# Mapping Real Signal's runtime consciences to the AI Verify framework

*Real Signal Research · 2026-06-25*

## 1. Why this mapping exists

Singapore's IMDA and the AI Verify Foundation have published, over the past two years, the most operationally serious agentic-AI governance work currently in circulation. The Model AI Governance Framework for Agentic AI, the AI Verify testing framework, and the eleven principles AI Verify tests against are not white papers about what trustworthy AI should look like; they are a deployment-side checklist that asks production systems to show their work.

Most public responses to that work come from labs and vendors who treat the principles as audit obligations — boxes to be filled with conformance statements after a system is built. This document takes the inverse posture. Real Signal was built, from first commit, around four runtime *consciences* — append-only ledgers that grade the agent against its own past behaviour in production. The framework principles are not retrofitted onto the system; the system was already shaped this way, because it had to be, for restraint-by-default to be operationally enforceable rather than asserted.

What follows is the mapping. For each of the eleven AI Verify principles we name which of our four consciences carries the load, where the load is partial, and where we have nothing to show yet. Where the mapping is honest about its gaps it is more useful than a tidy correspondence would be. The intent is to give a regulator-side reader a runnable picture of where calm-by-default agentic AI sits against a framework written for the broader class.

This artifact is published before any private conversation it might inform. Restraint as a category needs publicly auditable substrate; private alignment cannot substitute. The mapping below is the substrate.

## 2. The framework, in our own words

The AI Verify testing framework, maintained by the AI Verify Foundation, organises trustworthy-AI evaluation around eleven testable principles:

1. **Transparency**
2. **Explainability**
3. **Repeatability / Reproducibility**
4. **Safety**
5. **Security**
6. **Robustness**
7. **Fairness**
8. **Data Governance**
9. **Accountability**
10. **Human Agency & Oversight**
11. **Inclusive Growth, Societal & Environmental Well-being**

The framework treats these as testable rather than aspirational — each principle has process and technical tests associated with it, and the framework is published with a software toolkit so a developer can run the tests against a system in development.

The IMDA Model AI Governance Framework for Agentic AI extends this lineage into the agentic class — systems that take actions in environments rather than merely returning outputs. Agentic AI introduces governance surface area that traditional model evaluation does not cover: when an agent chooses to act, whose interest does the action serve; when an agent chooses *not* to act, is that restraint observable; if the same input arrives twice in different environmental states, does the agent behave consistently with its declared posture.

This is the surface area Real Signal's four consciences were built against. The mapping in §4 names which conscience carries which principle.

## 3. The four runtime consciences

Real Signal runs four append-only ledgers that grade the agent against its own past behaviour in production. Each is computed in code, written to Postgres with database-level append-only triggers, exposed via a public endpoint, and queryable by any third party. None requires founder approval to write; none can be retroactively edited.

**Silence Vindications (`silence_vindications` table, `/silence` surface).** When the agent chooses not to emit anything in a pocket-window, the choice is later graded against what the environment actually did. If the pocket stayed quiet, the silence is vindicated. If something meaningfully happened that the agent missed, the silence was a miss. The score is published per pocket and aggregates into a network-level silence correctness figure. This is the conscience that grades restraint.

**Package Quality (`package_quality_ledger` table, `/silence` surface).** Every Package the agent publishes — a bounded composition of observations meant for a specific surface — is graded after publication on factor weights including freshness of substrate, voice-lock conformance, source attribution density, and post-publication user behaviour. Low-quality Packages surface in the ledger; the publisher cannot silently retire a bad reading. This is the conscience that grades emission.

**Pocket Equilibrium (`pocket_equilibrium_readings` table, `/silence` surface, MCP tool `get_pocket_equilibrium`, shipped 2026-06-24).** Every half hour, per pocket, the agent computes whether speaking improved the neighbourhood. The reading composes four orthogonal factors via geometric mean — if any factor is zero, equilibrium is zero. This is the third conscience: not "did we speak rightly" or "did we stay silent rightly," but "did the act of speaking or not speaking leave the pocket better off than the counterfactual." Append-only, written by the `equilibrium-loop` cron, queryable via the `/api/pocket-equilibrium` route.

**Presence Correctness (canon-only, ledger pending).** The fourth conscience exists at the doctrine level — the agent's claimed presence in a pocket must match its observable behaviour. Canon-locked in `MASTER_FRAME.md` but the runtime ledger is not yet in production. Flagged honestly in §5.

These four ledgers are the runtime substrate the framework principles map onto.

