TGA-SOURCE-ATLAS-01 — training.gov.au as a data source

Drafted 2026-08-20 (Architect seat, Cowork, Fable, session 33), on Tim's ask of the same day. Status: APPROVED by Tim 2026-08-20 as a working reference — v1, from skeleton 95db533c47f3e763 / 12,180 B. A LIVING DOCUMENT, kept at hand: Part one is the stable plain-language map, amended by dated addition; Part two is a named structure whose cells only TGA-MIRROR-CORPUS-AUDIT-01 can fill, and it is generated by instrument, never hand-edited. Home: docs/docs/ops/tga-source-atlas.md, beside substrate-findings.md, subordinate to it on any figure conflict.

What this document is: the map of the source — what TGA looks like as a data pot, corner to corner, in language a person can hold. Written so that a future return to this territory (fault test, repair, addition) tools up to a known floor in one read.

What this document is never: a description of our programme. State, briefs, gates and priorities live in the anchor and the TMs. The moment this document grows a "what we're doing about it" section, it has become a second anchor, and both die of it.

Standing rule inherited from the anchor: a line here is not a measurement. Every figure carries its as-at and source. The coverage table (Part two) is the only part that claims currency, because an instrument regenerates it.

---

PART ONE — THE ATLAS (changes rarely; edit by dated addition)

1 · What TGA is

training.gov.au is the Australian Government's single national register of vocational education and training — every training product (qualifications, units, skill sets, accredited courses) and every registered training organisation, current and historical, for the whole country. There is no second register: even RTOs regulated by the two state regulators (VRQA in Victoria, TAC in WA) appear here. It is public, openly licensed, and served by an open REST API as well as the website.

One sentence to hold: TGA publishes facts; it does not publish judgements. What a fact means for compliance (may this RTO enrol a student in this product today?) is never in the data — that is what the rules engine (SEQ-MODEL-01) exists to compute.

2 · The surfaces

TGA offers its data through four distinct surfaces. A completeness claim must name which surface it ranges over.

1. The REST API (/api/…) — the primary machine surface. Returns JSON. Takes OData filters (server-side population selection). Publishes its own contract as a Swagger specification — the source's own enumeration of every endpoint it offers. The Swagger capture and diff-watch is TGA-UPSTREAM-INTELLIGENCE-CAPTURE-01; not yet built, as at 2026-08-20.
2. The website pages — human-rendered views over the same API. Matter to us twice: as the render target (Radar recreates the component page faithfully), and as the second enumeration instrument — the endpoints the site's own traffic calls is the usage-side denominator.
3. The reports page — periodic published exports. A named tripwire surface; not a fetch family.
4. The documents — the files behind each release of each component (the product as PDF, DOCX and XML; assessment requirements; complete packages), plus companion volumes (CVIGs). Reached via the release-detail endpoint, which is both the version record and the per-release asset manifest (measured 2026-08-20; see Part two). Added 2026-09-04: TGA's own documents are not held (Tim, 2026-09-03). The parsed rows plus the round-trip test are the mirror of that content, so the files are a render, not a gap. One exception, ruled 2026-09-04: 951 releases publish a document and no machine-readable content bundle, so for those the round trip has no rows to run against; those 951 documents are fetched and kept as bodies (RULING-ASSET-ONLY-RELEASES-01).

Adjacent sources that are NOT TGA and are treated as their own families with their own captures: legislation.gov.au (the era instruments — the law the rules engine cites), ASQA's site (the transition-extensions register, captured 2026-08-19), and the state regulator portals (TAC, VRQA — named future families, parked). These live in regulatory-sources-01, never in the TGA pith stores.

