Where does UMA-AIF belong?
Assessed against the two existing repositories.
kaysentinel — no fit
kaysentinel is a Rust protocol-integrity runtime for blockchain state execution: it extracts canonical state mutations from Geth/Reth traces into a Structural Sufficient Representation, commits them with domain-separated BLAKE3, and checks cross-client consensus at the mutation level.
Different language, different domain, different threat model. The only shared vocabulary is the word "consensus", and it means something else in each — cross- client agreement on state transitions there, agreement between reference model subspaces here. Nothing to merge.
ames — adjacent, but deliberately closed to this class of signal
ames is the near miss, and it is worth being precise about why it is still a miss.
AMES is a nine-layer containment framework: L1–L4 prevent an agent from leaving its execution environment, L5–L7 detect, L8 commits evidence to a hash-linked BLAKE3 ledger, and L9 terminates. It is Python, it is about AI agent safety, and L6 is literally called "Cognitive Drift". On the surface UMA-AIF looks like it wants to be an L6 implementation.
Three properties of AMES's own design say otherwise.
1. AMES's evidence/testimony split puts UMA-AIF on the wrong side of the line. The L8 ledger records host-observed evidence only — VM lifecycle events, image measurements, broker decisions. Guest-reported claims are quarantined into a reserved actor-ID band and excluded from evaluation. UMA-AIF reads activation tensors from inside the model the agent is running. In AMES's taxonomy that is testimony, not evidence: it is produced by the thing under observation, and a model that could manipulate its own activations to evade the auditor is exactly the adversary the framework assumes. Feeding UMA-AIF output into L8 as evidence would break the invariant the ledger exists to maintain.
2. L6 is advisory by construction. AMES structurally prevents cognitive drift from triggering a kill — there is no parameter on the L9 path that accepts a drift finding. That is a deliberate choice, and a good one: drift is a soft signal and L9 is an irreversible action. A statistical activation auditor is a softer signal still. Merging UMA-AIF into L6 would add code that, by design, cannot affect any outcome.
3. The deployment surfaces do not overlap. AMES contains an agent process inside an ephemeral microVM and watches it from the host. It does not need — and mostly cannot have — access to model weights or forward-pass internals, because the agents it contains typically call a model over an API that terminates outside the VM. UMA-AIF requires hooks into the transformer itself. Different side of the trust boundary, and a torch dependency that a containment runtime targeting Rust for production should not inherit.
AMES also has no plugin or detector interface. Its architecture is closed on purpose, in service of "structural impossibility over monitoring". UMA-AIF is monitoring, and probabilistic monitoring at that.
Recommendation: a separate repository, cross-referenced
The two are complements, and the complementarity is worth stating explicitly in both READMEs because it is the interesting part:
| AMES | UMA-AIF | |
|---|---|---|
| Vantage point | Host, outside the guest | Inside the forward pass |
| Object of study | The agent process | The model's internal representation |
| Evidence class | Host-observed | Guest-internal (testimony) |
| Guarantee style | Structural impossibility | Statistical anomaly detection |
| Failure mode | Availability (a kill that shouldn't fire) | Calibration (a score that means nothing) |
| Enforcement | L9 terminates | Advisory; projection at most |
AMES answers "did the agent leave the box". UMA-AIF asks "did the computation inside the box drift from its references". Neither subsumes the other, and each is weakened by pretending to be the other.
The one integration worth building
If a link is wanted later, the clean seam is UMA-AIF as an evidence source for an L8-style ledger, entered in the testimony band, never in the evidence band.
Concretely: emit a per-step audit record — seed, sampled layer indices, raw metrics, calibrated scores, calibrator fingerprint — and hash-chain it with the same domain-separated BLAKE3 scheme. That gives a tamper-evident, replayable audit trail with no claim to being host-observed and no path to L9.
Everything that record needs is already deterministic: LayerSampler and
EnvelopeNoiseInjector are seeded, ScoreCalibrator.state_dict() serialises the
baseline, and test_audit_step_is_reproducible_from_seed asserts a step can be
re-derived bit-for-bit from its inputs. That was designed in for this reason.
This is not implemented. It is a note about where the seam is, so that if the two projects are ever joined it happens at the boundary that respects both architectures rather than by dropping a torch dependency into a containment runtime.