wiseorder 0.1.0

WiseOrder — the kernel that governs how an AI's output is allowed to become an action, and seals every decision. One of three peers under Wise.Est Systems.

wiseorder-protocol

WiseOrder governs how cognition becomes consequence — the kernel that decides whether an AI's output is allowed to become an action, and seals every decision. It classifies each claim into a verification regime, runs a separate authorization gate before any action takes effect, and seals the resulting transitions in a hash-chained audit memory you can verify yourself, offline, with no account.

One of three peers under Wise.Est Systems. A judgment can never be recorded as a proven fact, and an AI cannot rubber-stamp its own actions.

---

1. Purpose

wiseorder-protocol is the kernel that governs how an AI's output is allowed to become an action — one of three peers under Wise.Est Systems. It defines:

• four artifact classes (A/B/C/D), each with a typed verification regime and a fixed set of status tokens
• an append-only audit memory that hash-chains every committed transition
• a hash-chained .win triple chain anchored at a NULLASIGN genesis
• an authorization gate that runs separately from kernel verification (the same transition can be kernel-passed and gate-denied; conformance vectors prove this)
• a refusal store that records the structure of every rejected proposal
• a CLI (intellagent) that exposes the runtime as a state machine: init → propose → transition → audit → refuse

This repo is NOT:

• A model. The runtime does not contain a neural network. Transformers serve as proposers under the kernel.
• A network service. The runtime is single-host with no outbound HTTP in the core path.
• A moral system. WiseOrder governs the form a claim must take to be admissible under its class; the substantive content of correct judgment lives in conduct artifacts (Class D) and is governed there.
• A claim of AGI. Every "Non-Goals" section in every doc says so explicitly.

---

2. Architecture

                  AI output
                      ↓
              ┌───────────────┐
              │   Wisernance  │   ← the kernel that governs what becomes an action (this repo)
              └───────────────┘
                      ↓
       ┌────────────────────────────┐
       │   Intellagent Runtime v0.1 │   ← state machine over the kernel (this repo)
       └────────────────────────────┘
                      ↓
       ┌────────────────────────────┐
       │     WIN       │   WISEATA  │   ← proof substrate: seal + verify, then inspect (separate repos)
       └────────────────────────────┘
                      ↓
                  execution systems

The kernel and runtime share a process; the runtime exposes the kernel through the intellagent CLI. There is no daemon, no service, no network listener.

Path	Purpose
SPEC.md	WiseOrder Protocol v0.1.0 canonical draft. Governs in case of conflict.
SPEC_LOCK_v0.1.md / SPEC_LOCK_v0.2.0.md	Frozen specifications per version.
STATUS-REGISTRY.md	Per-class status tokens and their meanings.
ARTIFACTS.md	JSON schema per artifact class.
CONFORMANCE.md	Conformance is class-scoped; vectors determine conformance, prose explains it.
IMPLEMENTATIONS.md	Registry of known implementations.
vectors/	Conformance test vectors (29 in vectors/v0.2.0/ + 23 in v0.1.0 baseline).
schemas/	JSON Schemas for vectors, manifests, fixtures, declarations.
chain/	The live v0.2.0 chain: genesis.win + 2 follow-up triples, CHAIN_VALID.
intellagent_runtime/	Kernel, gate, audit memory, refusal store, proposers, chain primitives, CLI.
tools/	Conformance + interop validators, pseudocode scanner, demo runner.
tests/	480 pytest tests covering protocol vectors, runtime semantics, chain integrity, transformer proposer, CLI, plus 7 crash-injection + 3 SIGKILL recovery tests.
interop/	Fixture manifests + interop checks tying implementations to vectors.
reports/	Committed conformance + interop reports (regenerated by make conformance / make interop).
rust_verifier/	First-party independent verifier track (Rust).
go_verifier/	First-party independent verifier track (Go).
docs/	Whitepapers, strategy, release packets, audits, runtime architecture.

---

3. Invariants

Core invariants enforced by tests and CI. Documented in spec and code; reproduced here at high level.

• Kernel ⟂ Gate separation. Kernel verifies structure; the authorization gate verifies policy. A transition can pass one and fail the other. See intellagent_runtime/kernel.py and intellagent_runtime/authorization.py.
• Audit memory is fail-closed. Loading the chain re-derives every entry's this_entry_sha256 and verifies every prev_entry_sha256 link. Any mismatch raises ChainCorrupt and the runtime refuses further operation.
• State is content-addressed. state_id = sha256(canonical_json(sorted_objects)). Loading state re-derives state_id and raises StateTampered on mismatch.
• apply_transition is crash-safe. Two-phase commit (stage → save → finalize) guarantees that after any crash, state and audit cannot diverge. Proven by 7 synthetic crash-injection tests + 3 real-process SIGKILL tests.
• TS-1 timestamp precision. Chain triples use microsecond precision; audit memory and conformance vectors use second precision. The two are intentionally NOT unified — chain filenames require sub-second uniqueness; audit/vectors require operator-readable consistency. Both formats are load-bearing.
• WiseDigest-3 / III parity with WOP. intellagent_runtime/iii.py is a byte-identical mirror of wop/src/wise/digest_v3.py. Any divergence breaks the cross-repository verifier compatibility claim.

