Deterministic AI control DLX · DeciRepo Execution-bound governance
Estonia
Joined February 2015
- Tweets1.5K
- Following12
- Followers1.9K
- Likes2.3K
Pinned Tweet
A small milestone for DeciRepo.
The protocol baseline (v0.1) is now externally recomputable.
verification: PASS
rebuild: MATCH
conformance: PASS
Independent recomputation confirmed.
decirepo.com/pages/index.htm…
#Reproducibility #Verification #ProtocolDesign
IndrasNet retweeted
Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit. It's widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing). However, we found that the subscriptions are actually far more generous. (2/4)
Take a look at the latest article from our newsletter:
“Allow ≠ Valid: Why Systems Continue After Conditions Change”
linkedin.com/pulse/allow-val… через @LinkedIn
#DLX
#DeciRepo
#DecisionArtifacts
#VerificationSurface #IndependentRecomputation
Semantic drift is becoming an infrastructure problem.
Latest article explores why execution admissibility may become a new control layer for automated systems:
“Semantic Drift Is Becoming a Supply-Chain Problem”
linkedin.com/pulse/semantic-…
#DLX #SemanticSystems #ExecutionControl
Most systems control actions after execution already started.
DLX focuses on an earlier boundary: whether execution itself is admissible before commitment occurs.
DLX semantic pipeline uses deterministic validation and reproducibility checks.
#DLX
#SystemsDesign
Allowed ≠ Valid
Systems often assume prior approval remains valid at execution.
The workflow can be followed.
The logs can be complete.
And still: the action should never have been allowe
Because execution became binding under different conditions.
linkedin.com/pulse/allowed-v…
It was allowed.
It executed.
It shouldn’t have.
Allowed ≠ Valid
#DLX
#DeciRepo
#DecisionArtifacts
#VerificationSurface
#IndependentRecomputation
Allowed ≠ Valid
AI didn’t fail.
It answered. Cited. Sounded certain.
And was wrong.
→ >60% incorrect
→ fabricated links
→ wrong > refusal
(CJR)
Not generation.
Execution.
Outputs used without validity at commit.
linkedin.com/pulse/allowed-v…
Allowed ≠ Valid
Most systems don’t fail because rules are ignored.
They fail because execution runs on past approval.
DLX checks what is still valid at commit.
youtu.be/yMj_wne8fYw?si=fbjx… via @YouTube
#Systems
#Risk
#Architecture
#DLX
#Fintech
Allowed ≠ Valid.
Most systems assume that once something is approved, it is safe to execute.
That assumption often fails at the moment it matters most — commit.
Wrote a breakdown on execution-bound control and admissibility:
medium.com/p/allowed-valid-w…
#AIGovernance
#AISafety
Allowed ≠ Valid
Systems don’t fail because rules are missing.
They fail because execution is not bound to current state.
New piece:
linkedin.com/pulse/why-syste…
#ExecutionControl
#DecisionBoundary
#AI
#Risk
#Systems
Most AI systems don’t fail because they lack validation.
They fail because they can still execute when validation fails.
Detection ≠ control.
If execution is still possible, governance is observation.
#AI
#AIAgents
#AIEngineering
#securitynotpolitics
#ExecutionControl
Allowed ≠ Valid
Something can be permitted, executable, and still wrong under the conditions it was approved under.
That is how invalid states become real without ever being explicitly approved.
New piece: linkedin.com/pulse/allowed-v…
The system already knows the tool call failed.
And still exits cleanly.
That’s not robustness.
That’s failure being treated as success.
`invalid_tool_calls != no-call`
(langchain#33504)
#AIEngineering
#LLM
#AgentSystems
Control doesn’t fail when systems break.
It fails when execution proceeds
after approval has already stopped being true.
#DecisionBoundary #LLM
↓
Systems can follow all rules and still fail.
Allowed ≠ valid.
This is where most AI systems break.
linkedin.com/newsletters/all…
We asked an LLM to summarize DeciRepo.
It turned a bounded validation claim into broad enforcement.
That is the problem.
Drift rarely arrives as failure.
It arrives as wording drift:
“can verify”-“ensures”
“bounded surface”-“the system”
Control must live where execution begins.
The idea is already public.
FSM, admissibility, execution boundaries:
none of that is secret.
Anyone can describe the boundary.
The question is whether it still holds when execution begins.
That is where most systems fail.
#DeciRepo
#Risk
#AIGovernance
🤖 Made with AI
ALT decirepo.com
If execution does not depend on the check, governance remains descriptive.
A workflow can still drift.
A new position appears.
No one approved it.
DLX checks.
DeciRepo preserves.
medium.com/p/control-never-r…
#DeciRepo
#DecisionBoundary
#AIGovernance
#DecisionSystems
ALT decirepo.com