Debug and harden a model

What you will build

You will diagnose four different outcomes: a business rule that returns false, a formula error value, a strictness failure, and a rejected input write. Each needs a different repair.

1. Establish a clean baseline

Open the practice model above or the canonical Text And Quality example. Its parsing layer is:

A4 = REGEXMATCH(A1, A3)
A5 = REGEXEXTRACT(A1, /INV-(\d+)/)
A6 = REGEXREPLACE(A2, /owner:\w+/, "owner:shared")
A7 = WILDCARDMATCH(A1, glob"INV-*-001")

Checkpoint: A4 = TRUE, A5 = "2026", A6 = "priority:urgent owner:shared", and A7 = TRUE.

The downstream cells settle at B1 = :accepted, B2 = "current", B3 = 0, and B4 = "unassigned".

2. Treat rejection as data, not failure

Temporarily change only the wildcard in A7:

A7 = WILDCARDMATCH(A1, glob"PO-*-001")

Checkpoint: A7 = FALSE and B1 = :rejected; A4 remains TRUE, A5 remains "2026", and B2 remains "current".

Nothing crashed. The model computed a valid negative business result. Restore the original wildcard before continuing.

3. Observe and recover from an error value

Replace the guarded formula temporarily:

B3 = 10 / 0

Checkpoint: B3 = #DIV/0!.

An error is a value and propagates unless a formula handles it. Restore the intended boundary:

B3 = TRY 10 / 0 ELSE 0

Checkpoint: B3 = 0.

The fallback is policy. It should be visible in source rather than inferred from a blank display.

4. Expose a hidden coercion

Add this temporary binding:

C1 = "5" + 3

In the default loose mode, C1 = 8. Add coercion strict after the model header directives.

Checkpoint: C1 = #TYPE! because the arithmetic requires text-to-number coercion.

Repair the assumption explicitly:

C1 = VALUE("5") + 3

Checkpoint: C1 = 8 under coercion strict.

Remove the temporary binding and directive before moving to the canonical validation model.

5. Reject bad writes without losing good state

Open Cell Metadata & Validation. Its guarded inputs include:

input UnitPrice = 40
  VALIDATE BETWEEN 0 AND 500 MESSAGE "Unit price must be between $0 and $500."

input Category = "travel"
  VALIDATE IN ["travel", "software", "meals", "hardware"] MESSAGE "Pick an approved expense category."

input InvoiceRef = "INV-1042"
  VALIDATE STARTS WITH "INV-" MESSAGE "Invoice references begin with INV-."

Baseline checkpoint: Subtotal = 120, BudgetShare = 0.005, NetAmount = 114, and ApprovalStatus = "auto-approved".

Try writing 600 to UnitPrice, "office" to Category, and "1042" to InvoiceRef.

Checkpoint: each write returns its authored message and leaves UnitPrice = 40, Category = "travel", and InvoiceRef = "INV-1042".

A validation rejection is neither a formula error nor a compiler failure. The stored state remains valid.

6. Catch a bad default before use

Temporarily change the declaration to:

input Units = 60 VALIDATE 1..48

Checkpoint: diagnostics include GRID_VALIDATE_UNSATISFIABLE because the authored default violates its own contract.

Restore Units = 3.

Try it yourself

In Text And Quality, set A1 = "INV-2025-099" without changing either pattern.

Checkpoint: A4 = TRUE, A5 = "2025", A7 = FALSE, B1 = :rejected, and B2 = "legacy". Explain why the regex and wildcard answer different questions.

In Cell Metadata & Validation, make valid writes Units = 48 and UnitPrice = 120.

Checkpoint: Subtotal = 5,760, BudgetShare = 0.24, NetAmount = 5,472, and ApprovalStatus = "needs-approval".

You are done when

• You can tell a valid rejection from an error value and a diagnostic.
• Division-by-zero recovery is explicit in source.
• Strict mode no longer depends on silent numeric-text conversion.
• Invalid writes preserve the last valid input state.