Themis / Jubal turns agent breadth into counsel-grade preparation — then stops at the exact line where licensed judgment, signature, and accountability begin.
The protocol does not cosplay as a law firm. It frames facts, binds claims to current authority, forces both readings, attacks its own work, and gives counsel the few questions that can actually change the decision.
Agents do the exhaustive preparation. A licensed lawyer or accountant decides the legally consequential point. Every difference between the two becomes the next improvement to the system.
The names are not theatre. They separate constitutional integrity from case execution, so a difficult answer cannot quietly rewrite the standard that produced it.
Themis decides what counts as a source, what must be attacked, which roles must be separate, what agents may never claim, and how the system may change.
A case result cannot weaken the citation standard. It can only propose a dated, explicit amendment for Lee to accept.
Jubal runs intake, opens the source shelves, frames the matrix, writes both readings, derives controls, survives attack, and records the licensed delta.
The Mysterium lawful-data case opened a 22-pin source shelf before the minimum-facts gate allowed a merits conclusion.
It owns the sources, method, attacks, learning loop, deterministic tools, and the boundary on what may happen next.
Each lower layer makes the layer above it harder to counterfeit. An answer without the stack is a model opinion, not a Jubal result.
Versioned primary law, case law, regulator material, counsel deltas, and lower-tier commentary kept visibly distinct.
Concrete example: Lithuanian provisions stay verbatim in Lithuanian with an English gloss, source URL, hierarchy, version date, and later-treatment check.
Facts, actors, jurisdictions, purposes, burdens, evidence, procedure, remedies, enforcement probability, impact, and reversibility.
Concrete example: “Can we use this data?” becomes a matrix of which data, which actor, which purpose, which jurisdiction, under which control, with which burden.
Every counted review round attacks citations, unopened law, hostile interpretation, practicability, jurisdiction conflict, and currency.
Concrete example: A technically sound data control still fails if MN cannot implement it with its actual systems.
Counsel deltas, outcome-blind backtests, and model-release comparisons turn misses into new sources, lenses, or protocol rules.
Concrete example: If counsel opens a body of law the agents missed, that miss becomes an unopened-shelf fixture for the next case.
Deterministic checks are built red-first after a real dossier proves the need: quote diffing, currency checks, and gate validation.
Concrete example: Case 1’s citation checker caught seven quotations that were not verbatim; the propositions stayed, the quotes were repaired from the pins.
Role separation, data handling, provider boundaries, privilege limits, provenance, mandatory output states, and human-control gates.
Concrete example: A passing internal gate does not authorize signature, filing, entity activation, spend, or disclosure.
No merits conclusion before the minimum facts exist. No confident proposition without current authority. No real-world legal change without licensed adjudication.
Reuse the canonical case record when the facts, decision, and legal domains already exist. Parallel dossiers create competing truths.
Researcher, citation verifier, opposing reviewer, and calibration custodian are named. The researcher is separate from the other three functions.
Record the objective in the principal’s words, tag every material fact by source and date, request the missing documents, and map procedure while facts are still arriving.
If the decision-changing facts are absent, the system may prepare the questions. It may not pretend the merits are ready.
Separate law, counsel interpretation, organisation policy, technical control, and commercial preference. Label each cell fact, hypothesis, assumption, or hunch.
Pin the current text, quote it verbatim, classify its authority for this proposition, record what was searched, and keep negative results.
Steelman the restrictive and permissive cases with equal force. One-sided reasoning is advocacy, not preparation.
Turn “allowed or forbidden?” into “under which control, at what cost?” A control that the organisation cannot run is not a control.
Rank the questions, give the proposed answer and confidence, and name the evidence that would change it. The memo becomes the appendix.
Run the six-lens loop to its stop rule, record the licensed delta by domain, and convert confirmed misses into system improvements.
The reviewer changes lens, not merely personality. Each lens names a different way legal work becomes dangerously convincing.
Re-fetch the cited provision. Diff it against the quotation. Check that the source actually supports the proposition.
It kills: authoritative-looking fabrication.
Name every body of law the dossier ignored and argue how each might bite the same facts.
It kills: elegant analysis of the wrong legal universe.
Write the most restrictive defensible interpretation as the regulator or opposing counsel would.
It kills: optimism disguised as balance.
Test whether this organisation, with these systems and people, can execute every proposed control.
