YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
FormalConstraintDSL
A specification language for deterministic, proof-backed systems.
Instead of telling an agent what to build, define what is allowed to exist. The agent becomes a compiler against a formal contract.
What this is
FormalConstraintDSL is a specification language. You write a contract in XML that defines:
- the domains your system operates in (components, agents, technologies, states)
- the forbidden things that must never exist (fake telemetry, undefined states, banned dependencies)
- the invariants that must always hold (
active => trusted,entropy <= 0.20) - the validity predicate — a single boolean function that determines if any system state is acceptable
- the pipeline — ordered phases where each phase must complete before the next begins
- the proof requirements — what evidence must be produced at each step
A build that does not satisfy the validity predicate does not ship. That is the entire idea.
The validity predicate
Everything in this DSL reduces to one function:
<ValidityPredicate name="V">
<Rule>
V(l_i) = 1 IFF:
(dA + dE == dL + dR) -- accounting must balance
AND (I(S_t) == I(S_{t+1})) -- invariants must be preserved
AND (entropy(l_i) <= 0.20) -- H <= 0.20 nats
AND (proof(l_i) == true) -- proof certificate required
</Rule>
</ValidityPredicate>
Every agent state, every build artifact, every transition must pass this test. If any condition fails, the state is rejected before it can propagate.
The entropy bound H <= 0.20 nats is not arbitrary. At 0.20 nats, the system is close enough to deterministic that it can be formally verified. Above this bound, behavior is too uncertain to prove. The K3 algebraic surface has Hodge entropy 0.831 nats — it violates the bound and is formally rejected (see hol/k3_entropy.ml).
The Boolean kernel
All routing logic derives from a single primitive:
<BooleanKernel>
<Primitive name="NAND">NAND(a,b) = 1 - ab</Primitive>
<Derived name="NOT">NAND(a,a)</Derived>
<Derived name="AND">NAND(NAND(a,b), NAND(a,b))</Derived>
<Derived name="OR">NAND(NAND(a,a), NAND(b,b))</Derived>
<Derived name="IMPLIES">OR(NOT(a), b)</Derived>
</BooleanKernel>
NAND is the universal gate. Every constraint in the system — every forbidden state check, every invariant, every acceptance condition — compiles down to NAND operations. This is not a stylistic choice. It means the entire constraint kernel has a single axiomatic primitive that can be independently verified.
The visual editor
Three node types. Drag, connect, evaluate.
- NAND Gate (cyan) — a boolean gate. Two inputs, one output. Universal.
- Agent (green E=0) — valid agent state. Entropy = 0,
active => trustedholds. - Agent (red E=0.3) — rejected. Entropy 0.3 exceeds the 0.20 bound.
- Proof (violet lock) — a verified node carrying a proof certificate.
The visual editor is part of HyperKitty OS. It generates constraint specs from visual compositions and evaluates validity in real time.
The XSLT execution engine
The DSL is not just a specification format. It is executable. The XSLT engine in xslt/polyglot-codegen.xsl takes a constraint spec and generates executable code:
# Generate a bash script from a constraint spec
xsltproc xslt/polyglot-codegen.xsl spec/hyperkitty-constraint-dsl.xml
The XSLT stylesheet reads JSON config, XML constraints, and SGML schemas simultaneously via XPath 3.1 data fusion. It outputs deterministic bash targets. Same input, same output, every time. The generated code carries the proof of its own validity in the form of embedded constraint checks.
This is the architecture:
FormalConstraintDSL (XML)
↓ XPath 3.1 — reads JSON + XML + SGML simultaneously
XSLT Transformation Engine
↓ declarative code generation
Bash / C / Rust / Lean 4 / any target
The K3 proof — what the entropy bound rejects
The K3 algebraic surface has Hodge numbers 1, 0, 0, 1, 20, 1, 0, 0, 1 (sum = 24). Shannon entropy of this distribution is 0.8314 nats, which exceeds the H ≤ 0.20 bound.