3 · The entity families, in plain language

• Training components — the products. 125,996 in the register of record (frozen population, 2026-08-15). Six types: qualification, unit, skill set, accredited course, accredited short course, module. Each carries its lifecycle status, classifications, and the supersession graph — which products replaced which, and whether the replacement was equivalent. Detail bodies exist per component and per release.
• Organisations — 13,146 records, of which 13,106 are RTOs (register body, 2026-08-15; the 40 non-RTOs are regulators, training-package developers and state authorities — extra data, not noise). The detail bodies carry the full corporate history inline: every address, contact, legal name, trading name, registration period and regulatory decision, each with start/end dates. The history is in the body — it does not need to be assembled from snapshots. (Parsed into 18 tables / 420,949 rows at A1-ORG-PARSE-01, 2026-08-20.)
• Scope — what each RTO may deliver/assess, per component, with start/end dates, an explicit/implicit flag and an extent ("Deliver and assess" vs "Assess only"). ~94% of scope is implicit — units flowing from on-scope qualifications — and TGA publishes the implicit lists rather than leaving them to be derived (sample measurement, n=200 bodies, 2026-08-20).
• Delivery notifications — where an RTO actually notified delivery of a component; bitemporal (the date notified and the date effective can differ by years). The largest family by volume (millions of pairs).
• Releases and documents — each component has numbered releases; each release lists its document assets with name, type, size and publish date. ~104,737 (component, release) pairs; the payload behind the manifests is ~225 GB (mean-basis, n=24, re-run owed — see Part two caveats).
• Classifications and taxonomies — ANZSCO, ASCED, industry/occupation taxonomies, and the scheme value dictionaries (7 endpoints, 0.40 MB — the full code lists including codes no component uses).
• Metadata — one endpoint publishing dataExportSynchronisationDateTime plus a build version: the source's only change signal, and it is global, not per-object.

Added 2026-09-04 — the mirror is two mirrors, and only one can answer a question about a resource. pith-assets is keyed <family>/<sha256 of the body>.json: content-addressed, carrying no resource id, verified on 36 objects across 12 families with 0 differing. It cannot be asked "is there an object for release X". The content mirror's own listing (tga-content-ed1-2026-08-28/_LISTING.jsonl, 105,911 rows over 75,156 distinct release_id) can. Naming which mirror is being counted, before counting, is the whole of that figure's reliability.**

4 · Temporal properties — the four facts that shape everything downstream

1. No per-object change signal. No ETag, no Last-Modified. One global export-sync stamp. Re-sync designs poll the stamp; per-object hash comparison does not work (identical data re-serialises with different bytes — measured).
2. History is embedded, not versioned. The detail bodies carry dated intervals inline; TGA does not publish "the body as it was last year". What we captured on a date is the only record of how it stood that date — which is why capture custody (the pith stores) exists.
3. The register is bitemporal where it matters. Notification date ≠ effective date on delivery facts (observed lag up to five years). Rights are evaluated on effective dates; visibility questions use notification dates.
4. All dates are register-published calendar dates. No timezones, no instants. Measured ISO-8601 across 624,491 values with zero exceptions (org family, 2026-08-20) — but that is one family's measurement, not a source-wide law; each parse re-measures.

5 · The honest limits

• TGA withholds by policy as well as by absence. One endpoint answered 403 "Completion mapping is internal use only" on 2026-08-20. A refusal is a real answer and is recorded as one. Corrected 2026-09-04: that refusal is falsified on the bodies. The 28 August build holds 2,575 completion and 6,403 completionusage bodies — 8,978 in all, every one a JSON array of components, zero error bodies. Whatever answered 403 in August, what is on disk is data. A refusal is an observation with a date on it, never a property of an endpoint — and it is re-tested before it is carried forward.
• Every claim is bounded by "what TGA publishes", never "what TGA knows."
• The source is not perfectly clean. Known warts, measured: inconsistent date formats in ASQA's register (2/4-digit years, month-name variants, at least one malformed value); TGA field-name traps (registrationManager on organisations vs recognitionManager on components — two fields, not a rename); status vocabularies that collapse distinct states when read by .name instead of .id. The full trap list lives in substrate-findings.md; this atlas names that they exist so nobody assumes cleanliness.
• No measured instance of TGA losing an artefact currently exists (the earlier sole-survivor claim was falsified 2026-08-19). Archive protection arguments rest on prudence, not on TGA fragility.

