Verification Strategy
Map the project's real risks to gates. Watch each gate fail before calling it
executable, and wire it to block the promotion it protects. Accept no gate on
reading the code, a coverage number, or the builder's opinion.
When to use
- The project floor is absent, stale, or cannot catch a real defect.
- A high-risk initiative introduces preservation, platform, data, security,
concurrency, resource, or rollout risks the current battery does not
cover.
- Skip a bounded feature: its plan's evaluators and
test-driven-development own the local proof. A missing project floor is
separate scope.
Step 1: Map the risks before writing any gate
- Inventory what the project promises and what could break it: committed
behaviour, platforms and build modes, data and operations, commands and CI
that already run, existing tests, escaped defects.
- For a transformation, add its migration facts, invariants, and approved
deviations.
- Start the matrix from
assets/assurance-matrix.md before writing any
gate. Fill Risk inventory and Risk-to-gate matrix.
- Read
references/battery-catalogue.md before filling Catalogue disposition for every category the catalogue lists: covered, or an
accountable expiring approval plus a compensating gate. An N/A needs
evidence and an owner.
Step 2: Make each gate real
- Fill
Cadence and promotion map.
- A "manual" gate gets a procedure, evidence, an owner, and a promotion it
blocks.
- Calibrate each gate in isolation you are authorized to use: pin down what
the gate touches and your authority to mutate, recover, and clean up.
- Watch the whole cycle: green, then red on an introduced divergence, then
full restoration, then green. Keep raw results outside the state under
check.
- If calibrating for real is unsafe, use a known-bad fixture. With neither,
the gate is uncalibrated: record it
planned or blocked, with the
reason under Known gaps and deviations.
- Record each gate as
executable, planned, or blocked. Only
executable satisfies an entry condition.
- To implement one bounded gate, invoke
test-driven-development. For gate
work that takes more than one step, invoke writing-plans.
Step 3: Protect the battery from its own erosion
- Fill
Control-plane independence and Test-inventory audit.
- Keep the inventory, the validator, and the wiring outside the tests they
protect.
- If a mutable control plane would guard its own invocation, record its
promotion as
planned or blocked until external enforcement exists.
Step 4: Challenge, then hand over the decision
Fill assets/assurance-challenger.md with assets/assurance-report.md before
the challenge.
REQUIRED SUB-SKILL: invoke requesting-code-review with the filled
challenger prompt, before approval. Keep the state under check read-only.
Read the returned assurance section. If its Matrix version is missing or
mismatched, or its verdict is anything other than clear, approval is
blocked.
On a blocker, correct a successor, reverify, and rechallenge, inside
receiving-code-review's repair budget. When the budget is spent, or no
supported correction remains, the loop ends unconverged: return the
open blockers to the user and present nothing for approval.
Present the matrix:
Assurance matrix {{path}} — {{n}} executable, {{n}} planned, {{n}} blocked
Thresholds: {{list}}. Omissions: {{list}}. Blocking cadence: {{summary}}.
1. Approve the battery
2. Request changes
3. Reject the strategy
4. Cancel
Recommendation: {{option}} — {{one sentence}}.
Ask this and end the turn; clarifying-intent owns what closes it.
Approve nothing yourself. Authority to draft gates is not authority to
accept unseen risks, thresholds, omissions, or exceptions.
Keep the matrix file's status at proposed.
Gotchas
- Marking a catalogue category
N/A with no evidence or owner leaves the
exact silent gap the matrix exists to surface; it only relabels the gap
instead of closing it.
- A gate that runs correctly and shows green still blocks nothing if it is
wired to run only on demand. Being correct and sitting on a path no
promotion can bypass are two different facts.
Hard stops
Calling a control executable when it is not
- It is only planned, or only prose.
- It has never been run.
- It has never been watched fail.
Writing gate code before the matrix and its catalogue decisions (the
Catalogue disposition rows) exist, including when the gate is small enough
to stay inline.
A gate that cannot fail for the right reason
- A presence or change detector: it goes red without any behaviour being
protected.
- An aggregate green that hides a platform, a build mode, a shard, a test, or
a cadence that never ran.
- A wildcard match, a file's presence, a set or count comparison, a named
receipt (a per-case record the run itself emits), a success marker, or an
outer timeout, presented as proof that an exact case or action occurred.
Calling the battery established before removing, skipping, duplicating, or
hollowing a test turns the project command red under an independent guard, and
restoring it returns green.
Self-protection that protects nothing
- A validator living only inside the tree it discovers.
- A mutable command asked to guard its own invocation. Without external
wiring, that promotion stays
planned or blocked.
Taking authority you were not given
- Deferring the independent challenge until after approval.
- Granting yourself an exception.
- Approving work the decision owner has not seen.
- Writing lifecycle state into the matrix under check.
Lowering a threshold, deleting a test, or accepting a deviation to make the
matrix green.