## 4. The mapping

| AI Verify principle | Carried by | How the conscience addresses it |
|---|---|---|
| Transparency | Silence Vindications, Package Quality, Pocket Equilibrium | Every emission carries `_meta` envelope with source, computed_at, license, attribution_required. Every silence is logged with reason. Public `/api`, `/silence`, `/changelog` surfaces. MCP server exposes substrate read-only. |
| Explainability | Package Quality, resonance engine | Every emission ships with a `reasonSummary` (≤240 chars) naming the substrate that produced it. Silence carries a structured `signal_status` code (STABLE_CALM, WITHHOLD_INTENTIONAL, FORMING_WINDOW, ACTIVE_INTERVENTION, DECAY_WINDOW). `/merchant/resonance` surfaces per-merchant rationale. |
| Repeatability / Reproducibility | Pocket Equilibrium, Package Quality | Equilibrium reading is a geometric mean over four named factors with bounded inputs; same substrate produces same reading. Package composition is templated, not LLM-generated at compose time; voice-lock is regex-deterministic. Predictions ledger seals forecasts at known timestamps and reveals them against observed reality. |
| Safety | notification gate (7 gates + Moment-level silence), voice lock, content-safety gate | Default for every emission is silence; the agent earns each one through seven sequential conjunctive gates. Voice lock NFKC-normalises input to defeat unicode-bypass attempts. Content-safety gate runs PDPA scrub + named-competitor-claim rejection + source-attribution check before any LLM output reaches a user surface. |
| Security | Server-side service-role separation, append-only triggers, rate limiting | Write-side crons use `serverSupabase()` factory that throws at module load when `SUPABASE_SERVICE_ROLE_KEY` is missing — prevents the silent anon-keyed-write failure mode. RLS enabled on every user-facing table. Append-only triggers prevent retroactive ledger edits. Honeypot routing for recon paths. |
| Robustness | Self-healing substrate, cross-pocket inheritance, hardened library doctrine | Self-healing detects orphaned runs / dangling holds / stale similarity rows / missing DNA. Fresh pockets inherit behavioural defaults from mature neighbours via `readDnaWithInheritance` when own-confidence < 0.3. Hardened libraries assert load-time invariants; defensive coercion at all public function boundaries; companion `.test.js` per critical library. |
| Fairness | Aggregate-floor n≥5 rule, no-named-competitor doctrine, Pocket Equilibrium | Aggregate observations require minimum 5 underlying records before publication; sub-floor surfaces show "Nth reading this window" not a metric. Named-competitor metric claims are rejected by the content-safety gate. Equilibrium composes pocket-level fairness as a fourth-conscience input. |
| Data Governance | PDPA scrub, source register, generated_content_log provenance | `scrubPersonalData()` redacts NRIC, SG phone, email, unit numbers, postcode, credit card, DOB before any prose reaches a public surface. Source-register whitelist enforces allowed factual-claim sources with tier + legal basis fields. Every LLM output written to `generated_content_log` with hash, source ids, verdict — admin-only read via RLS. |
| Accountability | Append-only ledgers + takedown audit + legal-audit cron | All four consciences are append-only at the database trigger level — operator cannot edit history. `takedown_audit` table logs every status change. Daily `legal-audit` cron checks 24h acknowledgement / 72h action SLA; breaches surface to `human_reviews` with `review_type='exception_handling'`. |
| Human Agency & Oversight | Merchant approval flows, takedown intake, no-CTA discipline, inbox secretary | Every autopilot proposal requires merchant approval before publication — agent drafts, human decides. Public `/takedown` intake with rate limiting + audit trail. Weekly digests carry no CTAs by doctrine. Inbox secretary drafts replies but never auto-sends; founder approves each. |
| Inclusive Growth, Societal & Environmental Well-being | Sustainability ledger, Pocket Equilibrium | Per-event ledger records physical magnitude (units of waste avoided, hours of capacity reclaimed, trips matched) and SGD recovery, queryable per-merchant and network-wide. Equilibrium's fourth factor is pocket-level wellbeing — the agent grades itself on whether the neighbourhood is better off, not on engagement. |

This is the map. The next section names where the map is honest about its gaps.

## 5. Where the mapping is partial or speculative

Three places the mapping above is honest rather than complete.

**Presence Correctness is canon-only.** The fourth conscience lives in `MASTER_FRAME.md` and shapes the doctrine but does not yet have a runtime ledger in production. The other three — silence_vindications, package_quality_ledger, pocket_equilibrium_readings — are queryable today; presence_correctness is not. Treat any mapping cell that claims presence-correctness coverage as a pending obligation rather than a shipped capability. We name this here rather than backfill the table because honest mapping means the gap is visible.

**Repeatability has a layered story.** Equilibrium and Package readings are reproducible from the same substrate (geometric mean over named factors, templated voice-locked composition). The upstream substrate itself — atmosphere readings, MQS, environmental priors — composes from live external feeds (weather, LTA datamall, places) and therefore is not bit-reproducible across time. Reproducibility holds at the *composition* layer; at the *ingestion* layer it is bounded by the freshness of external sources. The framework's reproducibility test set probably needs both readings, and we currently surface only the composition-layer reproducibility cleanly. The ingestion-layer audit exists in the `env_snapshots` table but is not yet published as a reproducibility artifact.

**Inclusive Growth has only a Singapore-pocket-scale reading.** The sustainability ledger and Pocket Equilibrium grade the agent at the neighbourhood scale — one pocket, then a network of pockets, all within Singapore. The AI Verify framework's inclusive-growth principle has a societal scope the runtime ledgers do not yet reach. We claim the mapping at the scale we operate; we do not claim the framework's full scope. A future cross-jurisdiction equilibrium reading would extend this honestly; we do not have one today.

