docs(sdc): threshold-tuning guidance + heuristic framing (#348)#408
Merged
Conversation
Signed-off-by: Tri Lam <tri@maydow.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Addresses #348 — pattern-13 (silent data corruption) shipped with
sdc_accuracy_drop_thresholddefaulting to0.005(0.5pp) but the spec only carried "raise" guidance and offered no citation for the 0.5pp basis. Real-world SDC can land sub-threshold (single-bit weight flips in small models; transient single-batch corruption in large training).Three prose-only changes:
docs/patterns/13-silent-data-corruption.mdcovering both directions (raise + lower) with concrete triggers, the relationship tosdc_accuracy_only_multiplier, a copy-pasteable YAML example, and a "Why the default stays at 0.005" note.docs/ATTRIBUTES.mdrow fortracecore.alert.silent_data_corruption.accuracy_dropnow notes the threshold is operator-tunable and points to the new §.Root cause
The spec asserted "typical SDC regressions land between 0.5pp and 3pp" with no citation, and the §"Edge cases" guard only documented when to raise the threshold above the recipe's variance band. Operators with high-precision recipes (variance bands <0.1pp) or paired vendor SDC counter coverage had no in-spec rationale to lower the knob — so legitimate sub-0.5pp regressions were silently skipped.
Claims verified against detector code
sdc_accuracy_drop_threshold = 0.005—module/processor/patterndetectorprocessor/config.go:178.sdc_accuracy_only_multiplier = 2.0—config.go:189.[0, 1]-outside —config.go:429-430.accuracy_drop < threshold—module/pkg/patterns/silent_data_corruption.go:286.accuracy_onlybranch gate isdrop >= threshold * multiplier—silent_data_corruption.go:295.processors.patterndetector—module/processor/patterndetectorprocessor/example_config.yaml.No new field names invented; no thresholds cited that don't exist in code.
Test plan
sdc_accuracy_drop_threshold,sdc_accuracy_only_multiplier) under the documentedprocessors.patterndetectorparent.Closes #348