KTIMA is a local-only screener for Cyprus land opportunities. It runs a 15-stage Python pipeline over public cadastral, planning, auction and terrain data and produces an evidence dossier per parcel. A FastAPI and React map terminal sits on top. Orchestrated coding agents built almost all of it. Projects built this way have a recurring cost: a new agent session does not remember the defects earlier sessions found. KTIMA handles this with a catalogue of failure modes. Each mode has a code, so a known defect can be cited in a few characters.
What does KTIMA's failure-mode taxonomy contain?
The catalogue has three coded families and one standalone rule. The extraction entries record a cause and a detection for each mode. The entries named in the source material are listed below.
| Code | Family | As recorded |
|---|---|---|
| E1–E8 | Extraction errors | 8 modes, each with its cause and its detection |
| E6 | Extraction | Fabricated rows that survive registry verification |
| E7 | Extraction | Registration-is-a-share |
| E8 | Extraction | VAT-exclusive reserve |
| L1 | Listing match | Pin-containment brittleness |
| L2 | Listing match | A bundle is not an ambiguity |
| L3 | Listing match | A differing registration is not a contradiction |
| M1 | Bulk pull | A smoke-test page can silently truncate a whole district |
| IDENTITY RULE | Standalone | Names the one canonical key |
The names take two forms. Some are plain labels, such as "VAT-exclusive reserve" and "pin-containment brittleness". Others are written as corrections. "A bundle is not an ambiguity" and "a differing registration is not a contradiction" tell the reader which wrong conclusion to avoid.
Why give each failure mode a code instead of a description?
A description has to be rewritten each time it is needed, and the wording changes from one rewrite to the next. Two agents can describe the same bug in different words and fail to see that it is the same bug. A code always points to the same entry. Writing "this is E7" refers to one place where the defect is already written down, so the reader does not have to rebuild that context from scratch.
A code is also short enough to fit in a code comment or a one-line handoff note, where a paragraph of explanation would not. In general, people keep using a vocabulary only if it is cheap to use, and a code like E7 takes almost no effort to write.
What does M1 show about silent failures?
M1 makes the strongest case for writing failures down. The runbook entry reads: a smoke-test page can silently truncate a whole district. Its worked example is Larnaka, which "would have been mirrored as 43 parcels instead of 185,605, silently, with a normal-looking log line."
The important words are silently and normal-looking. A failure that leaves no visible trace is easy to miss and easy to forget. Giving it a code and a catalogue entry keeps the knowledge in the project after the session that found it has ended.
What does KTIMA's identity rule say?
The E, L and M codes describe things that go wrong. The IDENTITY RULE is different: it names the one canonical key. The runbook's statement of the rule includes two explicit warnings: "The composite id is NOT unique" and "Any join on parcel_uid alone is a bug." The rule is kept as one named item, apart from the coded error lists, so it can be cited by name in the same way the codes are.
How are the codes used across handoff docs and code comments?
The codes are part of the project's working vocabulary, not only reference material. The main handoff document has a gotchas section titled "GOTCHAS (beyond RUNBOOK E1-E8 / L1-L3 / M1)". The title assumes the reader already knows the coded set, and it places new issues outside that set. The same document's read-order instruction cites the error modes by code. The codes also appear as shorthand inside code comments.
How do you start a failure-mode taxonomy for your own agent pipeline?
The following is general practice suggested by this approach, not a description of how KTIMA's catalogue was built.
- Group by family. Sort failures into families that match how your team talks about the system. KTIMA uses extraction, listing match and bulk pull.
- Give each failure a code and a short name. Where the trap is a wrong inference, consider naming it as a correction.
- Record the cause and the detection. Write down how the failure arises and how to spot it.
- Keep core rules separate. Hold any standing rule, such as which key is canonical, apart from the bug list.
- Use the codes everywhere. Cite them in handoff notes, reading instructions and comments.
A catalogue that sits in one document does little. It becomes useful when agents and humans use its codes as their normal way of talking about bugs.