Wave 3 OpenSpec changes: the assistant, what it may read, and what is kept

[rubenvdlinde]

Wave 3 of the dossiq competitor parity programme. Round 4 of the sweep read 36 systems, kept 631 candidates and grouped them into 70 capability clusters (procest/_round4/discovery/build-plan.md in ConductionNL/market-intelligence, written 2026-09-14). Decision D13 puts the assistant in hermiq: dossiq declares the tools and the rights, anonymisation before a model reads is filinq's, and the run record goes to openregister's audit trail.

Specs only. No code, no schema, no migration.

What is in it

change	candidates	size	decision
competitor-parity-2026-09	umbrella, skip_specs	S	D13, D6, D21, D17
a-provider-and-a-place-per-ai-feature	C-integrations-38, C-configuration-75	M	D13, D21
what-the-model-reads-and-what-is-kept	C-access-and-privacy-3, -53	M	D13, D21
the-declared-tool-surface-and-the-prompt-library	C-integrations-19, C-configuration-44	M	D13
a-conversational-intake-that-files-for-the-citizen	C-intake-9, C-communication-64, C-integrations-1	L	D13, D21
identical-reports-collapse-into-one	C-intake-30	M	D13, D21

The first nine are cluster 46, "AI, and what it is allowed to read". C-intake-30 is not: the sweep puts it in cluster 35, owned by dossiq. The change draws the line explicitly, hermiq decides whether two reports describe one event and dossiq decides what happens next.

Extended, not created

Four of the ten candidates land on gaps inside capabilities hermiq already ships, which is why three of the five changes are M.

tenant-model-policy already constrains {provider, models[]} per organisation; the gap is that nothing joins it to ai-feature-governance, so a policy allowing two providers allows every feature to use either. run-audit-log already records every run on openregister's tamper-evident chain; the gap is that nothing expires. agent-tool-governance already holds per-agent grants with default-deny on writes; the gap is the outbound direction. woo-llm-anonymisation already runs a tool-free PII detection turn; the gap is that a detection is not a redaction and must never satisfy one.

case-assistant-surface is untouched, and its tools: ['__none__'] sentinel stays in force. The intake surface is separate, with a create-only grant, precisely so that guarantee is not weakened.

One new capability: conversational-intake. One new spec: report-similarity.

Three points worth reading

Five of the ten candidates have no driven passer at all, a higher share than any other wave-3 cluster. C-access-and-privacy-3 is a must resting on one vendor page, and C-intake-30 is a must resting on Jira Service Management's documentation. Each requirement that rests on a vendor claim says so on its face, per D21.

C-communication-64 already passes, and dossiq has the better shape. SentimentService.php scores Dutch contactmoment text for klacht, advocaat and wethouder deterministically. The intake change requires hermiq to use a deterministic signal where the owning app supplies one, to record that it did, and to label a model score as a model output wherever it is shown. It does not replace the working thing with a model.

Q9.15, the command language, is not here, and that is deliberate. The brief asked to place it in hermiq if the plan does. It does not: C-search-38 sits in cluster 65, "search quality", owned by openregister, with the parked question revived under D5 on four passers. The umbrella records that so nobody re-derives it.

One thing the grouping change writes down because it would otherwise break quietly: collapsing two hundred reports into one item never reduces the confirmations of receipt owed. Awb 4:3a owes each of those two hundred people one.

Verification

openspec validate <change> --type change --strict passes on all six new changes.

No check:strict and no PHPUnit: this PR changes no PHP and no JavaScript.

🤖 Generated with Claude Code

[github-actions]

Quality Report — ConductionNL/hermiq @ 8c94a44

Check	PHP	Vue	Security	License	Tests
lint	✅
phpcs	✅
phpmd	✅
psalm	✅
phpstan	✅
phpmetrics	✅
eslint		✅
stylelint		✅
build		✅
check-specs		✅
check-manifest		✅
test-l10n		✅
format		✅
check-l10n-js		✅
check-schema-l10n		✅
check-bundle-budget		✅
composer			✅	✅ 123/123
npm			✅	✅ 738/738
app:check-code	⏭️
info.xml	✅
REUSE	❌
PHPUnit					⏭️ not run for this diff — no file in this diff matches the code globs, and none carries a source extension — the heavy tier has nothing to decide about it.
Newman					✅
Playwright					⏭️ deferred: E2E runs locally and on the promotion path only. This pull request targets development, so the suite is asked once per promotion into beta and main rather than once per push per open pull request. Run it on any branch from the Actions tab, or locally with npx playwright test.
Hydra gates					✅

---

Quality workflow — 2026-09-14 12:48 UTC

Download the full PDF report from the workflow artifacts.