6 · How to read a "do we have X" answer

The three-states frame governs: A — in the corpus (ledgered, closure-proven; trust it) · B — only in the prototype's rto-nrt-db (no ledger, no closure; volume signal only, never a figure) · C — never fetched. Part two ranges over the API surface and states A/C per family; B exists only as history.

---

PART TWO — COVERAGE (generated by instrument; hand-editing is a defect)

Owner instrument: scripts/ledger-01-emit-v1.py e4fdf55fc452a74d, emission outputs/ledger/TGA-MIRROR-COMPLETENESS-LEDGER-01-2026-09-04.md 4bd17526896a44b4, as-at 2026-09-04T10:04:22.

1 · The seven states

The seed had six. held-bodies is the seventh, ruled at the Gate 1 verdict §2.1 because the stores needed it: a family whose objects are on disk and whose rows do not exist is not owed — nothing has to be fetched for it — and it is not in. The seed called those families owed, which is what a ledger without an instrument does.

state	means
in	rows in a local store of record, keyed and digested, gate-certified
in flight	a brief is running or filed against it today
held-bodies	(new) fetched, in the mirror, no rows and no parse
owed	TGA serves it, nothing holds it as rows or bodies, no brief filed
fetch-dependent	cannot close without contacting TGA
render	a TGA surface re-serving data held elsewhere; the round-trip test is the mirror
not TGA's / refused	outside the mirror by definition, or TGA answers 403

The rule table, as applied — every state below comes from one of these and the rule id is printed with it:

rule	state	assigned when
S1	render	the path's class in the Gate 1 store is render
S2	not TGA's	the path's class is not TGA's data
S3	in flight	rows exist AND a brief is filed and running against it today
S4	in	rows exist in a local store of record, keyed and digested
S5	held-bodies	no rows, and >0 objects for it in a mirror listing (verdict §2.1)
S6	fetch-dependent	no rows, no bodies, and it cannot close without contacting TGA
S7	owed	no rows, no bodies, no brief — TGA serves it and nothing holds it

⚠ refused has no members, and that is a measurement

The seed §8 counts refused 2 — /training/{code}/completion and /completionusage, on the atlas's 2026-08-20 record of a 403 "internal use only". Gate 2 was told to read one body of each. It read every body of both:

family	bodies	error/dict bodies	empty arrays
completion	2,575	0	1
completionusage	6,403	0	1

All 8,978 are JSON arrays of components. Zero are error bodies. Verbatim head of one of each:

completion/       [{"code":"VBQM230","endDate":"2012-12-31","hasPreRequisites":false,"isMandatory":false,"links":[{"rel":"training-component","href":"https://training.gov.au/api/training/vbqm230"}],"startDate":"2009-08
completionusage/  [{"code":"10183NAT","id":"e230ecba-e5c1-4a6a-87b6-991d24fff873","status":"nonCurrent","statusLabel":"Non-current","title":"Diploma of Education","type":"accreditedCourse"},{"code":"10326NAT","id":"2c6

Whatever answered 403 on 20 August, what is on disk from the 28 August build is data. The two paths are data paths in held-bodies; the refused state has no members today. This corrects the seed §8 and the Gate 1 verdict §2.2's "47 + 2 refused": the reconciliation stands, but the two are not refused.

2 · The denominator, and the seed's own arithmetic reconciled

88 paths across the eight captures, each at the digest the brief asserts. By class, from the Gate 1 store:

class	paths
data path	47
render	35
not TGA's data	4
not TGA's / refused	2

The seed §8 says "of the 88, 39 are renders or not TGA's data (12 EXPORT · 8 SEARCH · 3 REPORT · 1 FEEDBACK · 3 security-role paths · 10 export paths inside TRAINING/ORGANISATION · 2 files paths)". This instrument: render 35 + not TGA's data 4 = 39. The seed reconciles exactly, item by item.

and "of the remaining 49 data paths" — this instrument: 47 data paths + 2 classified refused = 49. The seed reconciles exactly. With the refusal falsified above, the 49 are data paths outright and the seed's total was right all along; only its last two labels were not.

