Evidence brief · the published exemplar · our own venture
AI Parliamentary Questions Analysis
One researched briefing. One publication. One audit. The Evidence brief product, run end-to-end on the hardest corpus we had: every parliamentary question of the 34th Dáil.
Product
Evidence brief — one researched briefing, published as a hash-pinned PDF, with an as-published audit, a corrections log and right of reply. This page is that product's published exemplar.
The briefing
"Artificial Intelligence and Digital Policy in Parliamentary Questions — 34th Dáil, First 18 Months": what share of Dáil attention goes to AI and digital policy, who drives it, and where it lands — from the full record to the 3 August 2026 cutoff, not a sample.
Duration
Fourteen days from research corpus to publication: corpus cut 3 August 2026, briefing frozen 15 August, published 17 August.
Reported outcome
We used this, on our own behalf, to reach out to Matthew Fenlon, who works with Albert Dolan on statespend.ie — a project which preceded our own PO Monitor and whose findings we were able to replicate independently. Our own account, not an independently verifiable record.
One question in every twenty asked in the Dáil is about AI or digital policy. The briefing that established that number was published with its limits printed beside its findings — which is the product.
What an Evidence brief is
The Evidence brief is the fixed-scope entry point to Soliber's evidence-publishing work: one argument researched to a verifiable standard, one briefing shipped as a hash-pinned PDF, one as-published audit confirming the live document matches the ratified text, with a corrections log and right of reply standing afterwards. AI drafts; humans edit every published word and check every number; humans decide what ships.
This page shows that anatomy executed in full, in public, on our own question — so a client can inspect exactly what the product produces before commissioning one.
The story
The question. Irish public debate asserts constantly that AI is consuming political attention. Nobody could say how much, because no classified measurement of the parliamentary record existed — only raw counts. We set out to answer it from the full population: every one of the 118,399 written and oral questions tabled in the 34th Dáil between its first sitting on 22 January 2025 and the 28 July 2026 record, collected to a 3 August cutoff.
The research base. The corpus was assembled from the Houses of the Oireachtas open-data API and the official daily XML, normalised, validated to zero duplicate keys and zero orphan links, and hash-manifested — so every downstream figure resolves to a dataset with a printed SHA-256. Each question was then classified under a fixed 15-category rubric with three relevance levels.
The discipline. A second, stronger model blind-reclassified a 14,729-question audit stratum. The two passes disagreed on 422 questions; a human adjudicated every one in a structured workbook, and the 386 applied corrections flow through an append-only log — the original model output is never edited. The briefing then went through two rounds of external editorial review: factual corrections, moderated interpretive language, and a sharper statement of what the audit does and does not establish.
The findings. 6,026 questions (5.1%) touch AI or digital policy, four in five directly. Attention rose by roughly two-thirds across the first 18 months. One deputy filed a fifth of all relevant questions — so every statistic in the briefing is stated with and without him. The questions concentrate where the casework is — Culture, Enterprise, Tourism, Health — not in the coordinating digital portfolios. Sustained scrutiny runs through about eight deputies of 158, across four parties.
The honesty line. The briefing states, in its own executive summary, that the audit measured model disagreement in a targeted stratum — not corpus-wide accuracy — and that the headline count is best read as a floor until a representative human-coded validation exists. That caveat was published as prominently as the findings. It is the sentence that makes the rest of the document quotable.
