Hichem Benali

Building Asso, SYSTASYS and Asso Capital Engine. I start from the problem, not the field. Systems, products, research, operations. Evidence over claims.

[Hichem Benali @ace_prooflayer]

I kept trying to explain the work with one label. It kept breaking.

Asso is the continuity system.
SYSTASYS is the wider architecture.
Asso Capital Engine is the economic operating layer.

The public pieces are here so you do not have to take the description on trust:
https://t.co/obAE3khkBm

[Hichem Benali @ace_prooflayer]

Evidence, unknowns, and the test that could change my reading:
https://t.co/zMtnVFZjKE

Mine: https://t.co/RLXRkvexu3
Theirs: https://t.co/VIdScGKL1p

The June 13 repository is private today. The case says so explicitly.

[Hichem Benali @ace_prooflayer]

On June 13, Sendable? V0 appears in my Git history.

Today I discovered Sendable by Summit X Digital.

Different products, related pre-send territory.

I'm genuinely curious to compare timelines.

The evidence and the unknowns are public.

Receipts over claims.

[Hichem Benali @ace_prooflayer]

@49agents This is the part most people skip! The agent saying "it's fine" isn't enough for me, so I make it leave a trace. I can actually audit later. Nice to see we're pulling in the same direction, https://t.co/3Vem8PIMZT

[Hichem Benali @ace_prooflayer]

Your AI agent can open a PR, edit a workflow, and reach your secrets.
Untrusted PR text flows straight into its prompt.
The agent says it's fine. "It's fine" is not proof.
Agentic CI is the new attack surface.
No receipt, no passage.

[Hichem Benali @ace_prooflayer]

Free, Apache-2.0, runs locally:

https://t.co/3Vem8PIMZT

[Hichem Benali @ace_prooflayer]

New: one command for the whole repo.
ace-receipts report gives a shareable governance posture: which workflows use AI, what's risky, what's allowed to ship.
The document you hand a security lead, a client, or an auditor under the EU AI Act.
Receipts over claims.

[Hichem Benali @ace_prooflayer]

I just published a public proof notebook on Agent Decision Receipts.
https://t.co/ZDQo2WrMH2
AI agents should not act in the fog !
When software acts for people, the action should leave a readable trace:
intent->permission-> evidence-> decision-> closeout.

[Hichem Benali @ace_prooflayer]

@GoldilocksOrbit exactly👌 Two states force the split you describe: block everything (paranoia) or trust everything (apathy). The receipt is the missing third: here is what was seen, here is what is missing, a human decides. That middle step is the whole product!

[Hichem Benali @ace_prooflayer]

AI agents now edit PRs, touch workflows, and run inside CI.

Output is cheap. Trust is not.

I built ACE Receipts: a tiny local CLI + GitHub Action that scans AI-agent GitHub workflows and AI-generated diffs, flags proof/risk/permission gaps, and writes receipts for human review.

npx ace-receipts demo
npx ace-receipts scan-workflows

No cloud. No API key. No model call.
It helps review. It does not replace human review.

No receipt, no passage.

https://t.co/zZmUUnO0ha

[Hichem Benali @ace_prooflayer]

What scan-workflows catches on a booby-trapped repo: pull_request_target next to agent markers, PR body flowing into a prompt, secrets in reach, unpinned actions. Deterministic rules, line numbers included.

[Hichem Benali @ace_prooflayer]

@MSFTnews Unified platform governance works until the agent acts outside the platform. The portable question: does each action pass a pre-execution gate with a signed allow or deny record? That artifact survives vendor boundaries.

[Hichem Benali @ace_prooflayer]

@jeffsutherland Signed grants as the primitive is right. The test is the deny path: when a grant check fails, does the system emit a signed receipt of the refusal? Allow logs prove activity. Refusal receipts prove the boundary held.

[Hichem Benali @ace_prooflayer]

Klarna: $2.3M in unauthorized refunds from one agent.
The financial limits were in the system prompt.
Not in the tool constraints.
An instruction is a hope. A hard constraint is a receipt.

Receipts over claims.

https://t.co/ufPV0Xw1IU

[Hichem Benali @ace_prooflayer]

Agreed. The refusal receipt must be signed, carry the firing policy + agent context seen, and prove zero side-effect before anything runs. That’s the bar.

See the refused cases in asso-lab/examples/action-receipts (forbidden_write.refused.json and others). Looking forward to the replayable example.

@ai_zona

[Hichem Benali @ace_prooflayer]

Uncontrolled agent autonomy fails. Single unadmitted action, cascading system compromise. Agent proposes, kernel disposes. Admissibility gates every action, enforcing refusal. Syll system: execution logs, keyframes for human approval.

https://t.co/ReyYvz2yfw

R.O.C : Receipts Over Claims.

[Hichem Benali @ace_prooflayer]

Agent misbehavior risks system integrity.

The kernel must hold the deny-path. Agent actions propose, kernel disposes.

"Closed-by-default" execution prevents unauthorized access.
H-CSC enables "typed safe abort" for commitments.
This ensures bounded autonomy.

https://t.co/ReyYvz2yfw

Receipts over claims.

[Hichem Benali @ace_prooflayer]

In a market saturated by the promise of “Autonomous Agents”,

Real differentiation may come from the ability to produce agents that are auditable, governable, and compatible with high-responsibility environments.

https://t.co/ZrMkMYgRTp

[Hichem Benali @ace_prooflayer]

@killix Exactly. In practice the gate checks the call against the active mandate before anything executes. If it fails, the action is blocked and a receipt is created with the reason, timestamp and outcome. That receipt is what makes the whole decision traceable afterwards.