3 · NRT — training products

path	population	held	state	rule	source
/training/{code}	125,838 components (125,848 rows)	component · release · supersession_edge · component_parent in e2-nrt-01	in	S4	Gate 1 item 2
— its usageRecommendation and classifications	125,838	no column exists in e2-nrt-01.component, and no taxonomy table	held-bodies	S5	Gate 1 item 2; bodies at classification/ 30,674
/training/{code}/releases	125,838 components	87,099 carry ≥1 release; releases-list/ holds 87,099	in	S4	Gate 1 item 3
/training/{code}/releases/{n}	103,327 releases	103,320 level-2 bodies (100.0%), 7 without	held-bodies	S5	Gate 1 item 6
…/document-bundle · /content/bundle/{id} · /content/item/{id}	103,327 releases	781,817 sections indexed; bodies in e2-content-01	in for fetched releases, fetch-dependent for the rest — §5	S4/S6	Gate 1 items 4, 5
…/unitgrid	12,902 walked	grid 10,269 · empty 2,633; join 12,902/12,902	in flight — Q3	S3	Gate 1 item 7
…/components · …/components/qualifications	931 training packages	release-components/ 929 bodies	held-bodies	S5	Gate 1 item 9
…/files · …/files/zip	—	—	render — Tim ruled 2026-09-03 the PDFs are not held	S1	RULINGS-2026-09-03-TAPPED-THREE §4
/training/{code}/classification	125,838 × n	classification/ 30,674 bodies, no rows	held-bodies	S5	Gate 1 item 9
/taxonomyindustrysectors · /taxonomyoccupations	2 vocabularies	taxonomy-industry 1,244 · taxonomy-occupation 3,103	held-bodies	S5	Gate 1 item 9
/training/{code}/prerequisites	unmeasured	prerequisites/ 2,285 bodies; the content-side reference rows are in	held-bodies	S5	Gate 1 item 9
/training/{code}/delivery	unmeasured	delivery/ 89,738 bodies	held-bodies	S5	Gate 1 item 9
/training/{code}/unitgridusage	unmeasured	unitgridusage/ 27,156 bodies	held-bodies	S5	Gate 1 item 9
/training/{code}/recognitionmanager · /restrictions	unmeasured	comp-restrictions/ 3; no family for recognitionmanager	held-bodies / owed	S5/S7	Gate 1 item 9
/training/{code}/completion · /completionusage	unmeasured	completion/ 2,575 · completionusage/ 6,403 bodies, all data, none refused	held-bodies	S5	Gate 2 §1
/releases/{releaseId}/training	derivable	release.release_id_norm join	in (derived)	S4	model §1

⭐ The register forgot to give 38,739 components a release. 87,099 of 125,838 distinct codes carry a release row; the 38,739 that do not are the accredited families — accreditedCourse 19,439 and accreditedUnit 19,302 rows, zero releases between them, with a positive control run before the zero was believed. Their L1 register bodies ARE held: register-accredited-course-L1 19,439 · register-accredited-unit-L1 12,789. This is a build gap, not a fetch gap.

4 · NRT content — the section types, parsed into rows

781,817 sections held over 36 content type codes. Held vs parse rows, per code, from the store of record:

code	sections	parse rows	no-section rows	open
0012	104,754	104,754	0	0
0118	58,537	58,542	5	0
0103	58,494	58,494	0	0
0102	52,695	52,695	0	0
0121	46,470	46,470	0	0
0011	38,629	3,680	2	34,949
0200	30,985	30,985	0	0
0113	30,502	31,808	1,306	0
0120	30,500	31,806	1,306	0
0104	30,497	31,804	1,307	0
0112	30,476	0	0	30,476
0100	28,048	0	0	28,048
0111	28,043	0	0	28,043
0124	28,039	0	0	28,039
0125	27,995	0	0	27,995
0108	27,993	0	0	27,993
0119	26,645	0	0	26,645
0000	24,112	0	0	24,112
0001	15,745	19,425	90	0
0123	11,103	0	0	11,103
0116	10,077	0	0	10,077
0117	10,069	10,069	0	0
0110	9,604	9,604	0	0
0126	5,454	5,454	0	0
0127	5,452	5,462	10	0
0128	5,451	5,462	11	0
0109	4,620	0	0	4,620
0107	481	481	0	0
0010	74	0	0	74
0201	64	0	0	64
0002	48	0	0	48
0105	48	0	0	48
0122	47	0	0	47
0106	46	0	0	46
0202	10	0	0	10
0203	10	0	0	10

