Joined February 2015
- Tweets 1,454
- Following 12
- Followers 1,874
- Likes 2,353
166 Photos and videos
Pinned Tweet
Mar 9
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
3
483
May 19
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
1
21
May 11
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
1
54
May 11
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
34
May 4
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…
2
34
Apr 28
It was allowed.
It executed.
It shouldn’t have.
Allowed ≠ Valid
#DLX
#DeciRepo
#DecisionArtifacts
#VerificationSurface
#IndependentRecomputation
1
62
Apr 27
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…
1
41
Apr 21
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
1
131
Apr 15
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
1
72
Apr 13
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
61
Apr 13
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
1
1
43
Apr 6
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…
1
58
Apr 1
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
1
3
58
Apr 1
Control doesn’t fail when systems break.
It fails when execution proceeds
after approval has already stopped being true.
#DecisionBoundary #LLM
↓
1
49
Mar 31
Systems can follow all rules and still fail.
Allowed ≠ valid.
This is where most AI systems break.
linkedin.com/newsletters/all…
1
58
Mar 30
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.
1
78
Mar 30
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
ALT decirepo.com
1
1
79
Mar 29
We let an LLM reconstruct our system from public materials.
It produced a stronger story than the sources supported.
Not communication.
Control failure.
Models remove conditions.
reviewed source != generated claim
Without a boundary check,
governance is just description.
1
3
77
Mar 29
A reviewed message said: no acceptance of liability, no settlement.
The version about to be sent implied settlement anyway.
That is not wording drift.
That is a system that cannot hold its own position.
1
3
66
Mar 28
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
1
3
87
Mar 27
The wording moves first.
That is how a reviewed message becomes a committed position no one actually approved.
DeciRepo asks a narrower question:
Can one workflow surface cross an approval boundary without authorization?
If yes, the boundary failed.
#DeciRepo #BoundaryCheck
1
2
76