Reference
Templates and checklists: hypothesis stress-testing
Part of the Hypothesis Stress Testing skill · loaded on demand from SKILL.md
Copy-ready structures for each stage. Placeholders in angle brackets are allowed here and nowhere in a client deliverable.
1. Day-one answer record (stage 1)
## Day-one answer: <engagement>
Question (as the decision-maker would ask it): <one sentence>
Day-one answer: <claim with at least one number or direction>
Held by: <names or roles>; strength: <strong / moderate / weak>
Dissenting views in the team: <name: alternative answer>
Killing observation for the day-one answer: if <observable>, this answer is wrong.
Assumptions made in the absence of input: <list, each with what changes if it is wrong>
2. Hypothesis statement worksheet (stage 2)
One per hypothesis. All six elements must be filled before the hypothesis enters the matrix.
### H<n>: <short label>
Subject: <the specific thing claimed to have changed or to be true>
Direction and magnitude: <how much, in the decision-maker's unit>
Mechanism: <the causal path; what else it predicts should be visible>
Time window: <when it started or applies>
Scope: <where it should appear and where it should not>
Killing observation: dies if <specific measurement and threshold>
Champion: <who would defend this hypothesis>
Origin: <issue-tree branch / perspective / artefact-null>
Hypothesis set table:
| ID | Hypothesis (one sentence in falsifiable form) | Killing observation | Champion | Origin |
|---|---|---|---|---|
| H1 | | Dies if | | |
| H2 | | Dies if | | |
| H3 | | Dies if | | |
| H4 | | Dies if | | |
| H5 (artefact or null) | | Dies if | | |
Set checks before proceeding:
- Four to six hypotheses.
- At least one artefact, timing or null hypothesis.
- Each has a champion.
- Close to mutually exclusive on the primary cause; "none of the above" is not comfortable.
3. Evidence ledger (stage 3)
| ID | Evidence statement (one sentence, with number where available) | Source | Date | Reliability | Type |
|---|---|---|---|---|---|
| E1 | | | | audited / measured / estimate / single interview | measurement |
| E2 | | | | | testimony |
| E3 | Absence: <what was expected and not seen> | | | | absence |
| A1 | Team assumption: <assumption underpinning which rows> | | | untested | assumption |
Type is one of: measurement, testimony, absence, assumption, deduction. Testimony and assumption rows are expected to rate Low diagnosticity; they are listed so that they are visible, not so that they count.
4. ACH matrix (stage 4)
## ACH matrix: <question>
Cell question: if this hypothesis were true, how likely would I be to see this evidence?
Scale: CC strongly consistent | C consistent | N no bearing | I inconsistent | II strongly inconsistent
| ID | Evidence (short) | H1 | H2 | H3 | H4 | H5 | Diagnosticity |
|---|---|---|---|---|---|---|---|
| E1 | | | | | | | High / Medium / Low |
| E2 | | | | | | | |
| ... | | | | | | | |
Notes on I and II cells (mandatory, one sentence each):
- E<n>/H<m> (<I or II>): <why, including magnitude where quantitative>
Diagnosticity rule: High if any cell is II, or the row holds both CC and I; Medium if the row holds at least one I, or a CC alongside N cells; Low if every cell is C or N.
5. Refinement log (stage 5)
| Change | Item | Reason | Killing observation retained? |
|---|---|---|---|
| Removed from inference | E<n> | Low diagnosticity | n/a |
| Split | H<n> into H<n>a "<whole>" and H<n>b "<part>" | <row that refuted the whole> | Yes: H<n>b dies if <threshold> |
| Reworded | H<n> | <reason> | Yes / No (if No, reject the change) |
| Added | H<n+1> | Suggested by E<n> | Dies if |
| Merged | H<a> and H<b> | Identical ratings on every row | Yes |
| Considered, parked | <candidate> | <sized; inside residual> | n/a |
6. Elimination table and decomposition (stage 6)
| Hypothesis | Inconsistent rows | Reliability of those rows | Standing |
|---|---|---|---|
| H<n> | none | | Leader (least evidence against) |
| H<n> | E<a> (II) | High | Refuted |
| H<n> | E<b> (I), E<c> (I) | High, Medium | Refuted as primary; contributor of about <x> bps |
Ranking (least evidence against first): 1. H_ 2. H_ 3. H_ ...