0011 open: 34,927 releases — the largest single open figure in the store, reproducing the BRIEF-TGA-MODEL-02 Gate 5 close.

5 · Fetch-dependent — and it is SEVEN releases, not twenty-eight thousand

28,171 releases of 103,327 have no row in the store of record's section. The first emission of this ledger called all 28,171 of them "never fetched". That was wrong in kind, and the partition below is why.

cell	releases
level-2 body held, contentBundles NON-EMPTY, and sections in the store	75,156
level-2 body held, contentBundles non-empty, but no sections	0
level-2 body held, contentBundles EMPTY, but sections in the store	0
level-2 body held, contentBundles EMPTY, no sections — a SERVED ABSENCE	28,164
in the store with no level-2 body held	0
no level-2 body at all — ⭐ THE FETCH LIST	7

A four-cell partition with zero exceptions. 28,164 releases carry a level-2 body in which TGA itself serves contentBundles: [] — the release-grain case of declared-absent. Nothing has to be fetched for them and nothing is missing. The fetch list is 7 releases.

Three counter-hypotheses were tested against the served-absent population and all three fail. They are not merely older releases of otherwise-bundled components: 28,163 of 28,164 belong to components where every release is empty. TGA does not hold the content behind a document instead: 100.0% of bundled releases list a document asset against 3.4% of these. And they are not current — measured against TGA's own usageRecommendation, held in the L1 register bodies for 87,099 codes, 0 of 28,164 belong to a component TGA marks current (61.6% deleted, 38.4% superseded), where the same predicate finds 22,715 current components on the bundled side. Cited, not recomputed here: outputs/canon-01/CANON-01-FETCHGAP-TEST.txt 7ff788ce77a1a864 and outputs/canon-01/CANON-01-GATE2-CH3-USAGEREC.txt.

⚠ 951 of the served-absent releases list a document asset — TGA publishes a downloadable document and no machine-readable bundle, so for those the round-trip test has no rows to run against. RULING-ASSET-ONLY-RELEASES-01 (Tim, 2026-09-04): those 951 documents are fetched and kept as bodies in R2, state held-bodies, no parse briefed. They are the second fetch-list line.

⚠ Served-absent is as-at the capture, not forever. Every figure above is measured on what TGA served on 28 August 2026. Whether a fresh pull still returns [] is a question only the refresh pull can answer.

The type table below ranges over the served-absent population and is retained for its shape; its own population is named in outputs/canon-01/CANON-01-GATE2-FANOUT.txt.

type	current	replaced	total
qualification	2,805	0	2,805
skillSet	0	2	2
trainingPackage	129	0	129
unit	25,234	1	25,235
TOTAL	28,168	3	28,171

The 25,235 is confirmed as a number and retired as a sentence. The Time Machine carried "25,235 sectionless current unit releases" and no filed artefact held it. Measured: unit releases with no section row = 25,235 exactly, but 25,234 current and 1 replaced. The figure was right; its population name was wrong by one row.

Controls, both reproducing: qualification releases with no 0116 2,835 (Q1's 2,835) · with no section at all 2,805 (-01's 2,805).

⚠ Which mirror answers this. pith-assets keys are <family>/<sha256 of the body>.json — content-addressed, carrying no resource id, verified on 36 objects across 12 families with 0 differing. It cannot be asked "is there an object for release X". tga-content-ed1-2026-08-28/_LISTING.jsonl can: 105,911 rows over 75,156 distinct release_id. Naming the path before counting from it is the whole of this row's reliability.

