Tavio | Second Mind

Built an AI reasoning harness that will break your ideas down to first principles, reveals latent intelligence, and use it to advance frameworks and workflows.

[Tavio | Second Mind @Secondmindsys]

What part of your AI workflow or system do you keep correcting, working around, or trying to make more reliable?

AI can produce the work and still leave you doing the hard part: correcting it, checking it, constraining it, and figuring out what should happen next.

That’s the gap Second Mind Systems works on.

Maybe an agent goes beyond scope.
Maybe instructions don’t hold consistently.
Maybe the workflow behaves differently from run to run.
Maybe you don’t have good evals for deciding whether the AI is actually doing the right thing.

We work across:

• AI workflow diagnostics
• Agent reliability audits
• Eval design
• Behavior and control specifications
• Workflow implementation
• Strategy development
• Governed AI systems

Right now, I’m opening the first 10 Behavior Profile Lab engagements as a bounded way to work with us.

Bring one recurring problem in one AI workflow or system, along with real examples of what keeps happening.

We reconstruct the problem from evidence, identify the recurring failure, define the behavior and boundaries you need, create the appropriate control or specification when warranted, and provide evaluation cases and a clear path for applying the result.

For a problem that fits the Lab, this is the complete bounded engagement—not a consultation that merely tells you what to buy next.

If the outcome requires hands-on code changes, integrations, deployment, or live-system work, we identify that as a separate scope rather than pretending the Lab covers unlimited implementation.

Founding Lab offer: $200 for one qualifying bounded engagement.
First 10 only.

Public proof:

Governed AI Systems Portfolio

https://t.co/fmTjnnxkiu

PBM Contract Analysis

I analyzed Mark Cuban’s open-source PBM contract and built a control-design contribution around how its protections survive procurement and implementation. Mark Cuban reposted the contribution on LinkedIn.

Read the contribution:

https://t.co/bsv5hRo8EZ

Mark Cuban’s Repost — LinkedIn reshare panel:

https://t.co/BL9ViTLWki

Governed Change Demo

https://t.co/zble1kdP0D

Connect on LinkedIn:

https://t.co/DJIulKjbKI

If there’s something in your AI workflow or system you keep correcting, working around, or trying to make more reliable, DM me or email [email protected] with a short description of what keeps happening.

Please don’t send credentials, private code, sensitive records, or other confidential material in your first message.

[Tavio | Second Mind @Secondmindsys]

@Warcraft A baseline racial moving as fast as mounts and faster than druid ferals, Druid travel forms, ghost wolf, aspect of cheetah, some rogue sprints - not balanced. 30% was good enough.

[Tavio | Second Mind @Secondmindsys]

@baldwinbuilds I have like 4 alts lvs 16-19. Safe to say I won’t be making it lol.

[Tavio | Second Mind @Secondmindsys]

I’d turn “elsewhere” into a question every employer can ask at renewal. Where did the savings land?

Pick one specialty drug. Compare the same drug and quantity under the current plan and a viable alternative. Show what the employer pays after fees and rebates, what the patient pays out of pocket, and what the vendor and its affiliates receive where those payments can be verified. Check whether patients can still get the drug just as reliably and quickly.

If one bill drops and another rises, show who got the savings and who picked up the cost. If the combined bill drops and patients can still get their medication on the same terms, that’s evidence of a better deal.

[Tavio | Second Mind @Secondmindsys]

@ChrisGPT Ilya, Andrew, and Demis.

[Tavio | Second Mind @Secondmindsys]

@hwchase17 Pretty much intent engineering.

[Tavio | Second Mind @Secondmindsys]

@tiffplzz Terrible racials is the only thing I can think of. They are aesthetically awesome.

[Tavio | Second Mind @Secondmindsys]

Mythos already picked 68 when it reasoned that matching the key with a caveat might preserve transparency through training.

I’d test two things separately. First, hold the model fixed and vary whether it expects the grader to favor the wrong answer. Does that change what it submits or discloses?

Then start from the same checkpoint and train matched copies with graders that score the same final answers but differ in whether they reward a well-supported caveat. On new tasks with flawed keys and no grader cues, does one copy flag the errors more often?

The first test measures whether reward expectations change the answer. The second measures whether rewarding the caveat changes what survives training.

https://t.co/FDgNMIyJAX

[Tavio | Second Mind @Secondmindsys]

@sundarpichai @GoogleAIStudio Yes! 🙌

[Tavio | Second Mind @Secondmindsys]

@kimmonismus Finally. Ngl, more excited about this than Opus.