If the answer is a decomposition:
| Component | Basis (evidence row and arithmetic) | Effect | Addressable within <horizon>? |
|---|---|---|---|
| <cause 1> | E<n>: <calculation> | <x> bps | Yes / No / Strategic choice |
| <cause 2> | | | |
| Unexplained residual | <total less components> | <x> bps | |
| Total | | <x> bps against <observed> | |
Reconciliation tolerance: <bps>. Residual within required accuracy: Yes / No.
Largest cause: <H>. Largest addressable cause: <H>.
7. Refutation plan (stage 7)
## Refutation plan: <question>
| Hypothesis | Refutation condition (must hold for H to survive) | Known? (Y/N, source) | Analysis and data | Result that kills H | Owner | Effort (days) | Also discriminates vs |
|---|---|---|---|---|---|---|---|
| H<n> | | | | | | | |
Decision rules agreed before data is pulled:
- If <result>, H<n> is dead.
- If <result>, H<n> is split into whole and part.
Sequence (by refutation power per day of effort):
1. <analysis> (<days>): can kill H<n> and H<m>
2. ...
Stop rule: the plan ends when every live hypothesis has either been refuted or has had all its conditions confirmed.
Hand-offs: <analysis> to value-driver-modeling / market-competitive-intelligence / document-synthesis / primary-research-synthesis, with the evidence row IDs to fill and the threshold.
8. Sensitivity check (stage 8)
## Sensitivity check
Linchpin rows (the I and II cells against the top two hypotheses, plus any row from a single source that several cells depend on):
| Row | Vulnerability (source error, definition, sampling, interested party, alternative reading) | Confirmation check | Distance to flip (how far the value must move to change the ranking) | Status |
|---|---|---|---|---|
| E<n> | | | | Held / Restated / Open |
Rows sharing one source are treated as one linchpin: <list>.
If <linchpin> were wrong, the conclusion would instead be: <alternative>.
9. Surviving hypothesis report (stage 9)
One page. This is the deliverable that goes to executive-communication.
## Surviving hypothesis: <question>
Question: <one sentence>
Day-one answer and what happened to it: <one or two sentences; say plainly if it was refuted and by which row>
Standing of all hypotheses:
- H<n> <label>: surviving leader, <size>; no evidence against; supported by E<..>.
- H<n> <label>: alive as component, <size>; refuted as primary by E<..>.
- H<n> <label>: refuted by E<..>.
- Unexplained: <size>, within / outside required accuracy.
Surviving hypothesis (in full, with numbers): "<statement>"
Confidence: <band: low / moderate / moderate-high / high>, roughly <n>%, because <number of linchpins, their reliability, size of residual, any extrapolation>.
Linchpin evidence: E<..>, E<..>, E<..> (see sensitivity check).
What would overturn it: <two or three specific observations>
Tripwires:
1. <observable, threshold, owner, review date, and what it would mean>
2. ...
Hand-off:
- To executive-communication: this page plus the matrix as backup.
- To strategic-choice: <if the conclusion implies a choice among options>.
10. Partner review checklist (run before anything leaves the team)
Hypotheses
- Four to six hypotheses, each with subject, direction and magnitude, mechanism, window, scope, killing observation.
- At least one artefact, timing or null hypothesis.
- Every hypothesis has a named plausible champion.
- The day-one answer is recorded verbatim with its holder, and its fate is stated.
Evidence and matrix
- Every evidence row has source, date, reliability and type.
- At least one absence row and one assumption row.
- Cells rated row by row; every I and II cell has a one-sentence justification.
- Quantitative rows are rated on magnitude, timing and scope, not direction alone.
- Every row has a diagnosticity rating; Low rows are out of the inference.
- Testimony rows are tagged and do not carry the conclusion.
Inference
- Ranking is by least reliable evidence against, not most consistent cells.
- Every wording change to a hypothesis is logged and keeps a killing observation.
- If a decomposition, components reconcile to the observed total within stated tolerance and the residual is named.
- Largest cause and largest addressable cause are both stated if they differ.
Plan and sensitivity
- Every analysis in the plan names the hypothesis it could kill and the threshold, agreed before data is pulled.
- Analyses are sequenced by refutation power per day; the plan has a stop rule.
- Linchpin rows are named with vulnerability, check, distance to flip and status.
- The report says what would be concluded if each linchpin fell.
Report and boundaries
- All hypotheses reported with their standing, not just the winner.
- Confidence has a band, a rough probability and a reason.
- At least two tripwires with owner, threshold and date.
- No failure modes, mitigations, kill criteria or competitor scenarios in the report; those belong in a pre-mortem or
war-gaming. - No placeholders, no emojis, no em dashes, none of the banned words.