---

4. Failure model

The runtime is fail-closed everywhere a sealed claim is at stake.

failure	response
Audit chain corrupt (link broken or hash mismatch)	ChainCorrupt raised; CLI exits 2; no further commits
State tampered (recomputed state_id differs from declared)	StateTampered raised; CLI exits 2
Process killed mid-apply_transition	startup reconcile_pending(state.audit_head_sha256) either finalizes the staged entry (state already references it) or discards it (state does not) — proven by SIGKILL tests
Malformed transition body	TransitionSchemaError raised at load time, not at kernel time — bad regime or path-traversal transition_id is rejected before kernel verification
Malformed policy JSON	PolicySchemaError raised — known-kind bodies with missing fields fail loudly, not silently
State / audit divergence after reconciliation	StateAuditDivergence raised; CLI refuses to proceed

See docs/REVIEWER_GUIDE.md for a 30-minute path to reproduce each one.

---

5. Quick start

git clone https://github.com/Wise-Est-Systems/wiseorder-protocol.git
cd wiseorder-protocol

python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Full local CI: documentation code standard + tests + protocol conformance + interop.
make ci

# Run the transformer-backed runtime demo (no real model needed).
python3 tools/demo_transformer_proposer.py

# Initialize a runtime, propose a transition, commit, audit.
pip install -e .
intellagent --help

---

6. Verification commands

make ci                  # the heavy pre-flight: pseudocode standard + tests + conformance + interop + verifiers
make test                # pytest only (480 tests, ~3s)
make chain-verify        # standalone chain integrity; prints head hash
make conformance         # regenerate + check vector pass rates
make interop             # regenerate + check fixture conformance
make canonicalization-check  # 10 canonical-form corpus entries
make rust-verifier-check # first-party Rust verifier — must agree on all 3 fingerprints
make go-verifier-check   # first-party Go verifier — must agree on all 3 fingerprints

Expected output of make chain-verify:

CHAIN_VALID count=3 head=5964497c48c877946e2c92d15e3116f5991c1d8a4c99dc7eadb477cec558dd81

Expected output of make test:

480 passed, 9 xfailed in ~3s

Frozen v0.1.0 fingerprints

A third party cloning this repo at the v0.1.0 tag and running the three commands below MUST observe these exact fingerprints. Divergence is a non-conformance.

Fingerprint	Required value
vectors_suite_sha256	sha256:6168d2075931baa98bba763f9cd537142554aae7b26f19e3e93a60de2ee5bb0f (33 vectors: 23 baseline + 10 adversarial)
manifests_suite_sha256	sha256:74eaaa62271f23172232bda43ba3340543e66c46ba1e2f5e6da18d3447a6ba29 (3 fixtures)
corpus_sha256	sha256:c95685bff48c15abc313c462908b05ac309250bc660c2d802ae51af5a8038b09 (10 corpus entries)

The v0.2.0 vector set under vectors/v0.2.0/ is published but not yet frozen; freezing happens at the next SPEC_LOCK cut.

---

7. CI status explanation

workflow	what it gates	matrix
tests	pytest under multiple Pythons and OSes	ubuntu / macos × py3.11 / py3.12
lint	ruff check + ruff format check on package + tools	ubuntu × py3.12
verify-chain	runs verify_chain over chain/; emits head hash to job summary	ubuntu × py3.12
conformance	full make ci (the heavy pre-flight: pseudocode + tests + conformance + interop + verifiers)	ubuntu × py3.12

Every push to main triggers all four. The verify-chain workflow has path filters so it runs only when chain/, chain.py, or iii.py change; the others run on every push.

---

8. Recovery guarantees

event	guarantee
Process killed mid-apply_transition	reconcile_pending at next startup recovers a consistent on-disk state. Proven by 3 real-process SIGKILL tests.
Mac power loss between stage and save	write_atomic guarantees current.json is either old or new, not partial. Staging file (if present) is orphan and discarded.
Mac power loss between save and finalize	current.json references the staged entry's hash; reconcile renames staging→final.
Independent verifier (rust/go) disagrees with python on the same input	Conformance failure. CI fails. Block release.
Cross-machine replay produces different bytes	A non-conformance. See docs/release/CROSS_MACHINE_REPLAY_REPORT.md for the determinism evidence.

There is no rollback of a sealed chain triple. The chain is an audit trail; withdrawing a prior statement is done by appending a new triple whose statement documents the withdrawal.