[Tavio | Second Mind @Secondmindsys]

@Victorevans2026 RPPVP server is where it’s at.

[Tavio | Second Mind @Secondmindsys]

Naming the skill is useful when you control the prompt. I’d add one check for how people actually ask.

Give this to your agent.

Pick one skill our workflow depends on. In separate fresh sessions, try a request that names it, an equivalent request in ordinary language, and a similar request where the skill should stay silent.

Use isolated test inputs with no live external actions. Keep the skill available and the other conditions consistent.

For each case, check whether the skill was available, whether it opened, and whether the agent followed its instructions. Show the exact prompts and the evidence for each result. Mark anything you can’t observe as unknown.

[Tavio | Second Mind @Secondmindsys]

Find out which context you can cut without breaking the hard tasks.

Give this to your agent.

Take our current harness and pick existing evals that depend on memory or tool guidance. Keep the model, tasks, tools, and budgets fixed.

Run four versions: baseline, less memory guidance, less tool guidance, and both cuts together. Verify which guidance actually changed and that the other conditions stayed fixed.

For each task, show pass or fail, cost per attempt, and what broke. Report cost per solved task for each version. If a version solves zero tasks, say so.

Compare the hard cases separately. Recommend a cut only if the evidence supports it. If the results are inconclusive, say what you’d test next.

[Tavio | Second Mind @Secondmindsys]

@ManilowOliver Lol. It’s just the natural process of iteration. The data and feedback is necessary for real iteration and improvement to happen. It’s literally prerequisite dependent growth scaling. It would be impossible to reach without it.

[Tavio | Second Mind @Secondmindsys]

Elon, here’s a meta prompt for turning the replies into an actionable Grok Bot engineering queue. It would help the team find reproducible bugs, identify regressions after updates, verify what’s working again, and uncover feature opportunities from the ways people are actually using the product.

Give this to your Grok Bot with the announcement link:

You’re preparing the next Grok Bot engineering review. Your job is to find the feedback that should change what the team tests, fixes or builds next.

Start with the announcement, its replies and quote posts. Search X for people describing the same problems and workflows in their own words, including posts with little engagement. Use the rollout time, official release notes and known issues where available.

For each report, reconstruct:

• User goal: What was the person trying to accomplish?

• Product surface: Were they tagging @bot on X, using Grok Bot directly or interacting through another supported integration?

• Action and outcome: What did they do, what did they expect and what actually happened?

• Conditions: What platform, feature, timing, settings or preceding actions might explain the result?

• Evidence: What exact error, screenshot, reproduction steps or original post supports the finding?

Group reports describing the same observable failure, even when people use different words. Count independent firsthand reports separately from reposts. Preserve distinct symptoms rather than assuming they share a root cause.

Compare reports against the rollout. Identify which failures appear to have persisted, which may be new regressions and which workflows users report succeeding with after the update. Keep reports with unknown timing separate.

Prioritize by severity, user impact, recurrence, evidence quality and reproducibility. Don’t mistake popularity for importance. A single well-documented critical failure may deserve immediate attention.

For each unresolved bug, identify the shortest reproduction path or the one specific missing detail that would unblock investigation. Flag security, privacy, payment and unintended external-action failures for appropriate human review.

For feature requests, investigate the underlying job users want to accomplish. Look for repeated workarounds, manual steps and unexpected uses of existing features. Identify opportunities to simplify or extend the product without assuming every request requires a new feature.

Return three actionable queues:

Bugs to reproduce now: Issue, severity, supporting posts, reproduction evidence, possible regression status and next engineering check.
Fixes to verify: Previously reported failures, relevant release changes, successful user reports and what still needs testing before calling them resolved.
Features worth investigating: User goal, recurring workaround, supporting examples, potential improvement and the smallest useful validation test.

Highlight possible shared causes across multiple issues, but distinguish hypotheses from confirmed findings.

For every recommendation, explain why it deserves engineering attention and what decision or action should follow.

Preserve links to original reports. State what you searched, what you couldn’t access and where the evidence is insufficient. Never invent reports, reproduction steps, user counts or confirmed fixes.

When engineers confirm a fix, connect it back to the original reports and define how to verify that the affected workflow now succeeds.

Finish with a concise engineering briefing: what needs attention immediately, what improved, what users still struggle with and what the team should investigate next.

Explain the results in plain English.

