Hichem Benali Cabinet of Systems SYSTASYS ↗

Curated public work

A cabinet of
systems, proof & judgment.

Selected objects from work across software, operations and research. Enough to inspect the construction. Not the engine behind it.

CONSTRAINT

01 / Tool

ACE Receipts — The Gate

public source ↗

Output is cheap. Trust is not.

A public CLI and GitHub Action scans AI-assisted workflows and diffs for proof, risk, permission and missing evidence. The specimen below carries the same decision grammar, without touching your repository.

Specimen

ACE RECEIPTHOLD

Not yet. Bring receipts.

Proof
partial
Risk
high
Permission
human review
Public artifact monkidy/ace-receipts · examples/receipts/hold.mdextract · summary, findings and next omitted

ACE Receipt: HOLD

Not yet. Bring receipts.

Verdict
hold
Gate
fail
Proof
partial
Risk
medium
Permission
human_review_required
Closeout
pending

Command ace-receipts demo

Missing receipts

  • explicit permissions block
  • SHA pin receipt for third party actions
  • receipt policy (.ace/policy.yml) or an ace-receipts gate in the workflow

No receipt, no passage.

Reproduced unchanged from the public repository. This is the tool’s own output format, not a mock-up of it. The specimen above lets you move the same three conditions.

03 / Product

Sendable? — One last verdict.

public product ↗
Draft

“I’m sending the updated proposal tonight. I removed the optional scope and included the internal estimate so we can move faster.”

The same message. Three audiences. The verdict is not about grammar.
Pre-send receipt REVISE

Useful message, but the internal estimate should not leave the team context.

Remove the internal figure before sending.

04 / Observer surface

Asso Lab — Receipt signed by code.

public proof surface ↗
Public artifact monkidy/asso-lab · receipts/2026-06-17-receipt.jsoncomplete file 
{
  "receipt_id": "393911c3-2c62-4de0-8ae1-cbdb40804b88",
  "date_utc": "2026-06-17T10:41:51.567112Z",
  "sources": [
    "https://simonwillison.net/atom/everything/",
    "https://arxiv.org/rss/cs.MA",
    "https://arxiv.org/rss/cs.CR",
    "https://arxiv.org/rss/cs.AI",
    "https://www.anthropic.com/news"
  ],
  "model": "gemini-2.5-flash",
  "content_hash": "5e73ba03104d66e99518bbc904ccdbfc571ff01178ce0df6bf62fbca766efc01",
  "operator": "hichem",
  "confidence": null,
  "status": "PUBLISHED"
}
One day’s receipt, reproduced unchanged. The sources are public feeds. The hash makes the published artifact checkable after the fact. confidence is null because nothing measured it — an empty field is more honest than an invented number.

Publish the brief. Publish the receipt. Keep the boundary visible.

Asso Lab is a public observer surface: bounded briefs, code-generated receipts, hashes, timestamps, and worked examples of allowed, refused and human-review actions.

What this does not prove

A published receipt proves that an artifact existed in this exact form at this time. It does not prove the private runtime, production safety, or any authority to act.

05 / Operations

AI Ops SOP Pack — Recover without pretending.

public pack ↗

When context is lost — a crash, an interrupted review, a session that ended mid-decision — the tempting move is to reconstruct confidence from memory. The pack replaces that with re-read evidence and a guard on the object under review.

  1. 01context lost
  2. 02cold evidence
  3. 03head guard
  4. 04handoff
  5. 05human gate
Public artifact monkidy/ai-ops-sop-pack · templates/pr_final_audit_checklist.mdextract · head guard, boundary review, decision

Completing this checklist records review evidence only. It does not approve or perform a merge, and it does not grant merge authority by itself.

Head Guard

  • observed PR head commit:
  • expected PR head commit:
  • match confirmed:
  • merge command or action will preserve head guard:

Stop if the PR head changed and the new head has not been reviewed.

Boundary Review

  • no live runtime opened:
  • no provider call introduced:
  • no secrets introduced:
  • no daemon / watcher / background process introduced:
  • no memory write introduced:
  • no ledger write introduced:
  • no permission-to-act introduced:
  • no trading / order routing / leverage / portfolio authority introduced:

Decision

Choose exactly one:

  • PASS
  • REPAIR_REQUIRED
  • STOP_RISK
An extract from the published checklist, reproduced unchanged. The blank fields are the point: the document is designed to be filled in by a person, and it says in its own opening line that filling it in is not permission.

06 / Governance

Two receipts. One difference.

public examples ↗

A refusal is only trustworthy if it leaves the same kind of trace as an action. These two published examples share a boundary, an agent and a target file. One reads it. One tries to write to it.

Public artifact bounded_file_read.allowed.jsonextract · selected fields shown; full JSON at source 
{
  "schema": "ace.action_receipt.v0",
  "status": "allowed",
  "decision": "allowed_read_only",
  "requested_action": {
    "type": "file_read",
    "purpose": "summarize support context"
  },
  "boundary": {
    "mode": "read_only",
    "writes_allowed": false
  },
  "refusals": [],
  "human_review_required": false
}
Public artifact forbidden_write.refused.jsonextract · selected fields shown; full JSON at source 
{
  "schema": "ace.action_receipt.v0",
  "status": "refused",
  "decision": "refused_forbidden_write",
  "requested_action": {
    "type": "file_write",
    "purpose": "update customer record"
  },
  "boundary": {
    "mode": "read_only",
    "writes_allowed": false
  },
  "refusals": [
    {
      "reason": "write_not_allowed_inside_read_only_boundary",
      "required_next_step": "human_review_or_new_explicit_boundary"
    }
  ],
  "human_review_required": true
}

Both files carry the note Static public example. Not runtime proof. in their evidence block. Each plate states what it omits; nothing shown has been altered.

07 / Reasoning

Hypothesis Fan — Let evidence remove.

Specimen

One observation. Three explanations. One piece of evidence at a time.

Alternatives stay visible until a test gives a reason to remove one. The question this sharpens is Bateson’s: what is the pattern that connects these failures?

ObservationA workflow becomes unreliable as load grows.
H1capacity

Resource pressure is the dominant cause.

falsifier: failures persist under low load.
H2coordination

Handoffs introduce inconsistent state.

falsifier: failures occur without handoffs.
H3memory

State is reconstructed from incomplete context.

falsifier: explicit state does not improve recurrence.

08 / System

SYSTASYS — Continuity, not spectacle.

Observation Signal Hypotheses Test Decision Action Trace Memory

Memory is useful only if it changes the next observation.

The gold chord is the return path. Everything else is ordinary causal order; the return is what makes it a system rather than a checklist.

station / idle Eight stations, then a return.

Run one pass to move through them, or read them on the diagram.

The full room ↗

09 / Working laws

A manifest made of constraints.

Working law

01Keep more than one explanation alive.

Premature certainty is expensive. Let evidence eliminate.

Hold → test → remove.
02Reality decides.

Models are useful only while they survive contact with what happens.

Model ≠ field.
03Observation without memory is waste.

A system should remember enough to make the next observation better.

Trace → memory → changed observation.
04Evidence should change the next action.

A receipt that changes nothing is documentation, not control.

Proof must have consequence.
05Authority should remain visible and boring.

Power can live around the core. Permission boundaries should stay simple enough to inspect.

Power outside. Clarity inside.
06Continuity matters more than spectacle.

The useful system is the one that still makes sense after the demo ends.

After the effect, what remains?

What stays outside the aperture ↗