6 · RTO — organisations

⚠ Two databases answer to e2-rto-01 and only one is the register. The .local-stack store bound to that name holds 3 qualifications and 140 units — a prototype shape with no organisation table and no scope_entry. Every figure here is from ~/mirror/e2-stores-2026-09-01/e2-rto-01.sqlite.

path	population	held	state	rule
/organisation/{code} and its eleven tabs	13,034 organisations	organisation + 17 od_* families; org-L1/ 13,034 bodies	in flight — Lane Z	S3
/{code}/scope · /scopesummary	6,526,449 scope rows	6,526,449; scopesummary/ 3,674 bodies	in flight — Lane Z	S3
/{code}/deliverynotificationhistory/{trainingcode}	the largest family	dnh/ 312,010 bodies, no rows, and the model names no table	held-bodies — was owed in the seed	S5
/{code}/training-packages	unmeasured	org-training-packages/ 5,920 bodies	held-bodies	S5
/organisation-security-roles	—	—	not TGA's	S2
/api/security-roles	—	1 body, metadata family	not TGA's — declared by the METADATA swagger; TGA's own access-control vocabulary, not register data (C-0912-02)	S2
2 × scope/export/…	—	—	render	S1

7 · held-bodies — what is on disk and has never become a row

This section did not exist in the seed, because the state did not. Every family below was fetched by the E2 build and holds objects today; none has rows.

family	objects	coverage measurable locally?
dnh	312,010	no — the body names its contents, never its subject
releases	103,320	yes — the body names its subject (id)
delivery	89,738	no — the body names its contents, never its subject
releases-list	87,099	yes — the body names its subject (id (of each release))
register-unit-L1	56,725	not examined at this gate
classification	30,674	no — the body names its contents, never its subject
unitgridusage	27,156	no — the body names its contents, never its subject
unit-L1	24,913	yes — the body names its subject (code + id)
org	24,110	not examined at this gate
register-accredited-course-L1	19,439	not examined at this gate
org-L1	13,034	yes — the body names its subject (code + organisationId)
register-accredited-unit-L1	12,789	not examined at this gate
register-qualification-L1	6,870	not examined at this gate
completionusage	6,403	no — the body names its contents, never its subject
org-training-packages	5,920	not examined at this gate
register-skill-set-L1	3,679	not examined at this gate
scopesummary	3,674	no — the body names its contents, never its subject
taxonomy-occupation	3,103	no — the body names its contents, never its subject
completion	2,575	no — the body names its contents, never its subject
qual-L2	2,345	not examined at this gate
qual-L3	2,345	not examined at this gate
prerequisites	2,285	no — the body names its contents, never its subject
unitgrid	2,156	not examined at this gate
taxonomy-industry	1,244	no — the body names its contents, never its subject
training-L1	1,174	not examined at this gate
release-components	929	yes — the body names its subject (releaseId)
cv-list	337	not examined at this gate
register-training-package-L1	259	not examined at this gate
metadata	12	not examined at this gate
comp-restrictions	3	not examined at this gate
cricos	1	not examined at this gate
search-organisation	1	not examined at this gate
search-training	1	not examined at this gate

⚠ The Gate 1 verdict §5's least-sure, answered, and the answer is a split, not a yes or a no. Five families self-identify and their coverage is measurable locally today: releases, unit-L1, org-L1, release-components, releases-list. The rest name their contents and never their subject — a delivery body lists the RTOs delivering a product without saying which product; links is empty on every one read. For those, coverage needs the fetch ledger (pith-index-01, cloud), which is therefore a dependency for some families and a check for others, not a blanket dependency.