---

9. Operational philosophy

• The chain is the public-facing artifact, not the code. External verifiers re-derive consequence_proof and previous_action from a .win file alone. The repo is provenance for the verifier; it is not the source of truth.
• Fail closed at every seam. ChainCorrupt, StateTampered, TransitionSchemaError, PolicySchemaError, StateAuditDivergence are not warnings — they halt the runtime.
• No part of the protocol depends on a network. Verification works offline from a USB drop. See /Volumes/T7/2026-05-24/verify.sh for the canonical "drop the drive, prove the chain" example.
• Cross-language verifier parity is non-negotiable. Python, Rust, and Go must all agree on the head hash for any release tag. Single-language verification is single-source assurance.
• Vectors are the contract; code is one implementation. A second implementation that passes the vectors is a peer; one that does not is non-conformant.

---

10. Current limitations

• External-validation status: NOT_COMPLETE. The Rust and Go verifiers in this repo are FIRST_PARTY_INDEPENDENT_IMPLEMENTATION_TRACKS, not third-party validation. A real third-party verifier is required for EXTERNALLY_VALIDATED status; see docs/release/THIRD_PARTY_VERIFIER_BRIEF_v0.1.md.
• WiseDigest-3 is novel. It is not SHA-2, not BLAKE3, not a permutation function used in production crypto. The math is in wop/research/WiseDigest-3.md; cryptographic analysis is not yet complete. Production use as a security-critical primitive requires that analysis.
• Lint and mypy run with continue-on-error: true. The baselines are not clean. Drift is reported but does not block CI yet.
• No license file. pyproject.toml references one. Choosing a license is a deliberate decision the project owner has not yet made; the repo is currently effectively closed-source by default copyright law.
• No CHANGELOG.md yet. Release history is in git log --oneline; the changelog template is in docs/RELEASE_PROCESS.md.

---

11. Roadmap

1. Wire branch protection on GitHub (the policy in docs/BRANCH_PROTECTION.md is currently aspirational; the UI does not enforce it).
2. Pick a license.
3. Tighten ruff + mypy baselines; remove continue-on-error: true.
4. Freeze vectors/v0.2.0/ and cut SPEC_LOCK_v0.2.0 if it is not already (verify against chain/2026-05-23T071437_...win).
5. First external engineering scrutiny (see docs/release/EXTERNAL_REVIEW_PACKET_v0.1.md).
6. Tag v0.1.0 once #1, #2, and the audit packet are ready.
7. Cryptanalysis of WiseDigest-3 (deferred to wop; tracked there).

---

12. Security posture

boundary	posture
Network	No outbound HTTP from the core runtime. Optional LLM providers (OpenAI / Anthropic) are opt-in dependencies; the default DeterministicMockProvider requires no network.
Filesystem	intellagent writes only under --dir (default cwd) into intellagent_state/, intellagent_audit/, intellagent_refusals/, intellagent_objects/. No writes elsewhere.
Untrusted JSON input	All transition + policy JSON is validated by typed schemas at load time. Unknown fields are tolerated (forward-compatible); known-field type mismatches raise.
Subprocess execution	governed-run --execute runs proposed commands under tools/os_isolation_runtime sandbox-exec isolation. Sandbox failures are recorded in the manifest.
Authentication	None at the protocol layer. The CLI is a local tool. Third-party verifier authentication is out of scope here (the chain is content-addressed).

The threat model is: an attacker with file-system access to the runtime root could tamper with current.json, the audit chain, or the queue, and the next invocation would either reject the tamper (StateTampered / ChainCorrupt) or surface it in CI (verify-chain workflow). No invariant claims to defend against an attacker who can also modify the runtime's source code.

---

13. Release discipline

Documented in docs/RELEASE_PROCESS.md:

• Semver with explicit MAJOR triggers (new chain root or non-backwards-compatible vector change).
• Pre-release checklist: clean tree → make ci → make chain-verify → rust + go verifier parity → signed tag → T7 evidence snapshot.
• No retagging. No re-sealing the genesis after a public release.
• Rollback is by appending a withdrawal triple, not by editing or removing existing triples.

No releases tagged yet — this repo is v0.1.0-pre. The first tag will follow the items in the Roadmap.

---

Further reading

• docs/REVIEWER_GUIDE.md — 30-minute path for an external reviewer
• docs/INTEGRATION.md — how another system integrates
• docs/BRANCH_PROTECTION.md — operational policy on main
• docs/RELEASE_PROCESS.md — semver, checklist, rollback
• docs/release/EXTERNAL_REVIEW_PACKET_v0.1.md — what v0.1.0 is, what it isn't, what counts as successful external validation
• docs/release/REVIEW_QUICKSTART.md — commands only; reproduce CI from a cold clone
• docs/release/THIRD_PARTY_VERIFIER_BRIEF_v0.1.md — formal requirements for a third-party verifier submission
• CHANGELOG.md — release history