118,399
questions in the research base — the full term, not a sample
6,026
AI/digital-relevant (5.1%); 79% directly relevant
×1.62
half-yearly attention rise, 3.9% → 6.3%; May 2026 peak 8.2%
422
model disagreements, each human-adjudicated before publication
10 pages
published edition; monthly base numbers in the document of record's appendix
2
rounds of external editorial review before freeze
Progression
| Date | Step | Evidence-brief anatomy |
|---|---|---|
| 3 Aug 2026 | Research corpus cut: full 34th-Dáil question record collected from the official API, validated, hash-manifested. | Research sweep |
| 3–15 Aug | Full-corpus classification under the fixed rubric; blind second-model audit; 422 disagreements human-adjudicated; corrections applied through an append-only log. | Verification passes |
| 12–15 Aug | Two rounds of external editorial review; reproducibility appendix written, including the one known gap, stated in print. | Human edit — every word, every number |
| 15 Aug | Briefing frozen as the revised edition (v3). Figures pinned to named datasets with printed SHA-256 hashes. | Hash-pinned PDF |
| 17 Aug | Published as a 10-page edition built for LinkedIn's in-feed document reader — large type for mobile, one message per page, figures identical to the frozen briefing. The artefact is fitted to the channel; the numbers are not re-derived for it. | One publication |
| 8 Oct 2026 | As-published audit, run 52 days after publication: the live post rendered in a browser and checked against the frozen edition. All ten pages match in text and in render; the post copy matches the ratified text in every word and figure. One typographic delta logged: six em dashes display as hyphens. | As-published audit |
| Since | Corrections log and right of reply standing. | Corrections log · right of reply |
The published document
The briefing exists in two editions with identical figures: the document of record — the full briefing, frozen 15 August with printed dataset hashes — and the published edition, rebuilt for the platform it shipped on: large type for mobile document readers, one message per page, the reading rule and the audit caveat carried over intact. Platform-optimised presentation, same numbers, same limits.
AI in the Dáil: What 118,399 Parliamentary Questions Reveal
Download the published edition (PDF, 564 KB) →
The LinkedIn publication of record ↗
SHA-256 of this file, printed so you can check it: 797570c2b9502c8d2d1b61a64f6bdea47f15b88c6e701332f674a9802a02a975
Methods and data access: the classification dataset is not public; reproduction runs under a data-sharing agreement — [email protected]
What it demonstrates, line by line of the offer
"One researched briefing" — a single 10-page document answering a single question from a verified research base, every figure resolving to a hashed dataset or an official Oireachtas record.
"Hash-pinned PDF" — the corpus index, the classification table of record, the corrections log and the audit stratum all carry printed SHA-256 values in the document of record's provenance section; a 42-check verification suite recomputes every cited figure.
"As-published audit" — the live post was rendered in a browser on 8 October 2026 and checked against the frozen edition: the document's text matches on all ten pages, every chart matches, and the post copy matches the ratified text in every word and figure. One typographic delta was logged, and no correction was required. The audit ran 52 days after publication rather than within days of it; this page states that lag rather than hiding it.
"Corrections log and right of reply" — append-only corrections, raw and corrected states both permanently inspectable, contact line printed in the document.
"Humans decide what ships" — two rounds of external editorial review before freeze, named in the document of record itself: factual corrections, moderated interpretive language and a sharper statement of the audit's limits.
Dated evidence
| Claim | Source | Date |
|---|---|---|
| Research corpus: 118,399 questions, validated, zero duplicate keys / orphan links | Corpus manifest + validation report; questions parquet SHA-256 bf124224071a1bf13ae76058253ce66f8ba781a2df57cdf04aefff190e9df52d | 3 Aug 2026 |
| Classification table of record (audit-corrected), 118,399 rows | Briefing provenance entry [1]; SHA-256 53e384a3607a91e1839bea3da525c53204213b10bbf23764e9686244d96156ae | 15 Aug 2026 |
| 422 disagreements adjudicated; 382 corrections accepted | Briefing §2 and corrections-log parquet, provenance entry [3]; SHA-256 eee3f1a0f8d42ee914fbc52c4fb4e5ad5265f8d82de0ce3081fab14780972679 | 15 Aug 2026 |
| Publication | LinkedIn activity 7495046835842052096 ↗ | 17 Aug 2026 |
| As-published audit completed: 10 of 10 pages match the frozen edition; one typographic delta logged; no correction required | Audit record: browser-rendered check of the live LinkedIn post against the frozen edition and the ratified post text | 8 Oct 2026 |
| Outreach to Matthew Fenlon (statespend.ie), on our own behalf (reported outcome above) | Our own account — not an independently verifiable record | — |
Every claim on this page resolves to an official Oireachtas record, a hash-printed dataset named in the briefing's provenance section, or a dated publication. Corpus absence is never presented as parliamentary fact; the briefing prints its coverage window on page one.
This is the product. One argument, researched to this standard, published under your name with the same apparatus standing behind it — Evidence brief, from €3,500. A different question gets a different briefing on the same discipline.