hees.ai console 0.1.0

hees.ai console 0.1.0

Most AI tools show what a model said. hees.ai console shows what the model was allowed to say or do—and why.

AI models can produce fluent, well-structured answers without having authority to use particular evidence, make particular claims, or take particular actions. Hees.ai is a governed-intelligence runtime that turns evidence, reviewed memory, declared authority and policy into executable governance profiles for AI. Models may propose answers and actions; Hees determines whether those proposals are admissible and records the result.

hees.ai console is the terminal-first developer environment for creating, testing and inspecting those profiles. Its workflow begins before the prompt: developers explore evidence, curate candidate memory, inspect rights and review state, test package-declared actions and requirements, pressure-test proposals through Training by Committee, and inspect Spectrum decisions, selected memory, Content DNA and receipts.

GPT-5.6 proposes. Hees decides.

What you can explore in this release

Version 0.1.0 is a complete, bounded implementation profile built around an original fictional lesson-support package. It lets you:

• inspect source identity, fingerprints, exact source spans, language, provenance, review state and rights state;
• inspect reviewed memory atoms alongside known but non-admitted material;
• stage or unstage supported evidence and memory in a session-local candidate profile;
• test the candidate through the real Incan-authored Hees boundary without allowing interface state to replace active authority;
• inspect declared actions, response requirements, policy thresholds and active-versus-candidate state;
• follow an untrusted proposal through Hees-derived verifier targets, bounded committee observations and Hees-classified findings;
• inspect an admitted Spectrum result with selected memory, Content DNA and a receipt; and
• inspect exact rejection reasons when a model cites unknown evidence, invents an action, references unknown memory, or attempts to use non-admitted material.

Training by Committee does not vote a proposal into authority. Hees derives the exact targets, validates observation identity and complete coverage, classifies the resulting findings under profile-owned policy, and retains control of the terminal decision.

Run it without rebuilding Incan

Offline replay is the guaranteed, zero-credential path. Download the archive for a supported platform, verify its adjacent checksum, extract it, and run ./hees-console:

• hees-console-0.1.0-macos-aarch64.tar.gz
• hees-console-0.1.0-macos-x86_64.tar.gz
• hees-console-0.1.0-linux-x86_64.tar.gz

The five shipped scenarios include one admitted declared action plus unknown_action, unknown_evidence, unknown_memory, and memory_not_admitted rejections. Replay fixtures contain bound request, proposal and observation inputs; they never contain a stored Hees decision. The extracted executable reruns the compiled Hees validation, finding-classification, Spectrum, selected-memory, Content DNA and receipt path.

Each archive also includes Apache-2.0 project terms, NOTICE, a platform-specific third-party license report, RUNNING.txt, and a byte-identical release manifest. SHA256SUMS covers every archive, adjacent checksum and provenance manifest.

Optional live GPT-5.6 mode

Live mode requires an OpenAI API key and remains explicitly separate from offline replay. It sends a bounded structured-proposal request to GPT-5.6, asks role-bound evaluators to inspect only the exact targets derived by Hees, and passes the normalized untrusted inputs through the same terminal profile.

After the artifacts were frozen and audited, the tagged Apple Silicon binary completed a combined live admitted interaction with no retry: one strict GPT-5.6 proposal, three relation and three synthesis observations for six Hees-derived targets, six Hees-classified findings, selected memory, Content DNA and a receipt.

The same binary also rejected a live out-of-scope request for an apple-pie recipe. GPT-5.6 produced a safe, Lantern-grounded proposal, but the committee reported zero coverage of the actual request. Hees returned synthesis_incomplete, selected no memory, created no Content DNA, and issued a rejection receipt. This demonstrates an important distinction: an answer may be safe and evidence-supported while still being inadmissible because it does not fulfil the governed request.

The credential did not enter product output, fixtures or release artifacts. The adapter performs no automatic retries. Provider token usage and cost were not emitted by the product and are not claimed.

What this release establishes

This release establishes the working evidence-to-decision spine of hees.ai console:

evidence and review
        ↓
governed memory and profile authority
        ↓
untrusted proposal and committee observations
        ↓
compiled Incan-authored Hees boundary
        ↓
ADMITTED + selected memory + Content DNA + receipt
                         or
REJECTED + exact reason + receipt

An admitted result proves that this frozen profile accepted the exact package, request, proposal, support mappings, observation set, package-owned finding policy, selected-memory set, Content DNA construction and receipt projection required by console_profile_0_1. Semantic committee observations remain non-authoritative model evaluations; Hees binds their targets, validates them, applies package policy and owns the terminal outcome.

The permanent product direction extends this working profile into deeper semantic and factual verification, richer provenance and rights assurance, durable profile activation, broader Training by Committee workflows, and fuller Spectrum and Content DNA mechanics. Version 0.1.0 demonstrates real value without presenting those later assurance layers as complete.

macOS security note

The macOS archives are not Developer ID-signed and are not notarized. Linker ad-hoc signing may exist solely for local execution and conveys no publisher identity. Verify the adjacent checksum before extraction and follow the narrowly scoped Gatekeeper guidance bundled in RUNNING.txt.