What Is Actually Running
The public preview is honest about live computation versus packaged proof artifacts.
Live in this session
API routes compute canonical hashes, reproducibility checks, tamper mutations, rollback identity, ledger entries, and JSON exports.
Packaged verified receipts
Source snapshots, demo expert encodings, the change report, and the proof receipt are bundled with hashes and checked at runtime.
Not live in public preview
No autonomous legal interpretation, no live Python modality execution, no fresh proof generation, and no live LLM provider call.
Section D / Technical Review
How Triaxis Atum stays deterministic.
Triaxis Atum is designed for repeatable compliance artifacts, not stochastic answers dressed up with extra logging. This page explains the practical checks behind the demo in plain terms.
D.1 - What You Can Trust In This Demo
Stable artifact
Same encoding, same commitment
A committed expert encoding is serialized the same way every time, producing the same SHA-256 hash.
Change control
Deltas can be replayed
The regulatory delta composes into a post-update policy and rolls back to the original baseline hash.
Tamper evidence
Silent edits show up
Changing one encoded field produces a different commitment, which gives reviewers a concrete artifact to inspect.
D.2 - How To Check The Claim Yourself
Run the same source encoding repeatedly and confirm the same output hash.
Open the audit bundle and verify the baseline, delta, composed policy, and bundle hashes.
Compare the rollback hash with the original baseline hash.
Use the proof panel to check packaged proof receipts and keys against their recorded hashes.
D.3 - Why This Is Not A Stochastic Wrapper
No stochastic replay
The demo is not asking a probabilistic model to regenerate the same text. It works from committed artifacts and exact bytes.
No GPU dependency
The public workflow is artifact verification: canonical bytes, hashes, composition, rollback, and proof receipts. Those checks are lightweight.
No hidden prompt wrapper
The deterministic claim lives in the artifact commitments, not in a long prompt, seed setting, or post-hoc audit log.
D.4 - What The Public Preview Does Not Reveal
Production formula constants and implementation recipes.
Learned checkpoint parameters and scalar values.
Training or fitting procedures for production components.
Any internal detail that is not needed to verify the public demo claims.
D.5 - Evidence Files Behind The Screens
Source receipts
data/corpus/source_manifest.json
3 public source files or snapshots are stored with SHA-256 hashes.
Change report
data/reports/atum_compprog_delta_report.json
The packaged change report passes the demo bound check with 1.0049x tightness.
Proof receipt
data/vi_runtime/proof_0.bin
128 receipt bytes are packaged with recorded proof hash sha256:f1560cbd31b249e9b6a823d2bdad49166f73af8a84ca66afe539c847baa53185.
D.6 - What This Demo Is Not Claiming
- The demo does not issue legal advice. Humans create and review the regulatory encodings.
- The change-bound number describes a demo measurement, not legal or operational risk.
- The public preview verifies packaged proof receipts and keys against recorded hashes; fresh proof generation is reserved for a deeper deployment mode.
- The current encodings still need founder or qualified regulatory review before public launch.
- Production formulas, checkpoint parameters, and learned values are intentionally not disclosed.
D.7 - Common Technical Questions
Is Triaxis Atum replacing the regulatory expert?
No. The expert decides what the source means for the demo encoding. Triaxis Atum makes that expert artifact reproducible, composable, and checkable.
Is this just a seeded LLM?
No. Triaxis Atum does not ask a stochastic model to produce the same answer and then wrap it with logging. The demo commits deterministic artifacts directly. If the encoded artifact changes, the commitment changes.
Why not show every technical detail?
Because a public demo should demonstrate the claims without disclosing production formulas, learned values, or checkpoint internals. This page shows receipts, not the recipe.
Does this need GPUs to run?
No. The public demo's deterministic checks are lightweight artifact operations: canonical serialization, hashing, composition, rollback checks, and packaged proof receipt verification. This class of workflow does not require GPU inference to stay stable.
Verification Ledger0 entries
Session checks can be reviewed here on mobile and exported as JSON.
Ledger is empty. Run a verification to begin.