t81-foundation

First Deterministic AI OS-Object Chain

This is the smallest current example of T81 acting like a governed AI OS-object substrate rather than only an AI runtime.

The canonical runnable example is:

bash examples/ai-and-inference/model-load-canonfs/run_assess_fixed_host_action.sh

Portable smoke path for the same chain:

./build/t81_ai_task_assess_fixed_composition_test ./build/t81

The shell script remains the human-readable demo. The compiled smoke path exercises the same object flow without relying on POSIX shell behavior.

For the current bounded family of reusable AI OS-object chains built on the same object model, see AI_OS_OBJECT_CHAIN_CATALOG.md.

What Happens

The current assess-fixed chain is:

1. a bounded governed AI task runs in the strict deterministic lane
2. the task stores a result artifact and a provenance artifact in CanonFS
3. a typed downstream record is created from that result and stored
4. a final bundle artifact is created and stored
5. the bundle becomes the canonical top-level object for the whole chain

This is the actual object flow:

assess-fixed task
  -> task result artifact
  -> provenance artifact
  -> downstream host-action record
  -> final bundle artifact

The Four Objects

• Task result artifact: the bounded AI result for the assessed input. This is an intermediate object.
• Provenance artifact: the retained execution and policy evidence for that task result. This is an intermediate object.
• Downstream record: the typed non-AI decision record created from the AI result. This is an intermediate object.
• Final bundle artifact: the CanonFS object that links the full chain together. This is the canonical top-level persisted object.

The bundle is canonical because it is the smallest single object that points to the whole completed chain: result, provenance, downstream record, and action identity.

Canonical Identity Invariant

For the currently validated assess-fixed, route-fixed, and classify-fixed chains, object identity at every layer is a function only of:

• task
• model
• policy
• input

For those chains, the following objects keep the same content-addressed identity and bytes for identical task/model/policy/input:

• task result artifact
• provenance artifact
• downstream record
• final bundle artifact

CanonFS root does not affect that identity. Execution location does not affect that identity.

Within this object model:

• result artifact: intermediate
• provenance artifact: intermediate
• downstream record: intermediate
• bundle: canonical top-level object

This claim is limited to the currently validated bounded family.

Key Helper Surfaces

The canonical example uses these typed surfaces:

• t81 ai task assess-fixed
• t81 ai task read-field
• t81 artifact write-store-record
• t81 artifact store-bundle
• t81 artifact read-field

These helpers do not form a general workflow engine. They only support the current narrow typed chain family.

What This Is

• one deterministic governed AI OS-object chain
• bounded
• policy-gated
• artifact-first
• CanonFS-backed

What This Is Not

• chat
• agents
• general inference serving
• workflow orchestration
• a general packaging engine

Core Claim

T81 does not just run AI here.

It produces governed, typed, persistent objects that other parts of the system can consume.

This site is open source. Improve this page.