Regulatory Reporting Toolkit

Covered regimes

Six experts, each holding one regime and refusing to speak for any other.

• 🇨🇭
  FinfraG / FMIA Art. 39

  Swiss trading venue transaction reporting

  sme-fmia-art-39
• 🇨🇭
  FinfraG / FMIA Art. 104

  Swiss derivatives reporting to a trade repository

  sme-fmia-art-104
• 🇪🇺
  EMIR + REFIT

  EU derivatives reporting, clearing, risk mitigation

  sme-emir
• 🇪🇺
  MiFIR Art. 26, RTS 22/23/24

  EU transaction reporting

  sme-mifir-transaction-reporting
• 🇪🇺
  MiFID II

  Client classification and investor protection

  sme-mifid2-client-classification
• 🇨🇭
  FinSA / FIDLEG Art. 4

  Swiss client segmentation and conduct duties

  sme-finsa-client-segmentation

Out of scope: FinIA/FINIG, AMLA/GwG, SFTR, CSDR, Dodd-Frank, UK EMIR/MiFIR, APAC regimes, AnaCredit, MMSR, Basel/CRR and the Swiss Banking Act. The orchestrator names the gap rather than routing to the nearest-looking expert.

The pipeline

Experts run in parallel; every downstream stage runs in sequence, gated on its inputs actually existing. Stages that are not needed are not run — a single-regulation question runs one expert and the output stage.

1. 1
   reg-orchestrator

   Routes the question

   Asks for any decisive fact you left out, decides which experts the question actually needs, and names the gap when a regime is out of scope rather than routing to the nearest-looking expert.

2. 2
   6 × SME subagentsin parallel

   Analyse in parallel, in isolation

   Each expert analyses its own regime and nothing else, producing sourced findings with stable IDs.

   sme-fmia-art-39sme-fmia-art-104sme-emirsme-mifir-transaction-reportingsme-mifid2-client-classificationsme-finsa-client-segmentation
3. 3
   reg-comparison

   Compares without inventing equivalence

   Functional comparison only. Legal equivalence is a determination for competent authorities, and the stage refuses to declare it.

4. 4
   reg-convergence

   Finds where the regimes genuinely align

   Structured reasoning over comparison findings that are already sourced — separating real convergence from surface similarity.

5. 5
   reg-output-quality

   Strips anything unattributable

   Checks organisation, attribution and coverage against the bundled checklists before a single sentence reaches you.

6. 6
   reg-executive-output

   Writes the brief

   Selection and compression over findings that are already settled — with the conflicts and gaps still visible.

What asking it looks like

You ask in plain language. The orchestrator asks for any decisive fact you left out, tells you which experts it is about to run, and writes everything to a dated folder.

How citations are obtained

This is the single biggest determinant of how much the output is worth, so it is a setting you choose rather than a behaviour you have to infer. The mode in force is recorded in every run manifest and in the footer of every output.

An empty corpus is fine — the default mode degrades gracefully and tags what it could not verify.

What's inside

Everything is plain Markdown in a public repo. The skills reference one shared copy of the method rather than each carrying its own, so there is one version to read and one version to change.

What it does not do

Stated here rather than buried in the repo, because knowing the boundary is part of using it safely.

• Decision support, not legal advice. Every output carries that statement. The final call belongs to a qualified compliance, legal or regulatory reporting professional.
• Functional comparison only. Legal equivalence is a determination for competent authorities. The pipeline compares what regimes require and refuses to declare them equivalent.
• The output is only as good as the grounding. In the default hybrid mode, check the [UNVERIFIED] tags before relying on anything. They are there so you can.
• Regulation moves. Every output states its as-of date and flags the pending changes it knows about. Check the horizon items rather than trusting them.