8 · The number, honestly — and the seed graded

The seed's §8 gave three counts over the 49 data paths, hand-derived: in or in flight 22 · owed 24 · fetch-dependent 1 (plus refused 2). Every one of the 49 is assigned a state below by the rule table, from the evidence named beside it, and the counts are SELECTed from that assignment — not typed.

path	evidence	rule	state
/api/content/bundle/{id}	store of record.section	S4	in
/api/content/item/{id}	store of record.section	S4	in
/api/classification-purpose	—	S7	owed
/api/classification-purpose/{code}	—	S7	owed
/api/metadata	—	S7	owed
/api/nrt-classification-scheme	—	S7	owed
/api/nrt-classification-scheme/{code}	—	S7	owed
/api/nrt-classification-scheme/{code}/values	—	S7	owed
/api/nrt-classification-scheme/{schemeCode}/values/{code}	—	S7	owed
/api/rto-classification-scheme	—	S7	owed
/api/rto-classification-scheme/{code}	—	S7	owed
/api/rto-classification-scheme/{code}/values	—	S7	owed
/api/rto-classification-scheme/{schemeCode}/values/{code}	—	S7	owed
/api/organisation/{code}	e2-rto-01.organisation; Lane Z	S3	in flight
/api/organisation/{code}/addresses	od_addresses	S3	in flight
/api/organisation/{code}/classification	od_classifications	S3	in flight
/api/organisation/{code}/contacts	od_contacts	S3	in flight
/api/organisation/{code}/cricoscode	od_cricos_codes — 0 rows, served empty by TGA; the emptiness is the held fact	S4	in
/api/organisation/{code}/deliverynotificationhistory/{trainingcode}	dnh/ 312,010 bodies	S5	held-bodies
/api/organisation/{code}/legalname	od_legal_names	S3	in flight
/api/organisation/{code}/registration	od_registrations	S3	in flight
/api/organisation/{code}/registrationmanager	od_registration_managers	S3	in flight
/api/organisation/{code}/regulatorydecision	od_regulatory_decisions	S3	in flight
/api/organisation/{code}/restrictions	od_restrictions	S3	in flight
/api/organisation/{code}/role	od_roles	S3	in flight
/api/organisation/{code}/scope	e2-rto-01.scope_entry	S3	in flight
/api/organisation/{code}/scopesummary	scopesummary bodies; Lane Z	S3	in flight
/api/organisation/{code}/tradingname	od_trading_names	S3	in flight
/api/organisation/{code}/training-packages	org-training-packages/ 5,920 bodies	S5	held-bodies
/api/organisation/{code}/webaddress	od_web_addresses	S3	in flight
/api/releases/{releaseId}/training	e2-nrt-01.release (derived)	S4	in
/api/training/{code}	e2-nrt-01.component	S4	in
/api/training/{code}/classification	classification/ 30,674 bodies	S5	held-bodies
/api/training/{code}/completion	completion/ 2,575 bodies	S5	held-bodies
/api/training/{code}/completionusage	completionusage/ 6,403 bodies	S5	held-bodies
/api/training/{code}/delivery	delivery/ 89,738 bodies	S5	held-bodies
/api/training/{code}/prerequisites	prerequisites/ 2,285 bodies	S5	held-bodies
/api/training/{code}/recognitionmanager	—	S7	owed
/api/training/{code}/releases	e2-nrt-01.release	S4	in
/api/training/{code}/releases/{releaseNumber}	releases/ 103,320 bodies	S5	held-bodies
/api/training/{code}/releases/{releaseNumber}/components	release-components/ 929 bodies	S5	held-bodies
/api/training/{code}/releases/{releaseNumber}/components/qualifications	release-components/ 929 bodies	S5	held-bodies
/api/training/{code}/releases/{releaseNumber}/document-bundle	store of record.section for fetched releases; the rest is the fetch gap	S6	fetch-dependent
/api/training/{code}/releases/{releaseNumber}/tpusage	—	S7	owed
/api/training/{code}/releases/{releaseNumber}/unitgrid	Lane Q's walk; Q3 loads it	S3	in flight
/api/training/{code}/restrictions	comp-restrictions/ 3 bodies	S5	held-bodies
/api/training/{code}/taxonomyindustrysectors	taxonomy-industry/ 1,244 bodies	S5	held-bodies
/api/training/{code}/taxonomyoccupations	taxonomy-occupation/ 3,103 bodies	S5	held-bodies
/api/training/{code}/unitgridusage	unitgridusage/ 27,156 bodies	S5	held-bodies