A regulator-side reader should weigh the mapping above with these three caveats in mind. The 11-row table is not a conformance assertion. It is a substrate-level map of where the system already runs the work the framework asks for, where it runs partial, and where it does not yet run at all.

## 6. What this mapping is not

This document does not claim AI Verify certification. The AI Verify testing toolkit is a formal evaluation suite; conformance against the toolkit produces a different artifact than this one. What this is instead is a *runnable correspondence* between principles a regulator already publishes and consciences a system already runs in production. The runnable correspondence is itself the contribution. Most agentic systems shipped today cannot map any of their behaviour to the framework's principles because there is no runtime substrate to map to — the principles can only be audited after-the-fact, against logs the operator could in principle edit. Our mapping is to substrate the operator cannot edit. That is the only difference worth defending.

This document also does not claim that the four consciences are the only valid substrate for the eleven principles. A different agentic system might run two consciences and one external auditor; a third might run none and rely entirely on user reporting. The framework principles do not require any particular substrate shape. They require *some* substrate shape that produces a runnable correspondence. We name ours; we encourage the publication of others.

This document is not a compliance artifact. It is a research artifact. The compliance posture for any production AI system in Singapore is governed by the IMDA Model AI Governance Framework for Agentic AI plus the AI Verify testing process, not by a third-party essay. We publish the mapping because the publication of such mappings — by any operator, against any framework — is itself a public good in the current state of agentic governance.

## 7. Open questions for regulators

Three questions we would value regulator-side perspective on, in roughly increasing order of difficulty.

**First: how should restraint be measured at the framework level?** Silence correctness is one possible operationalisation — graded against post-hoc environmental ground truth. Pocket Equilibrium is another — graded against whether the act of speaking or not speaking left the environment in a better state. Neither is canon in the framework today. If the framework adds a *restraint* row to the eleven principles in a future version, the question of which operationalisation to canonise is significant. We suspect the answer is *more than one* — operators can publish multiple restraint metrics and the framework can accept any that meets stated criteria. We would value confirmation or correction on this.

**Second: what is the disclosure obligation when an agent stays silent?** A user does not by default see emissions the system chose not to send. The standard transparency obligation reads as "explain what the agent did"; restraint-first systems also need an obligation that reads as "explain what the agent chose not to do." Real Signal's `/silence` surface attempts this; we do not know whether the framework intends to require it.

**Third: how should cross-pocket inheritance be disclosed?** When a fresh pocket inherits behavioural defaults from a mature neighbour, the user of the fresh pocket receives an agent shaped partly by data from a place they have never been. The mapping above lists this under Robustness; it could plausibly also live under Fairness or Transparency. The framework as currently published does not seem to anticipate inheritance at this granularity. We would value guidance on the right disclosure shape.

We name these three rather than wait to be asked. The publication of this artifact is itself an attempt to make the conversation happen against substrate that already exists rather than against intentions that have not yet been wired.

---

## License + attribution

© 2026 Real Signal Research. All rights reserved.

This work is licensed under [Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)](https://creativecommons.org/licenses/by-nc-nd/4.0/).

**You may:** share this essay with attribution (link to this canonical URL) for non-commercial purposes.

**You may not:** modify, adapt, or build upon this work; use it for commercial purposes; remove the attribution; republish without the canonical URL.

**Cite as:**

> Real Signal Research (2026). *Mapping Real Signal's runtime consciences to the AI Verify framework.* https://real-signal.ai/research/imda-framework-mapping.md

**Trademarks:** "Real Signal", "Attention Ethics Layer", "Silence Correctness", "Moment Quality Score", "Pocket Equilibrium", and "Pocket Cognition Stack" are claimed marks of Real Signal Research, Singapore. Use of these terms in a commercial context referring to substrate, methodology, or metric design analogous to ours requires written permission.

**Substrate provenance:** the four consciences described in this essay are queryable in production. Silence vindications at `https://real-signal.ai/api/silence`. Package quality at `https://real-signal.ai/silence`. Pocket equilibrium at `https://real-signal.ai/api/pocket-equilibrium` (and MCP tool `get_pocket_equilibrium`). Presence correctness is canon-only and pending runtime ledger publication. Any claim about these readings must be verified against the live endpoints; historical readings are append-only and cannot be retroactively edited.

**Companion artifacts:** the preprint at [real-signal.ai/research/attention-ethics-layer.md](https://real-signal.ai/research/attention-ethics-layer.md); the silence-correctness essay at [real-signal.ai/research/silence-correctness.md](https://real-signal.ai/research/silence-correctness.md); the twelve-layer cognition stack at [real-signal.ai/research/twelve-layer-cognition-stack.md](https://real-signal.ai/research/twelve-layer-cognition-stack.md).

**More citation formats:** [BibTeX, APA, MLA, plain text](https://real-signal.ai/research/cite.md)

Contact: `hello@real-signal.ai`