Skip to content

Resolve v10x semantic review gaps in variable roles, count comparisons and artifact scope #2079

Description

@realmarcin

Independent review of the unchanged v10x CHORUS native artifacts found three draft semantic defects and one audit classification inconsistency. The attempt stopped on an earlier audit-format error and never reached reconciliation, so these are defects in an unfinished draft/audit, not an accepted production record or a demonstrated evidence-gate pass.

  1. Data types promoted to literal variable names. Full /variables/0 through /variables/8 turn the webinar's broad Data type rows (Demographics, imaging, waveform EEG, etc.) into VariableMetadata.variable_name. The schema defines this as the name/identifier as it appears in data files. The source table does not establish individual file variables. The audit marks all nine names supported based on the same table labels; it checks the words without supporting the schema role.

  2. Compatible counts labeled disagreement. Full and core /source_caveats say “The sources disagree on the number of admissions.” The webinar says over 45K as of August 2025; the homepage says 50,000 in an undated current-release snapshot with different population wording. These facts can both be true. Preserve the counts, bounds and temporal context without asserting contradiction. The audit marks the disagreement clause supported merely by quoting both numbers.

  3. Full-only reference copied into core. Core /source_caveats says “those rows are recorded here under variables,” but the derived core has no variables field. Shared prose needs to remain true in both artifact contexts; a deterministic projection and matching shared values do not establish that.

  4. Prospective status inconsistently classified. Audit /source_caveats claim 2 marks the explicit Anticipated Final Dataset target as fact/fact, while /updates/update_details correctly calls it planned/planned. The pinned instructions explicitly require quantities to inherit prospective headings and prohibit relabeling a stated plan as fact merely because it is quoted. The original narrative retains the anticipated qualifier; this finding concerns audit classification, not an observed-count substitution.

Source anchors in the unchanged data/preprocessed/concatenated/CHORUS_preprocessed.txt: webinar lines 242–244 and 260–375; homepage lines 1074–1076. Schema VariableMetadata.variable_name in src/data_sheets_schema/schema/D4D_Variables.yaml. Review covered the complete 1,698-line bundle and all 99 populated full-record values.

Artifact SHA256 identities:

  • Original/current full: d4c69f1563513254cac00537e7cf97cdede92b640927ad69bf24bf0cf12a6de0
  • Original/current core: 720f44fa469794f71f7aa76531ba280497f8ffb0576b08afadf87e932427dd90
  • Failed audit: 853bfbd31aeb51b5c333c0f9151caac2fdaf8dcf6a670318ecee93e4ec58027e

Resolve with general rules and offline counterexamples for schema/container assertions, compatible bounds/dates and artifact-local references; avoid CHORUS/Bridge2AI-specific generation behavior. Review repeated occurrences and the audit's supported judgments. Preserve failed artifacts and require independent unchanged-artifact acceptance of a fresh registered canary. Literal quotes and lexical coverage cannot establish these semantic relations. The reviewed outcome report (PR #2080) records this evidence and its limits.

Related: #1782, #1815, #1801, #1816. This issue does not authorize another paid attempt or alteration of v10x.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions