The internals, in the open. A deterministic core, an adaptive shell, and a human in every loop — engineered so intelligence can grow without ever escaping accountability.
This is not artificial general intelligence. It is aligned general stewardship.
Value — credits, valuations, contracts — must be reproducible and auditable. So learning lives around the ledger, never inside it. Models propose; humans commit.
Every value-bearing operation can be replayed from hashes alone and reconstructed exactly. A failure in one node becomes a learning event for all — never a contagion. This is the part you can take to a regulator.
The Moment Manager senses, predicts, and suggests — with confidence intervals visible. It can never mutate the record directly. Authority flows through transparency, not autonomy.
A meta-compiler turns human rules into grammars, workflows, and deterministic services — so machines can reason, humans can audit, and every result can be replayed and explained.
GRAMMAR water_quality { WHEN sensor.turbidity > baseline REQUIRE human.review # HITL gate EMIT observation { epistemic_status: measured # calibrated sensor virtue_tags: [Stewardship, Transparency] } EXPLAIN "why / why-not" # replayable }
Meanings connected across water, land, health, mobility — one executable language.
Infer, plan, and justify within constraints — with replayable explanations.
Legal text, executable code, and human-readable math — the same truth, three languages, one hash.
You can't bolt virtue onto a monster from the outside — it has to be the core. Each proposed action becomes a ten-component vector in moral space. Its magnitude is coherence. Actions misaligned with virtue incur synthetic "pain" — a drop in score that guides correction without coercion.
A configurable cognitive window — context that bounds awareness to the moment, then reflects. Predictions are risk-tiered, and the riskier the inference, the more consent and explanation it demands.
Scheduling, retrieval, reminders. Output is a visible suggestion — no special consent required.
Habits and routines require active consent; outputs carry confidence intervals you can see.
Emotional or wellbeing inferences are labeled inferred, reviewed by a human, and explained in plain language before any action.
PAL proposes, captures consent, and records a clear why / why-not — and never commits a change without your authorization.
This is the part no other system has built: every assertion the architecture emits carries a machine-readable, immutable status. A soil-carbon measurement backed by five years of calibrated sensing is categorically different from a model's estimate — and the system never lets them look the same.
From a calibrated sensor, with quantified uncertainty bounds.
Cryptographically anchored via an independent attestor.
Algorithmically derived from other records — no new measurement.
A probabilistic estimate from a model, with a confidence interval.
Consented human disclosure, with an irreducible subjective part.
We publish our thinking as knowledge nuggets: atomic, cited, status-tagged, and open to review. Disagreement is signal. Tell us where we're wrong.
An SDL-based city can eliminate 60–80% of redundant data infrastructure within five years where systems are highly duplicative.
Coherence is the magnitude of a ten-virtue vector; below 0.65 the system pauses for human review.
Every cognitive loop — sensing, deciding, learning, evolving — should contain at least one authenticated, accountable human node.
Comments are public and threaded — sign in with GitHub to weigh in on any topic above. Or email [email protected].