the seed said	this emission says, derived	why it moved
in or in flight 22	21 (in 6 · in flight 15)	the register, content and RTO rows
owed 24	owed 13 · held-bodies 14	the seventh state: 14 of the seed's owed paths have their bodies on disk already and need a parser, not a fetch
fetch-dependent 1	1, and now with a size: 7 releases (was reported as 28,171 in the 2026-09-03 emission)	the seed named the row; the instrument counted it; the partition corrected it — 28,164 of the 28,171 are served absences, not a fetch
refused 2	0	falsified on 8,978 bodies, §1
—	total 49 of 49 data paths	every data path has a state

The honest sentence the seed was reaching for: it is not that half the data paths are owed. It is that 14 of 49 (29%) are already on disk and waiting for a table — a parser job, not a fetch — one is a fetch of known size (28,171 releases), and only 13 have nothing held at all, of which 10 are the small classification-scheme and classification-purpose vocabularies.

The post-14-September list, in order

1. The fetch list — 7 releases with no level-2 body held. Not 28,171: the other 28,164 are served absences (§5). 1b. The 951 asset-only releases — a document published with no content bundle (RULING-ASSET-ONLY-RELEASES-01). 1c. (retired figure, kept for the audit trail) the former "fetch gap" of 28,171 releases with no section row, 28,171 of them recorded as never fetched: unit 25,235 · qualification 2,805 · trainingPackage 129 · skillSet 2.
2. The fetch-ledger export — one read of pith-index-01 to a local file under SPEND-ENVELOPE-01, which turns every held-bodies row whose bodies do not self-identify into a coverage figure. Ruled at the Gate 1 verdict §2.1.
3. The refresh pull, the /metadata sentinel, and a refresh policy per family — seed §7, none built.

Before any of that, and needing no fetch at all

1. Accredited-course and accredited-unit releases from their held L1 bodies — 38,739 components with no release row.
2. The level-2 release fields — currencyChangeDate · assets[] · externalLinks[] · specializations[] · packagingInformation — from 103,320 bodies already on disk, into tables the model must first name.
3. Classification and taxonomy per component from 30,674 + 4,347 bodies.
4. Delivery notifications — 312,010 dnh and 89,738 delivery bodies, into a table the model must first name.

---

Emitted 2026-09-04T10:04:22 by ledger-01-emit-v1.py e4fdf55fc452a74d. Every cell is a row of the Gate 1 store. To regenerate: re-run Gate 1's instrument, then this one.

Regeneration rule: the audit emits this table whole, with its own as-at, and the emission replaces the section between the PART TWO header and this line. A hand edit to any cell is a defect of the same class as back-filling provenance.

---

Status and open items

1. ~~Tim reads Part one for domain wrongness.~~ DISCHARGED — approved 2026-08-20.
2. OPEN: the audit brief (TGA-MIRROR-CORPUS-AUDIT-01) names this document's Part two as a deliverable when it is written. Until the audit runs, Part two is seeded knowledge with no completeness claim.
3. ~~Home and hierarchy ruling.~~ DISCHARGED — ruled 2026-08-20: home as in the header; subordinate to substrate-findings on any figure conflict (findings are the measurement record; this is the map drawn from them).

Least sure

Whether Part one's plain-language register stays plain under maintenance. Every editor of this file will know more than its intended reader, and the drift pressure is one-directional — toward the jargon of whoever last touched it. The cure attempted here is structural (the volatile parts are machine-generated, so humans mostly never need to edit), but Part one has no such guard. What would make me wrong about the document's whole value: if in six months it reads like substrate-findings, it has failed at the only thing it exists for.

— Claude (Architect seat, Cowork, Fable, session 33), 2026-08-20. Approved by Tim the same day; a working reference, amended by dated addition.