quality-research

name

quality-research

description

Answer a quality-standards question (AIAG FMEA/SPC/MSA thresholds, ISO/NIST concepts, why a rule like the ndc threshold of 5 or an RPN/%GRR band exists) by calling the quality-platform qdb_answer_question MCP tool over HTTP for a cited, page-located answer; falls back to this project's own public ASSUMPTIONS_LOG citations when the endpoint is unavailable. Use for "why" and "what does the standard say" questions, not for running fmea_score, an SPC chart, msa_gage_rr or controlplan_build.

quality-research — ask the standards

Answer a question about a quality standard by calling the quality-platform qdb_answer_question MCP tool over HTTP. The endpoint owns retrieval, refusal and citation; this skill decides when to call it, how to present what came back, and what to do when it is unreachable. See skills/CONVENTIONS.md for the rules this skill is written to.

When to use

The user is asking why a standard says what it says, where a threshold comes from, or what AIAG / ISO / NIST actually publishes about a topic. Not a request to compute anything.

What the user is asking	Where it goes
"Why is the ndc threshold 5?"	here
"Does AIAG publish an RPN number that forces corrective action?"	here
"Is a %GRR of 12% acceptable under AIAG MSA?"	here
"Why check control-chart stability before trusting Cpk?"	here
"What does ISO 9001 require for a documented quality manual?"	here
"Score this failure mode: severity 9, occurrence 8, detection 5."	fmea
"What does a severity rating of 9 mean?"	fmea (fmea_get_scale)
"Run a Gage R&R on this study."	msa
"Run a capability study against LSL 9, USL 11."	spc
"Build a control plan from this FMEA."	control-plan

This is not the skill for running an engine. A request that carries data — ratings, a study, readings, an FMEA table — is a tool call, and belongs to fmea, msa, spc or control-plan. If this skill fires on one anyway, hand it to the owning skill rather than answering it from a citation. That boundary does not move when a project is loaded: with a project on disk this skill may read an already-computed result (step 1b) to say what the project's own value is, but "run a Gage R&R on this study" is still msa's job, not this one's.

Steps

1. Call the endpoint first. Attempt qdb_answer_question with the user's question, k=5, and — only if the user named one — standard, source_id or region to narrow retrieval. scripts/call_qdb_answer_question.py is the same call as a runnable script; it reads QDB_MCP_URL and QDB_MCP_TOKEN from the environment. This is always tried first: it is the primary path, not an optional enhancement.

   Step 1b — if a project is loaded, also read the one artifact the question is about. A project is loaded when the user named a project directory (or one is already the working project) and <project-root>/project.yaml exists and parses; note its project_id. If it is absent, or no directory was named, there is no project — skip this step entirely and answer the standards question exactly as the four worked examples below do. With a project loaded, map the question's topic to a single artifact and read only that file with the host's Read tool, taking only the fields the question needs:

   Topic	Artifact	Field(s)
   %GRR, Gage R&R, ndc, EV/AV/PV	msa/gage-rr.json	pgrr_study, pgrr_tolerance, ndc, verdict, characteristic
   RPN, Action Priority, severity/occurrence/detection	fmea/fmea.json	the relevant FailureLink row's ratings
   Cp/Cpk/Pp/Ppk, capability	spc/results/<characteristic>.json	capability, oos_signal_count

   The full table, the detection procedure and the end-to-end worked example are in references/project-context.md. Three rules hold in all cases. Never recompute: the values in these files are the engines' own output — read them as written, do not derive them from components, re-round them, or re-run the verdict logic by hand; if the file does not carry the number the question needs, say so. Read-only, and only this file: this skill never writes a project artifact and never proposes a change to one — that is the post-1.0 project-aware co-pilot's job — and it opens no file beyond the one the topic maps to. If that file does not exist, say the project has no data for that artifact and fall through to standards-only. the project read augments step 1, it does not replace it — the answer must ground in both the corpus and the artifact. Compose the two halves separately: the project's own value and the file it came from, then the standard's threshold with its citation, stating which published band the project's value falls into — never a verdict the standard does not publish.

Whichever path answered, keep the two halves apart: what the standard publishes and what this platform additionally does. Blending them is the failure mode all four examples below are chosen to demonstrate.

Worked examples

Each runs once through the endpoint shape (a successful CandidateAnswer) and once through the fallback shape (the ASSUMPTIONS_LOG.md RULE that answers it offline).

1. MSA — the ndc threshold of 5

User: "Why is the ndc threshold 5?"

