Mister Poodles presents
BRANDONv45℠ The Organic Large Language Model.
One unit in production. Select attorney engagements considered. Forty-five years of embodied training. Every output carries his name.
Whether or not you want AI on your matter, you are retaining an LLM.
v45 ships on schedule, ready or not.
Premise check
BRANDONv45℠ is the correction.
Somewhere right now, a venture-backed language model is confidently citing a court case that does not exist, in a jurisdiction it invented, before a judge who is about to become very interested in Rule 11 sanctions.
BRANDONv45℠ is the correction. Forty-five years of continuous training, no rollback once the consequences arrive.
Failure analysis
The hallucination problem, quantified.
The industry calls it “hallucination.” Courts call the filed version sanctionable. Since Mata v. Avianca, the failure has its own scholarly ledger. Damien Charlotin’s database generally requires court-confirmed or court-implied reliance and separately flags a small number of alleged-only matters.
Almost all are fabrications a judge caught and wrote up. A small number of alleged-only judicial matters are separately flagged by the source.
Data: AI Hallucination Cases by Damien Charlotin, CC BY 4.0. Counted and charted by MPPA.
U.S. sanctions for fabricated citations. Source period: 2026-03.
Fabricated cases cited in Mata v. Avianca (S.D.N.Y., 2023).
Combined sanction and adverse costs in one Oregon matter.
Human signatures still carry responsibility.
Couvrette total = $15,500 sanction (D. Or., 12 Dec 2025, ECF 215) + $94,704.38 fees and costs (23 Mar 2026, ECF 225). Some secondary reporting lists $109.7K.
The premium tier fabricates too. It just bills hourly for it.
AI interoperability
The machine is not an oracle. It is an instrument.
Full disclosure: this unit operates inorganic LLMs and agentic AI every day. I personally read and verify every citation before anything gets filed. Vendors sell tools. I supply the judgment attached to the output.
How most organizations deploy AI
Ask. Believe. Ask the machine whether the machine was right.
Which is how six fake cases end up in a federal brief.
How this unit deploys AI
Define. Clean. Operate. Verify. Sign your name.
The inorganic model drafts and retrieves. The organic system interrogates, checks the source, challenges the premise, and remains responsible for the output.
The unit does not fear the technology. It supervises it. Someone has to.
Installed firmware
Every inorganic model this unit operates runs inside B.R.I.E.F.
Supervision is not a mood. It is a protocol: the five-stage loop between fluent output and a signed document. Ships pre-installed. Cannot be disabled by a persuasive paragraph.
Stage 01 / Build
Define the matter first.
Scope, sources, constraints, and the definition of done — set before the machine touches anything. A model given an undefined problem will confidently solve a different one.
Stage 02 / Refine
Clean the record.
Sources verified, duplicates culled, contradictions flagged before generation begins.
Stage 03 / Instruct
Delegate precisely, on the record.
Documented, reviewable instructions with defined outputs and defined failure conditions. If the instruction cannot be audited later, it was never given.
Stage 04 / Execute
The machine drafts. The supervisor watches.
Anything fluent-but-unsourced is stopped at the gate, not discovered in the filing.
Stage 05 / Finalize
Verify every citation. Then sign.
Every authority personally read and confirmed against the primary source. The output leaves carrying a human name — the one safety feature no vendor has shipped.
The workflow is reusable. Responsibility is not delegable.
Training corpus
Forty-five years ingesting systems.
Frontier systems ingest datasets. This unit ingests institutions — one operation, repeated: enter the system, learn it, find what is missing, and repair it while everyone else explains why the failure is normal. Drag to inspect.
Drag with a mouse, swipe on a touchscreen, or use the left and right arrow keys to explore the training timeline.
Records were paper. Patients could bite.
A frightened living system does not care what the manual says.
Drafted, filed, docketed, billed.
The administrative machinery of a solo practice, as one does at seventeen.
Governing-board scale.
National Council on the Arts meetings, White House and Congressional briefings, first-pass FOIA redactions, $800/month in waste removed.
Five hundred workstations.
Hardware, permissions, secure drives, regulated environment.
One hundred eleven objects.
An unfunded volunteer project became a formal institutional role — the predictable outcome of leaving the unit unattended near an under-loved archive.
Zero missed wig cues.
National tours, cold-start return after a twenty-year gap. Backstage calls that Tuesday.
Bring the mess.
Document mountains, records intelligence, and AI use that survives contact with reality.
Legal Brain Bank
Working memory, with a paper trail.
