Template · Research & Innovation
Research and Innovation: Templates
Four working templates: the question intake, the experiment design with its kill criterion, the negative-result record and the handover to product.
Markdown. No sign-up, no email.
1. Question intake#
| Field | Entry |
|---|---|
| Question | [Answerable. "Is AI useful for X" is not; "does approach Y beat our current approach on Z by more than 20%" is] |
| Which decision does the answer change? | [Name it. If nobody can, decline the question] |
| Who is waiting on it | |
| Asked by | [If this is always us, the queue has drifted from what the company needs] |
| Prior work checked | [Against the negative-results record. Have we paid for this answer before?] |
| Can it be settled in six weeks? | [If not, decompose it or decline] |
| Accepted | [Yes / No, and why] |
2. Experiment design#
Completed before any work begins.
| Field | Entry |
|---|---|
| Question | |
| Cheapest sufficient test | [Not the most rigorous. The cheapest that could settle it] |
| Kill criterion | [What result makes us stop. Written now, agreed now] |
| Time box | [A date it ends regardless of progress] |
| What we predict | [Recorded before the run. A prediction written afterwards always matches] |
| Success would mean | [What we would do differently] |
| Cost estimate | |
| Run by | |
| Interpreted by | [Ideally not the same person. Execution and interpretation are separated on purpose] |
3. Negative-result record#
Completed on every killed experiment, before it closes. The most valuable document this function produces.
| Field | Entry |
|---|---|
| Question | [Phrased as it will be asked again, so the search finds it] |
| What we tried | [Enough detail that the next attempt is genuinely different] |
| What happened | |
| Why it failed | [Distinguish "does not work" from "did not work for us, this way, at this time"] |
| What it cost | [Time and money. Makes the value of not repeating it visible] |
| What would have to change to revisit | [Specific. "Better models" is not a condition; "context windows above 2M at under half current cost" is] |
| Confidence in the conclusion | [High, medium, low. Low is a legitimate and useful answer] |
| Recorded by, and date |
4. Handover to product#
| Field | Entry |
|---|---|
| Finding | |
| Replicated by | [Someone who did not run it first. Most reversals happen here] |
| What it enables | |
| What it does not | [The limits. This row prevents the finding being oversold downstream] |
| Evidence | [Where the data is] |
| Known risks | |
| Recommended next step | [Prototype, pilot, or nothing yet] |
| Handed to | [AI Engineering and Product, or AI Strategy] |
| Research does not productise this | ⬜ [Acknowledged. The team that proved it works is the worst judge of whether it is ready] |
Using these together#
The charter sets out what this function owns, the SOPs say when each is produced, the KPIs define what the negative-result record feeds, and the workflows name who receives each output. An experiment design with a blank kill criterion is not an experiment, it is an activity with a budget.