Failure Mode Annotation Card¶
Template — instantiates Correspondence Violation Detection and Theory Refinement
A per-violation template that binds one failing case to its suspected cause, diagnostic evidence, a proposed refinement, and the follow-up test that will confirm the fix.
A Failure Mode Annotation Card is a per-violation template — one card per failing case — that captures the full working dossier of a single anomaly: what diverged, the suspected cause and its fingerprint, the diagnostic evidence behind that guess, the proposed refinement, and the follow-up test that will confirm or kill it. Its defining quality is that it is the atomic unit of an investigation-in-progress: it binds this specific failing case to this specific proposed fix and this specific test, so a refinement can always be traced back to the exact violation that motivated it. It is not a queue of many cards with no per-card depth, and it is not a settled scope statement — it is the one place where a single violation's cause, hypothesis, and verification live together while the work is still open.
Example¶
A new end-to-end speech-recognition model beats the previous hybrid system on overall word error rate but regresses on very short utterances — transcribing an isolated "no" as "know," dropping a standalone "uh." Each regression gets a Failure Mode Annotation Card. One reads: Case — utterance set S-0.5 (clips under half a second); Divergence — old model correct, new model wrong on 43 of 50 (illustrative); Suspected cause — the end-to-end model's built-in language prior overrides weak acoustic evidence when the clip is too short to accumulate signal; Signature — failures concentrate entirely below ~0.5 s and vanish above it; Evidence — error rate versus clip length shows a sharp knee; Proposed refinement — add short-utterance examples and down-weight the language prior for sub-second inputs; Follow-up test — rerun S-0.5 plus the full benchmark. The card is the object that carries this one failure from guess to verified fix.
How it works¶
- One card per distinct failing case. Each card is scoped to a single coherent violation, not a bundle, so its cause and fix stay specific.
- Bind signature to hypothesis to test. The card's fields force a chain: here is where and how it fails, here is why we think so, here is the change we propose, here is what will confirm it.
- Lifecycle. A card moves open → cause proposed → refinement proposed → tested → closed, and cannot close without its follow-up test having run.
- Aggregate into patterns. Many cards sharing a signature reveal a class of failure that a single card would miss, feeding the eventual scope statement.
Tuning parameters¶
- Field set — how much each card captures. Rich fields make the dossier thorough but raise the cost of filing one; thin fields are fast but lose the evidence trail.
- Case granularity — one card per individual case versus one per failure class. Per-case is precise but proliferates; per-class is compact but can blur distinct causes.
- Evidence bar — how much proof the suspected cause field demands before a fix is proposed. A high bar prevents jumping to a fix; a low bar speeds filing but invites guesswork.
- Linkage requirement — whether a card must name a follow-up test to advance. Requiring it enforces verification; relaxing it lets untested fixes slip through.
When it helps, and when it misleads¶
Its strength is traceability: every refinement traces to a card, every card to evidence and a test, so the archetype's ad hoc-patch failure mode is blocked at the unit level — you cannot make a change without a case that motivated it and a test that will check it.
Its failure mode is the card graveyard — cards filed and never worked — and, worse, the confidently-guessed suspected cause that skips the evidence and jumps straight to a fix. The discipline that guards against this is genuine root-cause practice: the "5 Whys"[n1] habit of pushing past the first plausible explanation, and the hard rule that a card cannot close until its follow-up test has actually run. A card whose fix was never tested is not closed; it is abandoned. Because the card does record a proposed cause and fix, the discipline here is an informal self-check on the evidence, not a formal adjudication — that heavier judgment belongs to the panel.
How it implements the components¶
The card fills the per-case capture-and-propose part of the machinery:
divergence_signature— the suspected cause and signature fields record the fingerprint of one violation: where it concentrates, how it varies, what pattern distinguishes it.refinement_hypothesis_generator— the proposed refinement field is a concrete candidate fix, generated for and bound to this specific case.
It does not implement boundary_condition_extractor or versioned_theory_change_log — turning the accumulated cards into a ratified, versioned statement of refined scope is the job of Boundary Condition Memo. The card is the working dossier of one violation; the memo is the settled conclusion drawn once a class of cards resolves.
Related¶
- Instantiates: Correspondence Violation Detection and Theory Refinement — the card is the per-violation record that keeps every refinement tied to a real failing case.
- Consumes: Residual Divergence Map — the located divergences it documents case by case.
- Sibling mechanisms: Anomaly Triage Board · Assumption Audit Worksheet · Boundary Condition Memo · Expert Adjudication Panel · Limiting-Case Test Suite · Overlap-Regime Benchmark Table · Parameter Sweep Matrix · Regression Correspondence Harness · Residual Divergence Map
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: The per-case template externalizes a failure signature, causal hypothesis, diagnostic evidence, proposed refinement, confirming test, and lifecycle state.
Nearest alternative: Interface, Display & Cue — Fields prompt disciplined completion, but the durable structured annotation is the operative artifact.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Engineering & Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Binding observed failures to causes, corrective refinements, and verification tests follows engineering failure analysis and corrective action.
Related originating lineages:
- Computer Science & Software Engineering — Bug reports and regression-test records materially shape the per-case annotation artifact.
- Statistics & Experimental Design — Hypothesis testing materially supplies the follow-up test that discriminates a proposed cause.
Review resolution: Both reviewers agree that engineering_design is primary. I retain statistics_experimental_design, computer_science only as formative origin lineages; cross_disciplinary_synthesis is appropriate because the final form materially combines the agreed primary with the retained formative lineages. Reach is multi_domain because the structure transfers across several fields but is not a near-universal human pattern, an applicability judgment kept separate from provenance. Encyclopedia synthesis is true because the exact generalized artifact is an encyclopedia-authored combination or refinement. No unresolved historical ambiguity remains after reconciling the secondary fields.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; medium confidence.
Notes¶
[n1] The 5 Whys, a root-cause technique originating in Sakichi Toyoda's Toyota production system, iteratively asks "why?" of each answer to push past the first surface explanation to an underlying cause. On an annotation card it is the guard against filling the suspected cause field with the nearest convenient guess. ↩