1---2name: verification-strategy3description: "Designs or repairs a project's correctness battery: which checks catch which defects and when they run. Use when a project's assurance is absent, stale, or unfalsifiable before high-risk work, when risk changes, after an escaped defect or a hollow gate, or when the user asks how to test this project, says the tests don't catch anything, or asks what CI should run. Skip a bounded feature, and skip writing or running an already-defined gate."4metadata:5 source: https://skillmd.com/skills/augments-labs/verification-strategy6---78# Verification Strategy910Map the project's real risks to gates. Watch each gate fail before calling it11executable, and wire it to block the promotion it protects. Accept no gate on12reading the code, a coverage number, or the builder's opinion.1314## When to use1516- The project floor is absent, stale, or cannot catch a real defect.17- A high-risk initiative introduces preservation, platform, data, security,18 concurrency, resource, or rollout risks the current battery does not19 cover.20- **Skip** a bounded feature: its plan's evaluators and21 `test-driven-development` own the local proof. A missing project floor is22 separate scope.2324## Step 1: Map the risks before writing any gate25261. Inventory what the project promises and what could break it: committed27 behaviour, platforms and build modes, data and operations, commands and CI28 that already run, existing tests, escaped defects.292. For a transformation, add its migration facts, invariants, and approved30 deviations.313. Start the matrix from `assets/assurance-matrix.md` before writing any32 gate. Fill `Risk inventory` and `Risk-to-gate matrix`.334. Read `references/battery-catalogue.md` before filling `Catalogue34 disposition` for every category the catalogue lists: covered, or an35 accountable expiring approval plus a compensating gate. An N/A needs36 evidence and an owner.3738## Step 2: Make each gate real39401. Fill `Cadence and promotion map`.412. A "manual" gate gets a procedure, evidence, an owner, and a promotion it42 blocks.433. Calibrate each gate in isolation you are authorized to use: pin down what44 the gate touches and your authority to mutate, recover, and clean up.454. Watch the whole cycle: green, then red on an introduced divergence, then46 full restoration, then green. Keep raw results outside the state under47 check.485. If calibrating for real is unsafe, use a known-bad fixture. With neither,49 the gate is uncalibrated: record it `planned` or `blocked`, with the50 reason under `Known gaps and deviations`.516. Record each gate as `executable`, `planned`, or `blocked`. Only52 `executable` satisfies an entry condition.537. To implement one bounded gate, invoke `test-driven-development`. For gate54 work that takes more than one step, invoke `writing-plans`.5556## Step 3: Protect the battery from its own erosion57581. Fill `Control-plane independence` and `Test-inventory audit`.592. Keep the inventory, the validator, and the wiring *outside* the tests they60 protect.613. If a mutable control plane would guard its own invocation, record its62 promotion as `planned` or `blocked` until external enforcement exists.6364## Step 4: Challenge, then hand over the decision65661. Fill `assets/assurance-challenger.md` with `assets/assurance-report.md` before67 the challenge.68 **REQUIRED SUB-SKILL:** invoke `requesting-code-review` with the filled69 challenger prompt, before approval. Keep the state under check read-only.70 Read the returned assurance section. If its Matrix version is missing or71 mismatched, or its verdict is anything other than `clear`, approval is72 blocked.732. On a blocker, correct a successor, reverify, and rechallenge, inside74 `receiving-code-review`'s repair budget. When the budget is spent, or no75 supported correction remains, the loop ends `unconverged`: return the76 open blockers to the user and present nothing for approval.773. Present the matrix:7879 ```text80 Assurance matrix {{path}} — {{n}} executable, {{n}} planned, {{n}} blocked81 Thresholds: {{list}}. Omissions: {{list}}. Blocking cadence: {{summary}}.8283 1. Approve the battery84 2. Request changes85 3. Reject the strategy86 4. Cancel8788 Recommendation: {{option}} — {{one sentence}}.89 ```9091 Ask this and end the turn; `clarifying-intent` owns what closes it.924. Approve nothing yourself. Authority to draft gates is not authority to93 accept unseen risks, thresholds, omissions, or exceptions.945. Keep the matrix file's status at `proposed`.9596## Gotchas9798- Marking a catalogue category `N/A` with no evidence or owner leaves the99 exact silent gap the matrix exists to surface; it only relabels the gap100 instead of closing it.101- A gate that runs correctly and shows green still blocks nothing if it is102 wired to run only on demand. Being correct and sitting on a path no103 promotion can bypass are two different facts.104105## Hard stops106107**Calling a control executable when it is not**108109- It is only planned, or only prose.110- It has never been run.111- It has never been watched fail.112113**Writing gate code before the matrix and its catalogue decisions (the114`Catalogue disposition` rows) exist**, including when the gate is small enough115to stay inline.116117**A gate that cannot fail for the right reason**118119- A presence or change detector: it goes red without any behaviour being120 protected.121- An aggregate green that hides a platform, a build mode, a shard, a test, or122 a cadence that never ran.123- A wildcard match, a file's presence, a set or count comparison, a named124 receipt (a per-case record the run itself emits), a success marker, or an125 outer timeout, presented as proof that an exact case or action occurred.126127**Calling the battery established** before removing, skipping, duplicating, or128hollowing a test turns the project command red under an independent guard, and129restoring it returns green.130131**Self-protection that protects nothing**132133- A validator living only inside the tree it discovers.134- A mutable command asked to guard its own invocation. Without external135 wiring, that promotion stays `planned` or `blocked`.136137**Taking authority you were not given**138139- Deferring the independent challenge until after approval.140- Granting yourself an exception.141- Approving work the decision owner has not seen.142- Writing lifecycle state into the matrix under check.143144**Lowering a threshold, deleting a test, or accepting a deviation to make the145matrix green.**