Search plans

What the planner does

Once claim features are accepted, the case owner chooses Generate plan in the Plan region. The plan says what it is: "A recommendation from the deterministic planner (planner_v1): which searches to run and why. It is not evidence and not a finding."

The planner is a rule engine with explanations. No language model is used anywhere in it, and it reads no clock, randomness or environment, so the same accepted features give the same plan, byte for byte. For each accepted feature it recommends searches:

• an interactive search of each recited SEQ ID NO, of its own molecule type;
• a deep search beside it for every identity floor, whatever its value, and for every recited sequence, fusion construct and fixed fragment span, unless the query is too short for alignment to reach what the claim recites;
• a search of a claimed fragment's residues;
• a pattern search across the full corpus for variants and open positions;
• a search of each component of a fusion that names a listed SEQ ID NO.

A nucleotide item states that both strands are searched. When the application number is known, every search excludes the application's own filing. Each item carries a completion estimate.

What every item shows

The plan version's items are listed in one table, "Searches in this plan version", with the columns Item, Search, Claim group, Tool and depth and Status, and Sent once an item has been sent. Show detail opens an item beside the list; on a narrow screen or at a high zoom level it opens below the list. Back to list returns focus to the item's row, and Hide detail closes it. A reason you type for one item is kept while you look at another.

An open item shows its SEQ ID NO, its depth (Interactive or Deep), the corpus snapshot it will run against, and its estimate. Why this search names the rules that fired, the inputs they read, and what would change the recommendation.

Every accepted feature is accounted for. One that no rule plans is listed under "Features with no search planned" with the reason, for example a fusion component written as text rather than as a SEQ ID NO.

Decide and send

For each item, open it and choose Accept, Reject or Revise, then Record decision. A revision can change the depth or the databases to search.

Execute accepted items sends the accepted items through the ordinary search path. Searches that use the deep-search quota are named in one confirmation, "Send these searches that use deep-search quota?", before Send these searches. If the answer is lost, executing again cannot send an item twice.

Deep-search limits and over-limit batches

Deep searches past the number you may run at once move to later batches, each sent once an earlier batch has finished. Deep searches you may not run at all, or that are past your remaining quota, move into a text-only request draft for a specialist search team. Nothing is sent from that draft; sending and tracking such requests is planned.

How long the plan should take gives the plan's estimate when every item has a measured range, and says so when one does not.

Limits

• The planner does not choose databases. Each search runs against the corpus snapshot active when it is sent, across all of its databases unless a revision narrows them.
• A query too short for the alignment search to find inside a longer record carries the known limit as a statement, word for word, on its item. For a short protein query, the item says that no limit is registered and that whether it is found inside a longer record is not measured.
• A fragment feature with no positions joins the full-length search, and its item says the plan does not claim that search finds a record holding only the fragment.
• Only the case owner generates, decides on or executes the plan: "Only the case owner can generate, decide on or execute this plan."