Endpoint shape — qdb_answer_question narrowed to MSA returns a populated CandidateAnswer. Illustrative field shape only — the endpoint is unprovisioned today, so the item_id and cited_page below are placeholders, not a real MSA locator; a live call fills them from the corpus:

item_id:         <opaque corpus item id>
text:            "AIAG MSA sets the number of distinct categories at five or more; below
                  that the measurement system cannot resolve enough levels of part
                  variation to be used for analysis."
cited_source_id: "aiag-msa-4e"
cited_region:    "Ch. II Sec. D"
cited_page:      <page number as returned>
refused:         false

Present that text and whatever locator actually came back — never the placeholders above.

Fallback shape — read apps/msa/docs/ASSUMPTIONS_LOG.md RULE 9, which quotes the manual directly:

For analysis, the ndc is the maximum of one or the calculated value truncated to the integer. This result should be greater than or equal to 5.

Nuance to preserve: that criterion is AIAG's only published ndc rule. The platform's "2–4 Marginal / below 2 Reject" sub-bands are this project's own internal design choice, not AIAG's — RULE 9 and RULE 10 say so explicitly. The manual's ndc formula itself lives in RULE 9; point at it, do not restate the arithmetic here.

2. FMEA — AIAG publishes no RPN action threshold

User: "Does AIAG say there's an RPN number that means I have to take corrective action?"

Endpoint shape — narrowed to the AIAG & VDA FMEA Handbook, a CandidateAnswer with cited_region around §3.5.9 and refused: false; present its text and locator verbatim.

Fallback shape — apps/fmea/docs/ASSUMPTIONS_LOG.md RULE 1, quoting the AIAG & VDA FMEA Handbook (1st Ed., 2019) §3.5.9:

The use of a Risk Priority Number (RPN) threshold is not a recommended practice for determining the need for actions.

Nuance to preserve: the answer is "no". What the handbook offers instead is the Action Priority lookup table (RULE 7). This platform still flags rows above 100 points (Flag_High_RPN) — that cutoff is a legacy carry-over kept for backward compatibility with FMEA-4-based workflows, explicitly not the current handbook's recommendation. Say which is which; do not report the platform's flag as AIAG's rule.

3. MSA — the %GRR acceptance bands

User: "Is a %GRR of 12% acceptable under AIAG MSA?"

Endpoint shape — narrowed to MSA, a CandidateAnswer citing Table II-D 1 "GRR Criteria" with a page locator; present it as returned.

Fallback shape — apps/msa/docs/ASSUMPTIONS_LOG.md RULE 10 carries the band table from AIAG MSA (4th Ed.), Ch. II Sec. D, Table II-D 1: under 10 percent → Accept, "10 percent to 30 percent" → Marginal, over 30 percent → Reject. 12% falls in the middle band, so the answer is Marginal — acceptable for some uses, consider improvement — not a pass.

Nuance to preserve: it is not a pass/fail binary, and AIAG cautions against reading it as one (RULE 8):

The use of the GRR guidelines as threshold criteria alone is NOT an acceptable practice for determining the acceptability of a measurement system

So do not hand back "12% — accept" or "12% — reject" as a verdict without that caveat. Same theme as example 2: do not over-mechanize one number into a decision the standard declines to make.

4. SPC — capability indices need a stable process first

User: "Why do I need to check control-chart stability before trusting Cpk?"

Endpoint shape — narrowed to SPC, a CandidateAnswer citing the AIAG SPC Reference Manual (4th Ed., 2005) with its page locator; present it as returned.

Fallback shape — apps/spc/docs/ASSUMPTIONS_LOG.md RULE 7: capability indices assume the process is in statistical control, and computing them on an unstable process is misleading. The same precondition is independently verifiable in the NIST/SEMATECH e-Handbook §6.1.6 — "Process capability compares the output of an in-control process to the specification limits" (https://www.itl.nist.gov/div898/handbook/pmc/section1/pmc16.htm).

Nuance to preserve: the platform warns, it does not refuse. quality_core.spc.capability still returns the indices on an unstable process and annotates them as indicative only (RULE 7's note). Do not overstate it as a hard gate, and note that stable is tri-state — null means stability was not assessed, never "in control".

Reference

• references/mcp-tool-contract.md — the qdb_answer_question request/response shape, the refusal string, the fields the server will never return, the error contract, and the QDB_MCP_URL/QDB_MCP_TOKEN env contract.
• references/fallback-sources.md — the five-step local fallback procedure, the per-app ASSUMPTIONS_LOG.md table, and the quote-cap / cite-and-point rule from docs/CORPUS_LEDGER.md.
• references/project-context.md — the project-detection procedure, the full topic → artifact → field table, the read-only/no-compute rules, and the end-to-end worked "is my %GRR acceptable?" example against a loaded project.