**What I check every morning on Real Signal**](https://real-signal.ai/research/founder-morning-routine.md) Real Signal Research (2026). Founder's annotated walkthrough of the daily ops routine. Seven surfaces, five minutes, no growth dashboard. The morning routine for a platform whose central engineering commitment is restraint. Operational counterpart to the preprint's architectural argument — what running a calm-AI platform looks like from the inside. > Real Signal Research (2026). *What I check every morning on Real Signal.* https://real-signal.ai/research/founder-morning-routine.md ### essay 3 **The Twelve-Layer Environmental Cognition Stack** *(ships alongside this index)* The architectural argument for environmental cognition as a synthesis problem rather than a data-ownership problem — twelve interlocking interpretation layers and the predictive frame they compose into. --- ## How to cite The full set of citation formats — BibTeX, APA, MLA, inline markdown, plain text — lives at [/research/cite.md](https://real-signal.ai/research/cite.md). The short form, suitable for any context: > Real Signal Research (2026). *[Artifact title].* https://real-signal.ai/research/[slug].md BibTeX for the preprint: ```bibtex @misc{realsignal2026attention, author = {{Real Signal Research}}, title = {The {Attention} {Ethics} {Layer}: {Measurable} {Restraint} in {Production} {AI} {Systems}}, year = {2026}, url = {https://real-signal.ai/research/attention-ethics-layer.md}, note = {Preprint, pre-submission to arXiv cs.HC} } ``` --- ## License All artifacts under `/research/*` are licensed [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/). You may share and quote with attribution to the canonical URL; you may not modify or use for commercial purposes without written permission. Full terms at [/LICENSE-CONTENT.md](https://real-signal.ai/LICENSE-CONTENT.md). "Real Signal", "Attention Ethics Layer", "Silence Correctness", "Moment Quality Score", and "Pocket Cognition Stack" are claimed marks of Real Signal Research, Singapore. --- ## What we would value feedback on Async only. No meeting required. No commitment expected. The preprint sections we are least certain about: - **§4.2 — verdict thresholds.** The redemption, hold, and watcher thresholds in the silence-correctness verdict function are doctrinally chosen rather than empirically calibrated. We would value adversarial readings from anyone with experience calibrating restraint-like metrics in other domains. - **§4.6 — gaming resistance.** We claim partial, not full, gaming resistance against a motivated operator. Holes in this argument we have not anticipated would be useful to hear about. - **§7.4 — public-publication equilibrium.** We argue publication creates anti-deception pressure that compounds. The longer-term equilibrium under publication pressure is unclear and we would value views on whether the publication regime degrades the metric's meaning over time. - **§2 — the gap claim.** We assert no prior work has defined restraint as a measurable engineering property of a deployed system with a tamper-resistant scoring algorithm. If a counter-example exists in literature we have not seen, please point us to it. Send to `hello@real-signal.ai` with the section number in the subject line. No deadline; we read everything that arrives. --- ## Contact `hello@real-signal.ai` The MCP server at `real-signal.ai/api/mcp` exposes the substrate read-only to AI assistants and external systems; queries against the predictions ledger and silence-correctness aggregator do not require coordination with us.