Input

Spec Ingestion

The device spec is where every test program starts, and it usually arrives as a PDF. ATE·IQ turns a datasheet or a Jama Connect project into structured requirements — each with its limits, its forced condition, and a test number that stays stable for the life of the program. Everything downstream joins on those rows.

01

Overview

A requirement in ATE·IQ is not a paragraph of prose. It is a database row with typed fields — minimum, maximum, units, forced condition, test number — that the generator, the traceability reconcile and the STDF join can all read mechanically.

Spec detail page showing the TMP126 requirement table
Fig. 1 — The TMP126 spec: 34 requirements, each carrying limits, forced conditions and a test number. Provenance is shown per row.

Sources are whatever the project actually has. File upload accepts .pdf, .docx, .csv, .xlsx, .txt and .md; Jama Connect syncs over its REST API with stored credentials. Structured formats parse column-by-column with no LLM involved. PDF and DOCX go through LLM extraction — and every extracted row is tagged as such, so an engineer knows which numbers to audit before trusting them.

02

What a requirement carries

The fields below are the contract the rest of the platform builds on. All numeric fields are optional — not every requirement is a parametric test.

FieldExample (TMP126)Why it exists
testNumber 1100 The stable join key. Names drift between the spec, the program and the datalog; the test number is the contract that links a requirement to its IG-XL test and to PTR.TEST_NUM in production STDF.
specMin / specMax / units — / 220 µA The measured limit, normalised for comparison — 65 uA and 65e-6 A reconcile as equal downstream.
forceMode / forceValue / forceUnits I · −100e-6 · A The mandated stimulus, distinct from the measured limit. "Force −100 µA, measure clamp voltage" becomes structured data — so the reconcile can catch a program that measures the right thing under the wrong drive.
forcePin VDD Which pin the force applies to. Real programs apply supply force through the Pin Levels sheet, not a VBA literal — the reconcile reads the right row because the requirement names the pin.
testMethod measurement-path note Free-text method, including documented measurement-path compensations (a series termination, a Kelvin drop). The reconcile respects a documented compensation instead of flagging it as limit drift.
provenance llm-extracted Where each row's numbers came from: csv-column, xlsx-column, llm-extracted, jama-field, regex-fallback or manual. LLM-extracted rows are flagged for audit before trust.
DATASHEET .pdf · .docx · .csv · .xlsx JAMA CONNECT REST sync · stored creds PARSE columns direct · PDF via LLM TestRequirement row testNumber specMin/Max units forcePin forceMode/Value testMethod provenance 1100 — / 220e-6 A VDD typed fields one row per requirement — 34 rows for the TMP126 — each downstream surface joins on testNumber
Fig. 2 — Ingestion flow: any source format reduces to the same typed requirement row.
03

How it works

  1. Upload or sync.

    Drop a file on the project's spec page, or connect Jama Connect once — the server URL and credentials are stored encrypted, and cleartext HTTP endpoints are refused.

  2. Parse by format.

    CSV and XLSX parse column-to-field with no LLM hop. PDF and DOCX are chunked and sent through LLM extraction. Plain text and Markdown fall back to a qualitative regex parser that deliberately extracts no numerics — a degraded parse is labelled degraded, not dressed up.

  3. Report the extraction honestly.

    The upload result states the method used, how many characters were read versus how many the document contains, and whether anything was truncated. Zero requirements extracted is a warning, not a silent empty table. The same pass extracts a device profile — pin names and rails — which the program generator later requires before it will run.

  4. Review with provenance.

    Each row shows where its numbers came from. LLM-extracted limits are tagged for audit; hand-edits are recorded as manual. Corrections made in the review UI persist on the row.

04

On the TMP126 demo

The TMP126 datasheet — a ±0.25 °C digital temperature sensor with a 3-wire SPI interface — ingests into 34 requirements: DC parametrics, SPI protocol checks, register read/write tests and temperature-accuracy points. Each carries its limits and forced condition; each has a test number that the generated program, the load board and the production STDF all join back to.

Two rows are worth showing a customer. The supply-current requirement carries a forced condition, so the reconcile verifies not just the limit but the drive. And the SIO output-level requirements carry a testMethod documenting the 33 Ω series-termination drop at the tester — which is why the traceability dashboard reports that compensation as documented rather than as limit drift.

Honest boundary Numbers extracted from a PDF by an LLM are tagged llm-extracted and should be audited before they gate a compliance verdict. Structured columns and Jama fields carry their own provenance tags. Truncated reads are reported as truncated — the tool never presents a partial extraction as a complete one.