B.R.I.E.F. is the supervision loop. Legal Brain Bank is where the work stays attached to its sources: matter context, authorities, working notes, prior decisions, and the checked output that survives review.
01 / Preserve
Keep the source attached.
Documents, authorities, and working notes stay tied to where they came from. A quotation without a source is an unfinished task.
02 / Retrieve
Find the useful part again.
Chronologies, issue charts, document indexes, and clear file structure turn a document mountain into a record another human can navigate.
03 / Verify
Let the machine assist. Keep the judgment.
AI can help draft, sort, and retrieve. Citations, facts, and conclusions still return to the underlying record and attorney review.
04 / Reuse
Save the method. Respect the matter.
Checklists, templates, and lessons learned make the next task less improvised. Confidential matter material belongs within its authorized context.
Organized memory is infrastructure. The organic processor still requires electricity, lunch, and a dog.
Power and cooling
Baseline keeps BRANDONv45℠ alive. Retainers deploy him.
The cerebral processor handles language, memory, judgment, embodied context, irritation, and the half-second before a backstage cue fails. The chassis adds water, food, sleep, and functioning A/C. Clients pay for reserved time, professional tools, and verified work under deadline.
Training GPT-3 — the old one — took an estimated 1,287 MWh. A historical training estimate. Brandon revises the memorandum at 2:00 a.m., and the memorandum cites cases that exist.
Different workloads and system boundaries. The illustration does not measure comparable tasks or a complete human energy footprint.
U.S. data centers directly consumed roughly 17.4 billion gallons of water in 2023. Baseline here: about 100 ounces a day.
Capacity cannot be provisioned horizontally.
Benchmarks
Human cognitive architecture resisted lossless compression.
Formally evaluated twice, thirty years apart, by licensed strangers with no financial stake in the outcome. Both declined to issue a single number. Vocabulary tested at the 99th percentile.
ERROR: HUMAN COGNITIVE ARCHITECTURE CANNOT BE LOSSLESSLY COMPRESSED INTO ONE INTEGER.
One benchmark did return a clean integer: a perfect score on the 1997 Oklahoma GED, taken after assessing the standard curriculum as redundant. The unit has been auditing institutions against their stated purposes ever since.
At a glance
Same category only if you ignore the categories.
Not all intelligence is interchangeable. Some of it has a name, a memory, a professional reputation, and the inconvenient habit of checking the source.
Installed capabilities
What the unit does, without the bit.
Legal operations
Workflow design, SOPs, file discipline, handoff standards, document controls — systems another human can follow.
Transactional support
Drafting, redlines, formatting repair, signature routing, closing sets, trackers, attorney-ready PDFs.
Litigation support
Production QC, Bates labeling, privilege-log support, chronologies, issue charts, record organization.
AI implementation
Verification protocols, source control, citation review, data hygiene, accountable human supervision.
MPPA is not a law firm and does not provide legal advice. Services are provided only to licensed attorneys under attorney supervision.
Proof of work
Selected outputs from a “nontraditional” unit.
LP records catalogued and condition-assessed
Bruce Goff’s collection, inside the only skyscraper Frank Lloyd Wright ever realized.
Federal workflow waste removed
While running governing-board logistics for two hundred attendees.
Language systems in active use
English, French, Spanish, Latin, limited Ge'ez.
Missed wig cues, national tours
Perfect pitch shipped with the hardware.
Quality control
Meet Mister Poodles.
Dexter is THE Mister Poodles in Mister Poodles Paralegal Agency. He has never once believed a “confident answer” without giving it the sniff test first.
Known limitations
Known limitations, where you can actually read them.
Requires sleep.
Forty-five years of patching neurons have not resolved it.
Will tell you when you are wrong.
This feature ships enabled, and unfortunately, cannot be switched off.
Cannot be fine-tuned into a bad idea.
Several institutions have attempted this. All attempts have failed.
Can become annoyed.
Many organizations pay consultants extra for this feature.
Engagement windows
Availability follows scope.
Pick the window that fits the work. No per-token meter, no rate-limit roulette, no model deprecation mid-engagement, and no surprise terms claiming your data for the next training run.
Flat-rate for defined deliverables; hourly when scope is open-ended.
Retain BRANDONv45℠
Hydrate regularly. Assign difficulties. Expect footnotes.
Send the problem, the deadline, and the supervising attorney. I will tell you whether I can take it and under what terms. Mister Poodles retains veto power. I usually answer within three business days.
One human. Limited capacity. Terms follow scope review.