failure-taxonomy

Failure Taxonomy

Dead branches are the continuous-learning asset — but only if their deaths are recorded precisely enough to retrieve later. The difference between "didn't work" and a useful taxonomy row is the whole discipline: a future frame-loader matches on shape and verbatim symptoms, not on summaries.

The row

pqa.memory.record_failure(conn, session, task, Failure(approach, death_reason, conviction))

• approach — the topology, not the implementation: "fixed-window counter", "trait-object state machine", "denormalized read model". A future task matches on shape; "used a dict" matches nothing.
• death_reason — VERBATIM from the killer: the first failing assertion from the verifier, or the adversary finding's title + trigger condition. "fails burst-at-boundary when two windows share an edge" retrieves; "had some test issues" does not.
• conviction — the branch's flagged level (high|medium|low|none). High-conviction deaths are the calibration loop's best rows.

Layer prefixes — who killed it

Prefix the death_reason with the killing layer when it is not obvious:

Prefix	Meaning	Retrieval semantics
verifier:	failed real tests/types/lint	approach refuted in this context
adversary:	critical unresolved finding	approach has a named hole
divergence:	superposition collapsed; branch was a duplicate	prompt/axis failure, NOT approach failure
budget:	run aborted before judgment	NOT a merit death — never retrieve as refutation
reconcile:	passed in branch, failed after merge	approach interacts badly with the tree

The budget: prefix is load-bearing: an aborted run must never teach the harness that an approach is bad. The prefix is what protects it from being read as evidence.

Protocol

1. Search before write. search_failures(conn, task, limit=5) (FTS5 relevance-ranked, LIKE fallback on FTS5-less builds). A hit with the same substance — same approach, same death reason — means update the story, not duplicate the row: record the new occurrence referencing the prior id ("second death, see failure:12"). A taxonomy is not a log.
2. Record every dead branch, including boring deaths. Boring deaths repeat the most; the row costs tokens once and saves a branch every time it retrieves.
3. Keep secrets out of approach and death_reason — registries persist across runs and exports. The secrets hook guards reads; the writer is the last line.
4. Retrieve by relevance at frame-load, never recency. prior_art(conn, task, max_tokens=400) is the sanctioned injection: ranked, token-capped, ids cited so the run report can name what shaped the frame (memories_injected).
5. Append-only. Wrong rows are superseded by sharper rows, never deleted — supersession is itself signal that the first description missed something.

Worked example

Bad row (useless tomorrow):

approach: "rate limiter v2"        death_reason: "tests failed"

Good row (retrieves and teaches):

approach:     "fixed-window per-tenant counter"
death_reason: "verifier: test_burst_at_boundary — 200 admitted in 200ms against
               100/min limit; windows sharing an edge each grant full quota"
conviction:   "none"

Next month, task "throttle webhook fan-out": search_failures matches the window/limit shape, prior_art injects the row with its id, the frame-loader surfaces failure:12 in the self-eval frame, and the spawn prompts either exclude fixed-window or demand the boundary case be handled. One recorded death = one branch not wasted, in every future run that rhymes with this one.

Budget contrast: the same branch dying because the run hit its token cap is recorded as budget: aborted at collide stage, 612k/800k tokens spent — and when retrieved it reads as "untested", never as "refuted".

Anti-patterns

• Paraphrasing the killer. Summaries feel cleaner and retrieve worse. Verbatim or nothing.
• Implementation-naming. "used asyncio" is not an approach; "single-writer event loop, no locks" is.
• Skipping boring deaths. The fifth duplicate death you didn't record is the sixth duplicate branch you will pay for.
• Letting budget deaths refute. An aborted run treated as evidence poisons every future frame that retrieves it.
• Deleting rows. Append-only; supersede with sharper rows.