Test Planning & Governance

Model the drivers rather than adding a flat percentage: expected defect arrival times re-test cost per defect, divided by the measured fraction of days the build and environment are usable. Keep the buffer named, visible and tied to stated assumptions.

On the requirement side it looks unverified though it was tested, so the work is repeated. On the case side nobody can justify the case, so it is never pruned. Only the first has a forcing moment.

It overstates because shared components pull in every requirement that reaches them, changed behaviour or not. It understates because real coupling — shared data, ordering, implicit contracts — was never written down as a link at all.

Use it as a floor, never a ceiling. What stale links include still names relationships that were genuinely real; what they omit proves nothing, because the record stopped following the product two releases ago.

The link still exists, so the record reads the criterion as verified while nothing runs against it. An unlinked criterion is visible and gets an owner; a permanently skipped case hides the same hole behind a joined-up chain.

Anything still stored can be recovered: which change addressed which requirement, which suite ran against which build, and what it produced. What cannot be rebuilt is what nobody recorded — intent, judgement, and contemporaneity itself.

Generate it wherever the links already live on the work itself: a generated view can only show what the work records, and costs nothing to reproduce. An authored document shows more, but is only as true as its last edit.

The person producing the thing being linked creates it, at the moment they create it: the case author when writing the case, the engineer when opening the change, whoever files a defect when filing it. No separate scribe pass afterwards.

Because that figure divides by the code that exists. Behaviour nobody implemented contributes no code, so it cannot appear as unexecuted code — cutting work can even raise it. Only the requirement-side denominator holds the missing behaviour.

Hand over evidence and options, not a verdict: what was covered, which criteria are unmet and by how much, what is unverified, and realistic choices. The accountable owner signs.

Sort the existing checks by what makes them safe to run constantly: speed, determinism, self-diagnosis and freedom from scarce resources. Those become standing assertions evaluated on every change and against the running system; the rest still needs a release boundary, and saying which is which is the decision.

Replace the ratio with statements a reader can act on: what was exercised and what was not, what the cycle learned about each area, where the remaining uncertainty sits, and how much of that is measurement rather than product. Keep the percentage as detail, never as the headline.

Work backwards from the feedback time each decision can tolerate, then place every signal in the earliest tier that can produce it reliably inside that tier's budget. Every tier needs an owner, a blocking action and a written claim about what green means.

Make it a decision, not an omission: name the failure being accepted, score it on the same scale as what you did verify, offer a cheaper containment, and have the consequence-owner accept it with a review date.

Shift a check left when its findings survive the move and the artefact it inspects is stable enough to be worth inspecting. Emergent behaviour — real load, real data volumes, real third-party systems, real users — produces signals that only exist later, and no amount of early review substitutes for them.

Decide from where the obligated behaviour actually reaches, not from the module diagram. One standard everywhere is simple but taxes work that never needed it and decays into form-filling; two are cheaper until the boundary erodes unnoticed.

Test point analysis is a formula-based method that sizes testing from a counted functional size of the system, adjusted for quality characteristics, risk and environment factors. It is rare because almost nobody maintains a functional size count to feed it.

Attach it at the acceptance-criterion grain in most products. A whole-requirement link is quick to write and too coarse to answer anything useful a year later; a per-scenario link answers precisely but multiplies rows faster than any team keeps them current.

Suspension criteria are pre-agreed conditions for stopping a test cycle in flight — a build too broken to yield information. Resumption criteria state what must be true to restart, including which earlier results the interruption invalidated.

The quadrants classify checks on two axes: business-facing versus technology-facing, and supporting the team versus critiquing the product. Plot the planned checks on the grid, and an empty cell shows a kind of evidence the release does not have.

Agree a written, machine-checkable description of the interface first — request and response shapes, field meanings, error cases — then have each side check itself against that same description in its own build: the provider that it can produce those responses, the consumer that it works against a stand-in generated from them.

Re-point each affected link deliberately: decide per case which of the two new requirements it now verifies, sometimes both, sometimes neither cleanly. Keep the retired identifier resolvable so older results and defect records still lead somewhere.

Confidence comes from observable properties of the walk: how recently the links in that area were maintained, how much of the set came from links rather than your own reading, and whether earlier sets caught what actually broke.

It measures the link set, not the system. With links entered by hand the findings track how well people keep those links current, so a clean result means nobody has recorded a gap, not that everything is verified.

A test policy is a short organisation-level statement of quality objectives and non-negotiables every team must meet - the frame a strategy obeys. It earns its cost only with several teams to align, or an external obligation.