Turn a scan into a strong invention conversation.
A scan hands you mechanisms written up in plain language, each tied to its code. A focused review turns that into the best starting point a patent practitioner can get: an account you have checked, corrected and can defend. Here is how, on a real concept.
Start with a mechanism you can explain.
Pick a result tied to a real design decision: an eviction rule, a recovery path, a scheduling choice, a way of keeping state consistent. Not the highest score; the list order is a starting point, not a ranking. A mechanism you know is one you can check quickly, and one you can explain to someone else.
The example: dual-cache with eviction prioritization
What the scan saidgroupcache keeps a main cache for the keys a server owns and a smaller hot cache of popular keys copied from other servers. When memory runs out, the code decides which cache gives up an entry. Its key insight, in the scan's words: it “prioritizes cache eviction based on relative cache sizes ... evicting from the larger cache when total usage exceeds limits.”
Where it livesUse the code to challenge the story.
Four steps. On the example, they take one short function and find one real mistake.
Open the lines
Every concept names its file and lines and links to them at the exact commit scanned. Read them at that commit: if the file has changed since, the window says so, and you want to review what the scan read, not today's version.
Trace one concrete case
Pick a state and follow it. Say the main cache holds 64 MB and the hot cache 10 MB, and the total is over the limit. Is 10 more than an eighth of 64? Yes, it is more than 8, so the hot cache gives up its oldest entry. Now say the hot cache holds 6 MB: the main cache gives one up, even though it is ten times larger than the hot cache.
Check the difference being claimed
Each concept lists comparable techniques. Here: a single LRU cache with a fixed eviction policy, or separate caches each evicting on its own. Ask what the code does that those do not. In this case it is one rule across both caches, which holds the hot cache to about an eighth of the main cache however the load shifts. That is the part worth talking about.
Correct the account
“Evicting from the larger cache when total usage exceeds limits.”
When the total is over the limit, evict from the hot cache if it holds more than an eighth as much as the main cache, and from the main cache otherwise.
Step 2 already showed the difference: with a 6 MB hot cache the larger cache gives up an entry, but at 10 MB the smaller one does. Write the corrected sentence in your review note; the written-up fields are the scan's record and stay as it wrote them. The figure you can fix in the window, and you should: a diagram makes a wrong account look convincing. In our walkthrough we added the two arrows the rule needs.
Approve, reject or skip.
Nothing is exported until you decide. From the keyboard: j and k move between concepts; a, r and s approve, reject and skip. Select several to decide them at once.
| Decision | Use it when | Record |
|---|---|---|
| Approve | The mechanism is worth carrying into the next conversation. | What you checked and the corrected sentence. Approved concepts are what export writes. |
| Reject | It is routine, a duplicate, inaccurate or in the wrong place. | Pick the reason. It stays on your computer and is never exported. In our walkthrough: one duplicate, one routine. |
| Skip | You need another engineer or another file first. | The question you are waiting on. |
| Held back tab | A mechanism you know is missing from the list. | Each hold says why. Mark whether you think the hold was right. |
Ask questions that get better answers.
The questions an inventor is usually asked produce opinions. These produce the technical material a practitioner can work with.
| Instead of | Ask |
|---|---|
| Is this innovative? | What would the straightforward implementation do here, and why did you change it? |
| Does it work? | Show one case that works and one at the boundary. What was measured? |
| Who wrote this file? | Who came up with this mechanism, what did each person add, and what records show it? |
| Can the diagram be broader? | Which parts or orderings could change while the mechanism still works? |
| Is it important? | What breaks, slows down or costs more without it? |
Why an eighth? The comment above the rule calls it “good-enough-for-now logic”. Was another ratio tried, and what happens to hit rates without the hot cache limit?
Who conceived the rule, and when? git blame gives 41 lines to one author on 23 July 2013, and the code has been public since then. What does that mean for anything built on it?
What to bring to the meeting.
The mechanism, corrected
The problem in a sentence, how the code solves it, and every correction you made against the source.
Where and when
The repository, the commit the scan read, and any changes not yet committed.
One example and one edge
An input and its result, a boundary case, and any measurements. The tests the scan paired with the concept are a good place to look.
Alternatives and people
What else was considered and why it lost, who conceived the mechanism, and any demos, posts or releases with their dates.
Export the concepts you approved and bring them with the scan's summary file and your note. For a filled-in example, here is our review note for the groupcache concept; for a blank one, the checklist.
Questions people ask.
Do I need to read every concept?
Read the whole list once. The order is not a ranking: in the release's labelled test it separated worth-reading mechanisms from routine ones at 0.54, where 0.50 is chance. Then go deep on the few you know best.
Can I edit the write-up in the app?
You can correct the figure: rename, add or remove its parts and draw arrows, and your arrows are exported with the concept. The written fields stay as the scan wrote them, so the corrected account goes in your note.
What if the scan got it wrong?
Then the review did its job. Reject it as inaccurate or wrong location, or approve it with the correction in your note if the mechanism itself is real. A wrong sentence about a real mechanism is still a good find.
Who should do the review?
The engineer who knows the code, ideally with the person who will talk to the practitioner. Practitioners: what a client's scan gives you.
Run it on code you know.
Running it on client code? Where your code goes, by setup.