CVIG-LICENCE-CENSUS-01 — REPORT — 2026-08-14¶

Seat: Alex (bridge) · Lane: dead-window recon · Blast radius: zero. Read-only throughout. No request to training.gov.au or any TGA host. No PUT/POST/PATCH/DELETE against any Cloudflare resource. Bytes read from R2 companion-volumes-ref (Prototype account, permitted morgue read) via op://RTOpacks/cloudflare-api-token, id 76eb6809…. No token minted.

No conclusions of law are drawn. Classification and verbatim quotation only.

Register: outputs/CVIG-LICENCE-CENSUS-01-REGISTER-2026-08-14.csv sha256 dd42c032e7e634c9909b8122240c567b7b82e11d1f167fa25af77f29d2ffbf66 · 964 rows × 18 columns.

1 · POPULATION — RECONCILED BEFORE ANYTHING WAS READ¶

Measure	Enumerated	Certificate	Delta
Objects in bucket	563	—	—
— content files	562	562	0
— acquisition manifest (`_manifest/2026-08-08/….jsonl`)	1	not counted	+1, explained
Bytes, content files only	1,001,808,227	1,001,808,227	0

Set-difference against the acquisition manifest, both directions: 0 and 0. Every bucket object has a manifest record; every manifest record has a bucket object. The census proceeds over the full 562 with no residual.

Format split reconciles, with one cosmetic difference: bucket keys carry 509 .pdf where the certificate records .pdf 499 + .PDF 10. Same 509 files; case was normalised on ingest. .xlsx 36, .docx 16, .zip 1 — all exact.

file_category: Companion Volume 501 · Companion Volume Supporting Material 61. state_advertised: Current 525 · Replaced 37. Packages 51. Developers 11.

The .zip was opened and its members censused individually — MEA/d066508e…/fe298e82….zip, 403 members, each listed in the register under the parent key with ::member appended.

Register population: 964 rows = 561 standalone files + 403 zip members. (561 not 562: the zip parent is represented by its members, not by itself.)

2 · INSTRUMENT — DEMONSTRATED, AND IT FAILED TWICE FIRST¶

The brief's nominated control does not exist in this corpus. HLT Release 11.0 is not present — the highest HLT release held is 10.0. This is not a gap in the acquisition: CVIG-LANDSCAPE-RECON-01 records HLT R11.0 as HumanAbility-hosted (JSC), and CVIG-ACQUIRE-01 states all 562 files are TGA-hosted. The control is structurally absent by design. Reported as a finding, not fetched.

Substituted controls, named by CVIG-LANDSCAPE-RECON-01-ADDENDUM-01 as carrying explicit CC BY-NC-SA 4.0: CHC R11.1 IG, HLT First Aid CV, HLT Work Placement CV.

First attempt: 0/3. The extractor could not find a notice the record says is present. Halted and diagnosed rather than proceeding. Two distinct defects, both mine:

The full-document fallback never fired. It triggered only when the hit list was empty, and a stray licence in body prose (occupational-licensing text, ubiquitous in these documents) suppressed it. Fixed: the fallback now triggers unless a licence-class term hits.
Glyph-corrupted PDFs defeated literal matching. CHC R11.0's text extracts as Modiﬁca/on, informaGon, and critically Crea9ve Commons. Fixed with a tolerant crea\w{0,3}ve\s+commons pattern plus a licen[cs]\w{0,3}\s+under term, and a per-file glyph-corruption counter. One file exceeds 50 corruption markers (CHC R11.0, 947).

After the fixes the extractor demonstrably finds a known-present notice — CHC R11.1 returns "© Commonwealth of Australia. Licenced under Creative Commons 4.0", and CHC R11.0 reclassified from copyright-no-licence to CC-BY-4.0 once the mangled spelling was matchable.

Method. PDFs: pdftotext over the first 6 and last 6 pages, escalating to the whole document when no licence-class term hits. OOXML: full text of document/sharedStrings/core/app/header/footer parts. Terms searched, with which one hit recorded per file: CC BY-NC-SA · CC BY-SA · CC BY-NC · CC BY-ND · CC BY · NonCommercial · ShareAlike · NoDerivatives · Creative Commons · all rights reserved · Commonwealth of Australia · Copyright · © · licence/license · "licenced under".

