provenance truth pass: resolve prior-art citations, correct pre-production/compression claims

[keithbinkly]

Summary

Two frozen adjudication rulings (R4, R5) from a provenance sweep of the live meta-context pages, applied to this repo. Ruling source: dc-tracks/dbt-summit-econ-open provenance-adjudication-2026-07-30.md.

R4 — vendor matrix citation now resolves

The live meta-context page's vendor comparison matrix cites spec/prior-art.md as its "full detail" source for every row, but 6 of 17 rows had no entry there. Added:

• Euno — SaaS metadata/governance graph platform (EQL query language, typed "metadata tags," free-text "AI Instructions" for deep business reasoning, MCP delivery)
• Kilo Code — open-source coding agent; git-tracked markdown (AGENTS.md, custom rules, deprecated Memory Bank), context bound to repo paths not data objects
• Cassis — hosted ontology aggregating dbt/Looker/Snowflake/docs; markdown-prose interpretive layer; delivery mechanism honestly hedged (not fully specified in its own public docs)
• WHOOP snowflake-semantic-tools — open-source CLI compiling git-versioned YAML to Snowflake Semantic Views; structured fields + verified-query exemplars, business rules as a free-text instruction string
• Ramp Research — internal Slack-based analyst agent; indexed metadata + a deliberate choice to keep deep domain knowledge as prose docs, not schema fields

Each entry describes what the tool stores and how it delivers context — no scoring against this schema's five layers, no L1–L5 labels applied to the added tools, no competitor language, no ✓/✗ marks. Each carries its source verification date (2026-06-26 or 2026-07-01, per the source documents).

Power BI (prep-for-AI) is intentionally skipped. The six-axis comparison draft (the only source naming it) itself flags Power BI as part of an unlinked "Set A citation gap" — pending a citation pass against current vendor docs, with no primary source fetched. Inventing a defensible entry from an admittedly-unverified row would repeat exactly the kind of provenance failure this adjudication exists to fix, so that row is left out rather than filled with guessed detail.

Sources used (read-only, not modified): analytics-workspace/data-centered/content/articles/drafts/six-axis-context-tool-comparison-DRAFT-v2.md (the v2 draft, per instruction — not the non-v2 version) and .../content/research/dbt-summit-2026/context-tools-research-report.md.

R5 — three factual corrections

(a) spec/known-gaps.md Gap 1 "Raised by" line

• Before: **Raised by:** Production deployment in regulated public-transport domain
• After: **Raised by:** A pre-production field report from a regulated public-transport deployment (design-and-development phase, not production metrics)
• The source (an external adopter) explicitly described pre-production, design-and-development experience — not production metrics. Adopter name/org withheld pending their own consent (unchanged from before — the original text didn't name them either).

(b) CHANGELOG.md 0.1.0 entry

• Before: ...validated in a production financial-services deployment.
• After: ...piloted in a financial-services deployment during its design phase.

(c) eval/results.md + CHANGELOG.md — 5.6× → 4.1× compression

• The eval's own word-count table gives C3 = 4,743 words, C1b = 1,157 words → 4,743/1,157 = 4.10×, not 5.6×. The live page already states 4.1×; only these two internal artifacts disagreed with it and with each other. Fixed both instances (grep confirmed no other repo occurrence of "5.6x" tied to this figure — the only other "5.6" hits in the repo are unrelated GPT-5.6 model-name references in eval/design-v2.md, left untouched).

Gates (pasted)

1. Validator test suite (from the validator/ directory):

$ cd validator && python3 -m pytest tests/test_rules.py -q
.........                                                                [100%]
9 passed in 0.04s

2. Grep — zero remaining overclaim strings:

$ grep -rn "5\.6" --include="*.md" .
eval/design-v2.md:43:- **Reader (6):** Fable 5 · Opus 4.8 · Sonnet 5 · Haiku 4.5 · GPT-5.6 (codex) · Gemini 3.1 Pro (agy).
eval/design-v2.md:85:   judges** (GPT-5.6 and Gemini 3.1 Pro). Blinding: responses are stripped of any mention of
eval/design-v2.md:108:- GPT-5.6: `codex exec -m gpt-5.6-luna` (stdin-hang guard: `< /dev/null` / stdin-file pattern).
# (all 3 hits are unrelated GPT-5.6 model-name references — zero "5.6x compression" hits remain)

$ grep -rn -i "production.*deployment\|validated in a production\|production financial" --include="*.md" .
spec/known-gaps.md:14:**Raised by:** A pre-production field report from a regulated public-transport deployment (design-and-development phase, not production metrics)
# (the only hit is the corrected line itself — describing PRE-production; zero overclaim strings remain)

3. All 6 tools present in spec/prior-art.md (Power BI skipped, reason above):

Euno, Kilo Code, Cassis, WHOOP `snowflake-semantic-tools`, Ramp Research — each has a "### <Tool>" entry.
Power BI — 0 hits (intentionally excluded).

4. git diff --stat (only intended files):

 CHANGELOG.md       |  4 ++--
 eval/results.md    |  2 +-
 spec/known-gaps.md |  2 +-
 spec/prior-art.md  | 20 ++++++++++++++++++++
 4 files changed, 24 insertions(+), 4 deletions(-)

Test plan

• Validator suite green (pasted above)
• Grep confirms zero remaining overclaim/wrong-figure strings
• Grep confirms all 6 R4 tools present (5 added, Power BI correctly absent)
• git diff --stat touches only the 4 intended files

Corrections to public claims; Keith merges.

https://claude.ai/code/session_017uymU7VUQ2ngpuASgzt3t4