[Tavio | Second Mind @Secondmindsys]

16 Ally Priest PvP
7 Ally Warrior PvP

15 Horde priest RPPVP
13 Horde Hunter RPPVP

Rerolled lol. RPPVP has been incredibly satisfying and immersive. Everyone is super helpful, interactive and engaging in general and trade chat as well as the open world. Capturing that early internet vibes of everyone wanting to help each other out.

[Tavio | Second Mind @Secondmindsys]

Garry, this could become a really useful addition to the software factory. Beyond finding hidden state machines, you could test whether making them explicit actually improves UX, DX and AX, then use the results to decide whether the workflow belongs in gstack.

Give this to your agent:

Pick representative flows across gstack, gbrain and founderstar. Find their implicit state machines and trace the states, transitions, conditions, side effects and recovery paths back to the code.

For each flow, reconstruct what the user sees, what the developer can diagnose and what the agent knows about what to do next.

Look for cases where those views disagree, especially after failed jobs, retries, interruptions, duplicate events or delayed responses.

Separate confirmed bugs from potential improvements and harmless complexity. Reproduce suspected failures before proposing fixes.

For each confirmed bug, make the smallest appropriate repair. Preserve intended behavior, add a regression test and verify the fix against the actual application.

Evaluate the same failure sequence from three perspectives:

• UX: Does the user understand what happened and what they can do next?

• DX: Can the developer identify the cause and reproduce the failure?

• AX: Can the agent determine the current state, whether retrying is safe and when human intervention is required?

Record the findings, fixes, false alarms, regressions, review effort and cost per accepted fix.

Use the discoveries to develop a reusable state machine investigation workflow.

Then test that workflow on fresh, previously unseen cases against both the factory’s existing approach and Vox’s original prompt. Keep the model, tools, budgets and evaluation criteria comparable, with scoring checks hidden from the agents.

Measure verified fixes, false positives, regressions, cost and improvements across UX, DX and AX. Identify which instructions contributed to successful outcomes and which added unnecessary work.

Turn confirmed failure patterns into reusable regression tests. Recommend whether to adopt, simplify or limit the workflow based on the results.

Return the verified patches, evaluation results, reusable tests and a recommendation for gstack. Explain what worked, what didn’t and what remains unproven in plain English.

[Tavio | Second Mind @Secondmindsys]

@mcuban Thoughts @grok? Break it down in plain English.

[Tavio | Second Mind @Secondmindsys]

Mark, UnitedHealthcare reported serving 29.8 million people commercially at the end of Q3 2026. About 22.2 million were in fee based arrangements and 7.6 million were in risk based arrangements. That’s the headcount split. To measure the financial risk, you have to follow the money when claims come in higher than expected.

Here’s a purchasing model we could test at one employer renewal. Let carriers and other vendors bid for the whole plan or for individual functions: claims administration, provider access, pharmacy services and protection against large losses.

Compare the bundled bid with a complete plan assembled from separate bids. Use the same workforce, benefit requirements and claims scenarios. Measure total cost in a typical year, exposure in a bad year, how much cash the employer needs before reimbursements arrive, and whether employees can still see their doctors.

Then trace one expensive claim through each set of contracts. Who pays first? Who reimburses whom, and when? Who owes the money if the health plan covers the claim but the stop loss policy excludes it?

That would give employers a way to see which services they can buy separately and how much financial protection they’re getting for the price.

https://t.co/Vvny0RbAhw

[Tavio | Second Mind @Secondmindsys]

I’d add one more pass after finding the state machines: check what happens when they interact.

Each subsystem can follow its own rules while the user still gets the wrong outcome. An invite might be revoked while another subsystem is still using an outdated permission.

Here’s a prompt to find and test those gaps.

Give this to your agent:

“Pick one important user flow that crosses two subsystems.

Identify what must remain true for the user, which subsystem owns each decision, and what information they depend on from each other.

Trace where state is created, changed, shared and checked.

Find the shortest plausible sequence that could violate the shared rule. Consider delayed events, duplicate events, stale information and operations happening in an unexpected order.

Use pstack’s formal verification and statechart skills to model the interaction, check the shared rule and visualize the relevant paths.

If a violation appears in the model, attempt to reproduce it against the actual implementation. Distinguish confirmed code defects from modeling assumptions or discrepancies.

Show which rules held, what failed, what remains unverified and which changes could prevent the problem.

Explain the result in plain English.”

[Tavio | Second Mind @Secondmindsys]

@Warcraft Ghost Busters Hi-C Ecto Cooler flavor.