The process
The process has eight steps. The numbers are used the same way in this document, the schema and the simulator.
| Step | What happens | Who does it |
|---|---|---|
| 1. Intake | The user's input becomes a brief. Gaps become questions to the user. | Code and the model fill in what they can read. The user answers the rest. |
| 2. Rule out techniques | Each technique is tested against the brief. Techniques that cannot run on the material are removed. The rest are eligible. | Code |
| 3. Choose analysis techniques | The model reads a list of the eligible techniques and chooses at most 10, with a reason for each. | The model. The user can change the list where the application shows it. |
| 4. Plan analysis sequence | The chosen techniques are placed in a sequence, with a description of what each technique does and a note of which need additional sources. Preparation techniques and checks are added. | Code |
| 5. Run the techniques | 5.1 Examine the question. 5.2 Prepare the material, including any additional sources. 5.3 Analyse the material. | The model, with code for arithmetic and text matching. The user confirms the request after 5.1. |
| 6. Check the findings | Verification and challenge techniques run on the findings. The library checks are applied to each result. | The model and code. Checks that need a person are listed as not done. |
| 7. Combine the findings | The findings become conclusions. Each conclusion gets a confidence label and a one-line reason. | The model. The user can choose the method. |
| 8. Answer | The answer is given with a list of what it is based on. The user can ask for a follow-up round. | The model. The user decides on a follow-up round. |
Terms
Each term below has one meaning in this document.
| Term | Meaning |
|---|---|
| Request | What the user asks, in their words. It can have several parts or be open-ended. |
| Settings | The three answers that shape the run: stakes, strength of statement needed, and kind of finding wanted. |
| Material | The content the user gives for analysis. |
| Description of the material | The properties of the material that requirements are tested against, such as content type and number of sources. |
| Brief | The request, the settings, the material and the description of the material. |
| Technique | Any of the 236 analytical techniques in the catalogue. The step names call them analysis techniques. |
| Requirement | A condition a technique places on the brief. A technique that fails a requirement is removed. |
| Eligible technique | A technique that meets every requirement that applies. |
| Chosen technique | An eligible technique the model selected in step 3. |
| Sequence | The order in which the chosen techniques run, in four named groups. |
| Reserve list | Techniques the model wanted beyond the limit. They are offered in a follow-up round. |
| Removed, dropped | Removed: failed a requirement in step 2. Dropped: chosen, then left out because the additional source it needs was not found, or because its run did not finish. |
| Result | Everything one technique produces. |
| Finding | One statement in a result. |
| Conclusion | One statement in the answer, drawn from the findings. It has a title that states the finding as a sentence, an explanation, a confidence label and a reason. The label and reason are recorded. The answer groups conclusions by the label and shows it only when the user asks (decision 46). |
| Insight | A finding that goes against, or beyond, what the request or the material would lead a reader to expect. It states the expectation and what the analysis found instead. |
| Summary | The first part of the answer: the direct answer to the request in plain sentences, with how far to trust it and what it leaves open. |
| Library check | One of the 18 checks in the check library, such as a spot check or repeat runs. It is applied to one result. |
| Work | Everything produced in steps 1 to 7: the description of the material, the eligible and chosen techniques, the sequence, each technique's result, the checks and the combined findings. |
| Output | The answer given to the user in step 8. It is the only output. |
| Follow-up round | A second pass through steps 2 to 8, run only when the user asks for it after reading the answer. |
Kinds of technique
Every technique belongs to one of five kinds. The kind decides what the technique works on and when it runs.
| Kind | Works on | Produces | Runs in |
|---|---|---|---|
| Question technique | The request | Findings about the request | 5.1 |
| Preparation technique | The material | Prepared material | 5.2 |
| Primary technique | The material, including prepared material | The findings that the answer is built on | 5.3 |
| Verification technique | The findings of the analysis | Findings about those findings | 6 |
| Challenge technique | The findings of the analysis | Findings that question them | 6 |
The kind comes from the technique's phase tag: frame gives a question technique, gather or verify and prepare gives a preparation technique, analyse gives a primary technique, and validate gives a verification technique or, in the Challenge family, a challenge technique.
Rules that apply throughout
- A primary technique works on the material. It never works on another primary technique's result. Primary techniques are independent of each other and can run in any order.
- Combining the findings (step 7) is a fixed step, not a technique. The catalogue lists techniques only.
- Stakes and strength of statement are set for the brief. They are never stored per technique.
- Only the user changes the request. The process can show the user what it found about the request and ask. It never rewrites the request itself.
- The process never starts a follow-up round on its own.
- A technique whose run does not finish is dropped, with the reason. The process never writes a result for it. The answer lists it among the techniques that were dropped.
Status
The process defines the states of a run so that whatever runs it (a chat, an agent, an editor assistant or a script) can show where the run is. The design defines the states. It does not define how they are delivered: an application can narrate them, send them as events, keep them in a record or map them to its own task list. A run does not have to be resumable.
| Applies to | States |
|---|---|
| A step | Waiting to start. Running. Waiting for you. Done. Skipped (the step does not apply, for example 5.1 when no question technique was chosen). |
| A technique | Removed (step 2). Eligible. Chosen. Reserve. Running. Done. Dropped (with the reason). |
| A library check | Done. Not done, because it needs a person and the user did not review the sample (high stakes only). Left out, because it needs a person and the stakes are below high-stakes. |
| The answer | Not yet given. Given. |
A step is waiting for you only at one of the decision points: confirm the brief, confirm the request after 5.1, review a sample of the results (high stakes only), choose the combining method, confirm that estimates measure the same thing, or ask for a follow-up round. The status names the decision point, so the application can show what the user needs to decide.
Rules for the status:
- The process sets the state before a step begins and after it ends, so the state is never behind the run.
- At most one step is waiting for you at a time.
- A step is done only when all its sub-steps are done or skipped.
- The status is work. The answer is the only output.