3 · LICENCE CLASS — COUNTS¶

Population: the 561 standalone files.

Class	Count
`no-notice-found`	334
`CC-BY-NC-SA`	122
`other-explicit`	52
`copyright-no-licence`	37
`CC-BY-4.0`	15
`CC-BY-SA`	1
`not-machine-readable`	0
Total	561

Population: the 403 zip members. no-notice-found 396 · not-machine-readable 7.

All 7 not-machine-readable files are inside the MEA zip. They yielded no extractable text and are not classified and not counted in any licence class, per the brief. They are the OCR/manual-read list in §6.

`no-notice-found` — verified, because 334 is where a false zero would cost most¶

Six deterministic samples (sorted by key, every 55th), independently re-read at full document text, counting only copyright-notice markers:

Package	Format	Extracted chars
ACM	PDF	185,725
AHC	PDF	264,896
BSB	PDF	32,632
FWP	OOXML	211,907
MSL	PDF	111,948
TLI	PDF	8,333

0/6. Text extraction is demonstrably working (8k–265k chars, fonts present); the notices are genuinely absent. Every licen occurrence in these files is occupational-licensing prose — "Licence to operate a forklift truck", "Regulation and licensing requirements" — not a copyright notice.

A correction to my own first verification pass: it initially counted licen and commonwealth as licence markers and reported 5/6 as suspected false zeros. That was a defect in the verification test, not in the classification. Recorded because the intermediate figure would have been badly wrong.

`other-explicit` — what it actually contains¶

Trigger	Count
Creative Commons + NoDerivatives (i.e. an ND variant)	39
`© <year> PricewaterhouseCoopers … All rights reserved`	12
Commonwealth + all rights reserved	1

Verbatim, the PwC class: "© 2018 PricewaterhouseCoopers Data and Analytics Services Pty Limited. All rights reserved." — a third-party commercial rights holder embedded inside a companion volume, not the JSC and not the Commonwealth.

4 · RIGHTS HOLDER × CLASS¶

current_rights_holder is TGA's developer attribution, which is already the current custodian — all 11 are live bodies (10 JSCs + DEWR Interim). No defunct SSO/ISC names appear in this field.

Custodian	no-notice	CC-BY-NC-SA	other-explicit	copyright-only	CC-BY-4.0	CC-BY-SA	Total
Skills Insight	121	0	14	2	6	1	144
Manufacturing Industry Skills Alliance	50	7	25	3	5	0	90
Future Skills Organisation	11	45	7	4	0	0	67
Mining and Automotive Skills Alliance	27	21	4	4	0	0	56
Industry Skills Australia	31	0	0	15	0	0	46
Powering Skills Organisation	40	0	0	0	0	0	40
BuildSkills Australia	3	29	0	0	0	0	32
Public Skills Australia	28	0	0	1	0	0	29
Service and Creative Skills Australia	9	12	1	5	0	0	27
HumanAbility	11	0	0	2	4	0	17
DEWR (Interim)	3	8	1	1	0	0	13

publisher_as_printed is extracted per-file into the register from the notice text where a notice exists (e.g. Commonwealth of Australia, 2018 PricewaterhouseCoopers Data and Analytics Services Pty Limited). For the 334 no-notice files it is necessarily blank — there is no printed publisher to record. No defunct predecessor body was found printed in any extracted notice.

5 · THE DERIVED LETTER LIST — GROUPED BY CURRENT CUSTODIAN¶

Per Tim's ruling 2026-08-14: current custodians only. Classification of need: permission = CC-BY-NC-SA / CC-BY-SA / other-explicit (NC, SA or ND restrictions, or third-party rights) · attribution-only = CC-BY-4.0 · clarification = no-notice-found / copyright-no-licence.