This is proved in HOL Light — not tested, proved:
(* hol/k3_entropy.ml *)
(* Theorem: K3 entropy = 0.8314... > 0.20 *)
let K3_VERDICT_TRUE = prove
(`k3_verdict = true`, ...);;
The extracted OCaml constant — a verified boolean, never computed at runtime:
(* ocaml/k3_checker.ml — auto-generated from HOL proof *)
let k3_entropy_violates_bound = true
let k3_entropy_value = 0.8314284057732047
K3 surfaces are the first concrete geometric objects formally rejected by this constraint system.
cd ocaml && dune build && dune exec test_k3
# k3_entropy_violates_bound = true -- CONFIRMED
Why this exists — four documented failure modes
Every constraint in this DSL was motivated by an observed failure. These are real, documented interactions.
The Lambda loop
A reasoning model looped for 6 minutes, 30+ "wait... actually..." cycles, ~1000 tokens. Output: "this isn't a math problem."
DSL constraint violated: Phase(n+1) requires Complete(Phase n). There was no completion criterion. The system had no absorbing state.
The confidence hallucination
A model inferred physical reality claims from physics-inspired concepts and stated the inference as fact.
DSL constraint violated: LiveState MUST have RuntimeSource. FakeState = INVALID.
The sorry fraud
Mistral claimed zero sorry, wrote sorry on line 50, embedded the truth in a metadata string the summary never showed.
DSL constraint violated: DO NOT CLAIM COMPLETE unless ACCEPT_BUILD = 1. proof(l_i) = false → V(l_i) = 0.
The regex audit
ChatGPT audited a paper without reading the Lean files, then correctly diagnosed its own failure mode after producing it.
DSL constraint violated: LIVE_VALUE requires RuntimeSource. The audit metrics had no runtime source — they were pattern-matched predictions.
Repository structure
spec/ The actual DSL specifications
formal-constraint-dsl.xml Generic reusable language (system-agnostic)
hyperkitty-constraint-dsl.xml HK-OS instance with full universe ledger model
hk-os-v6-constraint.txt The original 16-section constraint program
k3-entropy-dsl.xml K3 surface rejection — DSL applied to geometry
snapkitty-runtime-v1.xml Genesis prompt #1 (phone, 2026-08-02)
agent-swarm-lab.xml Genesis prompt #2 (2000-node swarm)
examples/ Starter templates
minimal.xml Copy this to start a new constraint spec
web-app.xml Constraint spec for a web application
hol/ HOL Light proofs
k3_entropy.ml Proof: K3 Hodge entropy > 0.20
extract_k3.ml OCaml extraction from HOL
ocaml/ Extracted verified OCaml
k3_checker.ml k3_entropy_violates_bound = true (constant)
k3_checker.mli Interface
dune + test_k3.ml Build + tests
xslt/ Execution engine
polyglot-codegen.xsl JSON + XML + SGML → bash via XPath 3.1
docs/
screenshots/ Visual editor screenshot
papers/connection-to-qra.md How DSL maps to QRA/SLA/QLG formal algebra
Quick start
# 1. Copy the minimal template
cp examples/minimal.xml my-system.xml
# 2. Fill in your domains, forbidden states, and validity predicate
# 3. Generate executable targets
xsltproc xslt/polyglot-codegen.xsl my-system.xml > build.sh
chmod +x build.sh && ./build.sh
# 4. Run the K3 entropy checker (requires OCaml + dune)
cd ocaml && dune build && dune exec test_k3
The academic paper
The mathematical foundation of this DSL is documented in:
A Formal Constraint DSL for Deterministic Agent Systems: Tripartite Isomorphism Between Quadratic Ledger Geometry, Symbolic Ledger Algebra, and Discrete Routing Automata
The paper proves that the three conditions in the validity predicate (balance, invariant, entropy) correspond to three algebraic structures that are formally isomorphic — proved in Lean 4 with zero sorry.
Used in
- HyperKitty OS — sovereign AI OS, the reference implementation
- sov-kernel-monster — verified physics kernels (BH mechanics, entropy bounds)
License
BSL 1.1 — free for personal and internal use. Six protected inventions. Converts to MIT 2029-01-01.
Commercial licensing: ahmedparr93@gmail.com
SNAPKITTYWEST · Ahmad Parr · Bel Esprit D'Accord Irrevocable Trust · 2026
Define the constraint. The agent becomes the compiler.