Resources

Unstructured documents, starting with resumes, become governed, cited context through the platform’s one existing pipeline. Bringing them in was configuration on that pipeline; a new source type entered the same machinery. Engineering reference, 2026-08-13; sources: the resource-ingestion v0.2 design record, the landed build (extraction + tab live and verified internally), graded run results. Named Resources, 2026-08-13.

corpus 30 people, clean wave extraction graded vs ground truth writes proposals only, one queue surface live on the window customer link untouched pending approval

Used by · consumed by · outcome

used by	the ingest pipeline (documents in), the ratification queue (bound facts land here for a person to edit or refuse, never a gate on entry), and the operator surface where resource managers monitor extraction and problems
consumed by	the trial page’s Resources tab today; next stage, the engine’s answer path, where ratified document facts join operational records for questions like who has a skill and is available
outcome	a person-level record where every skill, tenure, and engagement claim cites its source line; extraction graded byte-exact on the five graded field classes over the 30-document corpus

1The pipeline: a document becomes a record with receipts

A resume enters as a raw document and leaves as facts bound into the same graph the customer’s operational systems feed. Every extracted value carries the verbatim source line it came from; a fact enters the record on arrival, with ratification open as a person’s permanent edit and refusal control, never a gate on entry; failures surface as their own rows.

Extraction writes facts that bind into the record on arrival. Ratification stays open as a person’s permanent edit and refusal control, the same lifecycle every other source uses, so provenance has one shape everywhere. The corpus’s ground-truth answer key is stripped before extraction, so extraction is graded blind.

2One extracted fact, anatomy

field	content	rule
field class	skills · roles held · employment tenure · education · certifications · project engagement history · leadership evidence · domains	closed set per wave; new classes enter as design rows
value	the extracted reading (e.g. “Test Automation, 8 years, last used 2026”)	graded against ground truth, byte-exact on the graded classes
from	the verbatim source line in the document	every fact cites its line; the provenance is the source text itself
state	ratified / open / problem	every state binds to the record on arrival; ratified/open/problem tracks review status, not entry; problems render as their own rows
key	content-derived per person + field class	re-running extraction updates the same fact in place

3Identity: clean first, dirt later

Sequencing is ruled: wave 1 is clean, well-formed documents with distinct people, so the pipeline proves working, stable, testable end to end. Only then come the dirt waves: name variants, messy layouts, hard identity cases, PII probes.

Identity posture (standing rulings): one identity implementation, estate-scoped memory; channels propose, and a person links a document to an operational record through the identity queue, on a human click.

4The surface: one mirrored tab, trust by visibility

Design decision: the surface lives in the trial, identical across roles — the internal team, the implementation partner, and the customer’s own operator. Rationale: “the customer’s operator will want to see the flow and monitor; the process never runs on trust in the machinery.” Role decides where writes land; every role sees the same page.

element	what it shows
strip	one slim line: 30 resumes · 29 extracted · 1 problem · 0 pending; live counts, matching ground truth at read time
rows	one collapsed row per person: name, corpus id, field count, state chip
expander	the full extraction table (field class / value / verbatim source / state), every row citing its line
problems	rows in the same list as everyone else, in reading order

5Measurement

Expected values come from the corpus’s ground-truth endpoint at test time; roles-held is additionally checked against ground-truth assignments. The grading rig is the same pass/fail shape the rest of the program uses; acceptance is the rendered screen, and the verification click-through fetches expected counts from the source system at run time.

The extraction step itself is machinery upstream of the declaration layer: an extraction model produces facts that bind into the record on arrival, with ratification open as a person’s ongoing edit and refusal control. The dictionary carries the same boundary.

6Where it runs, and where it goes next

Swapping the synthetic corpus for the customer’s real document sources changes the left box only. The pipeline, governance, and surface are already the production shape. That substitution is a later phase, gated on internal approval.

7State: built and next

built, verified	next, gated
clean 30-person corpus landed · extraction end to end, graded byte-exact on the 5 graded classes · proposal-queue write path with content-derived keys · the tab live and verified internally (30/29/1/0 verified against ground truth at click time) · problems surface · grading rig	dirt waves (name variants, messy layouts, PII probes) · partner-side questions in flight (supply API shape, source format, ratings location) · joins from document facts to operational records on the answer path · real-source intake through the connect fabric · shipping the tab to the customer-facing link, pending internal approval