It kills: compliance theatre.
Find where EU, national, contractual, and third-country rules disagree. Establish which rule wins.
It kills: single-jurisdiction tunnel vision.
Check whether the act, judgment, guidance, or interpretation changed after the pinned version.
It kills: perfectly cited obsolete law.
Stop only after two consecutive rounds produce no confirmed new finding. A lower-evidence exception is not a clean round.
Builds the shelf, facts, matrix, readings, controls, and counsel questions.
Re-fetches sources and checks exact quotes, currency, and proposition fit.
Attacks the dossier through every legal-specific lens and records findings.
Owns the denominators, outcome records, amendments, and cross-case learning.
“A verbatim quote proves the words, not the law.”
The protocol also records hierarchy, binding force, subsequent treatment, applicability, search scope, negative results, and the proposition the source is actually being asked to carry.
A longer memo is not learning. A changed source shelf, attack lens, deterministic guard, or constitutional rule is learning.
Records what counsel added, rejected, accepted, and how confident they were. Counterparty position stays distinct from independent adjudication.
Run the protocol on admissible post-cutoff cases without the outcome. Leakage controls stop recognition from posing as judgment.
A new model inherits the role only after it clears the defined non-inferiority and zero-regression gates. This measurement appendix remains frozen pending its own review.
The protocol must become denser, not merely longer. If a model now performs a check natively, the check is tested for retirement.
Talos and Ieva independently converged on every explicit control proposition in Ieva’s reply. That is strong evidence that the control shape is useful. It is not evidence that either answer is legally correct, or that MN’s missing facts satisfy the controls.
All explicit propositions appear in the pre-reveal Talos dossier. This denominator measures issue spotting and control shape—not correctness.
Purpose scope. Ieva answered a narrow organisational-summary question; Talos modelled four broader analytical purpose groups.
No proposition directly conflicts. The broader Talos perimeter remains conditional because Ieva did not adjudicate those wider purposes.
Talos’s dossier also covers revenue perimeter, customer or cohort quality, cost / margin / headcount, and acquisition / retention / payment-recovery work. The same controls may apply, but the factual need, legal basis, fields, notices and provider path can change by purpose.
Operational consequence: start narrow. Every broader purpose gets its own purpose card before fields or access expand.
SHA-256 ecf68aed1d10c8ffb39d0c1463adc27ba363a2c1322e445a1ef519105cfb8a1a
Talos completed the abstract-data dossier without reading Ieva’s reply, passed the 13/13 citation checker, and froze the exact bytes.
Ieva’s framing was opened only after the freeze. The raw WhatsApp text remains in the private source record and is not copied into this deployment.
The two answers were compared, not merged. Shared controls, Talos-only mechanics, Ieva-only authority claims and unresolved facts remain distinct.
Rounds 13–52 kept exposing defects in the competence-claim and model-swap statistics. The accepted operating core was separated from that machinery so case work could proceed without publishing an unearned competence claim.
Accepted core: source, method, adversarial obligations, governance, roles, case roadmap. Frozen appendix: fixture statistics, claim gates, model-swap adoption statistics. Open seam: N61, critical-miss zero-regression rule.
Every dossier must lead with one of four exact states. No amount of polish upgrades the state.
JUBAL PREPARATION COMPLETEInternal protocol work passed; licensed adjudication is still pending.
JUBAL ADJUDICATEDNamed legal or tax domains were adjudicated by named licensed evidence on a stated date.
JUBAL LOWER-EVIDENCE EXCEPTIONMissing components, the authority ceiling, consequence, and next owner are visible on the artifact.
JUBAL BLOCKEDA failed gate and the exact next owner are named. Work does not masquerade as a conclusion.
Accepted 24 Aug 2026. It governs current case work.
The blind comparison is complete: 13/13 control themes overlap, with no direct contradiction. The narrow purpose is framed; broader purposes remain conditional.
Conditional preparation exists. Two clean independent rounds and licensed Lithuanian / activity-matched Vietnamese adjudication remain pending.
Not because the system is infallible. Because it makes error visible, authority explicit, counsel efficient, and every real miss useful to the next case.
Open or continue the dossier. Name the facts. Pin the authority. Attack all six lenses. Stop at the licensed line.
This presentation explains the protocol. It is not a legal opinion and does not change any case’s authority state.
Themis binds. Jubal learns.