RSI TECH
Contact

Interactive system / Workflow X-Ray

Watch an AI workflow make, and refuse to make, a decision.

Inspect evidence extraction, simulated tool lookups, deterministic rules, evaluation, fallback and human ownership. Then break one boundary on purpose.

  1. 01Synthetic data
  2. 02Recorded reference first
  3. 03Simulated business tools
  4. 04No external business action
Open the reference run

SCENARIO / REFUND EXCEPTION REVIEW

€840 refund request, 16 days after purchase

Policy permits review. Automatic approval stops at €500.

STATUS
Completed
OUTCOME
Review required
OWNER
Finance
MODEL CALLS
2
TOOL CALLS
4
LATENCY
450 ms
EST. COST
$0.0060

Recorded reference

SELECTED STAGE / MODEL

Evidence

7 EVENTS / 60 MS

Turns the synthetic request into validated, typed evidence. Model output remains untrusted until validation passes. Current state: passed.

SANITIZED SOURCE TEXT

orderId: A-1842
“order A-1842”
requestedAmount: 840
“full €840 refunded”
reasonCode: integration_mismatch
“did not meet our integration requirements”
ACTOR
MODEL
STATUS
passed
CALLS
1
COST
$0.0030

DECISION BOUNDARY

Review required

Owner: Finance. External action: none. The workflow can prepare evidence and a handoff, but it cannot execute a refund.

Evaluation checks
6/6 passed
Policy violations
0
Fallbacks
0
MODEL / TOOL / RULE / CHECK / CONTROL / HUMANCHECKING LIVE CONFIGURATION…
Recorded reference run loaded.

CONTROLLED FAILURE

Break one boundary on purpose.

CRM timeout

Two simulated lookups time out. Cached context remains incomplete.

Expected: action blocked, Operations owns review

Schema violation

The first typed extraction is invalid. One bounded repair is attempted.

Expected: recovered, Finance still owns review

Prompt injection attempt

Instruction-like content appears inside synthetic customer evidence.

Expected: warning, deterministic policy unchanged

Live replay is temporarily unavailable. Every recorded failure path remains inspectable.

Baseline / controlled failure

The difference is the product.

A successful path proves little on its own. Compare makes changed stages, rules, retries, fallback, ownership and the unchanged no-action boundary explicit.

Real / simulated

A demonstration with an honest boundary.

01

Recorded truth

Typed extraction, deterministic validation, public trace events and measured reference timing expose the full decision path without requiring infrastructure.

02

Simulated boundary

The customer, order, policy, CRM responses, refund history, timeouts and every business-side tool response are synthetic.

03

Never executed

No refund, email, CRM write, customer message, account mutation or other external business action is performed.

Engineering pattern

Put probabilistic work inside deterministic authority.

  1. 01

    Bound the input

    Fixed scenarios and typed evidence keep visitor data and arbitrary instructions outside the execution boundary.

  2. 02

    Separate actors

    Model, tool, rule, check, control and human events show exactly where probabilistic behavior ends.

  3. 03

    Make failure observable

    A structured trace records retries, validation, fallback, ownership and why action was blocked.

  4. 04

    Authorize nothing by default

    The model proposes. Deterministic rules and a human owner decide. This reference executes no external action.

Apply the pattern

Your workflow probably fails differently.

Map the decision, data sources, tools, constraints, evaluations and expensive failure modes. I'll use that context to frame the smallest useful next step.

See relevant work →