Custodian	Total	Permission	Attribution-only	Needs clarification	Ask
Skills Insight	144	15	6	123	PERMISSION (+ large clarification)
Manufacturing Industry Skills Alliance	90	32	5	53	PERMISSION
Future Skills Organisation	67	52	0	15	PERMISSION
Mining and Automotive Skills Alliance	56	25	0	31	PERMISSION
BuildSkills Australia	32	29	0	3	PERMISSION
Service and Creative Skills Australia	27	13	0	14	PERMISSION
DEWR (Interim)	13	9	0	4	PERMISSION
Industry Skills Australia	46	0	0	46	CLARIFICATION
Powering Skills Organisation	40	0	0	40	CLARIFICATION
Public Skills Australia	29	0	0	29	CLARIFICATION
HumanAbility	17	0	4	13	CLARIFICATION

Seven custodians need a permission ask. Four need a clarification ask. None is attribution-only.

6 · NOT-MACHINE-READABLE — the OCR / manual-read list¶

7 members, all inside MEA/d066508e-cfa8-44f7-b94b-6a60f914a67b/fe298e82….zip ("MEA release 1.4 Companion Volume V1.0"). Unclassified and excluded from every licence count. Exact member names are in the register's zip_member column.

7 · WHAT SURPRISED ME¶

1 · The licence posture is worse than "inconsistent" — it is mostly absent. 334 of 561 files (59.5%) carry no copyright or licence notice of any kind in their TGA-hosted form. This is not an extraction failure; it was verified at full document text on a deterministic sample with zero notice markers found.

2 · The dossier's licence findings do not reproduce against the TGA-hosted corpus. The addendum records CHC R11.1 and both HLT topic volumes as carrying explicit CC BY-NC-SA 4.0. Measured here: CHC R11.1's TGA copy says "Licenced under Creative Commons 4.0" — no NC, no SA — and both HLT topic volumes carry no notice at all. The dossier read JSC-hosted copies; this census reads TGA-hosted copies. The same document can carry different licence text depending on where it was downloaded from, which is a materially harder problem than an inconsistent publisher.

3 · A third-party commercial rights holder is embedded in the corpus. Twelve files carry "© PricewaterhouseCoopers Data and Analytics Services Pty Limited. All rights reserved". A JSC permission letter does not reach PwC's rights.

4 · Thirty-nine files carry a NoDerivatives variant. ND is the one restriction that directly conflicts with generating derived material, which is the whole use case.

5 · Skills Insight is the largest holder and the least documented — 144 files, 121 with no notice.

6 · The single MEA zip is 403 of the corpus's 964 register rows — 42% of all rows sit inside one archive whose parent counts as one file in every prior census.

8 · WHAT WAS NOT DONE¶

No printed-publisher extraction for no-notice files — there is nothing printed to extract.
No successor mapping was required. All 11 developers are current bodies; no defunct SSO/ISC name appeared. Had one appeared, printed-name and successor would be separate register columns.
The 7 not-machine-readable members were not OCR'd. Out of scope; flagged as their own category.
No JSC-hosted copy was fetched for comparison against the TGA-hosted text, which is what §7.2 would need to close. Forbidden by the guards and correctly so.
CC-BY-4.0 classification does not distinguish licence versions or jurisdictions. A file saying "Creative Commons 4.0" without a variant is classed CC-BY-4.0; whether it is BY, or BY with unstated conditions, is not determinable from the notice text.

Least sure, and what would make this wrong. The no-notice-found class rests on absence of four marker strings in extracted text. If a notice is rendered as an image, or lives in PDF metadata rather than page content, or appears only in a cover page that pdftotext drops, the class is overstated — the six-sample verification tests extraction completeness but not those three failure modes, and I did not check XMP/DocInfo metadata at all, which is where a publisher tool would plausibly write a copyright field. Second, other-explicit bundles two genuinely different things — Creative Commons ND variants (39) and third-party all-rights-reserved (13) — and the letter implications are not the same; the register distinguishes them by search_term_hit but the class does not. Third, §7.2's claim that the same document differs by host rests on comparing this census against the dossier's description of JSC copies rather than against those bytes, which the guards prohibit fetching.