How we actually do this.
The mechanism, the math, and the code
Most AI services sell effort: a person, working harder, with better tools. We sell a governed system: one whose quality is proven before it ships and whose proof you can read. This page is the engine room: the harness, the statistics that replace the dashboard, and the open-source repositories that back every claim. No mechanism, no claim.
How we actually do this. The mechanism, the math, and the code Most AI services sell effort: a person, working harder, with better tools. We sell a governed system: one whose quality is proven before it ships and whose proof you can read. This page is the engine room: the harness, the statistics that replace the dashboard, and the open-source repositories that back every claim. No mechanism, no claim. ← back to proof Read the source → Talk to an engineer →
What the system is built to minimize.
the objective
Every engagement points at one objective: deliver the outcome with the fewest moving parts and the least surprise. Stated formally, we minimize a free-energy functional: the gap between what the system predicts and what actually happens, plus the cost of complexity.
free-energy objective F = 𝔼q[ −ln p(o | s) ] + D KL( q ‖ p ) The first term is surprise, a measure of how badly the system's model of "done" predicts the real outcome o given state s. The second is complexity, how far our chosen approach q strays from the simplest one that works. Minimizing both at once is the whole discipline: maximum fidelity, minimum machinery. Effort that doesn't reduce surprise or complexity isn't progress.
What the system is built to minimize. the objective Every engagement points at one objective: deliver the outcome with the fewest moving parts and the least surprise. Stated formally, we minimize a free-energy functional: the gap between what the system predicts and what actually happens, plus the cost of complexity. free-energy objective F = 𝔼q[ −ln p(o | s) ] + D KL( q ‖ p ) The first term is surprise, a measure of how badly the system's model of "done" predicts the real outcome o given state s. The second is complexity, how far our chosen approach q strays from the simplest one that works. Minimizing both at once is the whole discipline: maximum fidelity, minimum machinery. Effort that doesn't reduce surprise or complexity isn't progress.The offering is one signed tuple.
We don't ship prompts. We ship a versioned, signed harness: six components that turn a vague "AI project" into a governed system with a contract.
The workspace and the real data sources the agent is allowed to touch. S coped, not the open internet.
An allowlist of executable procedures. Anything external or destructive (sends, deploys, payments) requires a human approval gate.
The proof obligation. E very cycle must produce a ProofPacket of measured results. T he agent never self-declares "done."
V alidators that block bad output before it ships, and ceilings that stop an agent from acting outside its bounds.
An inner quality loop per deliverable, a weekly evolution loop, and a self-breeding loop that crystallizes repeated wins into reusable skills.
An independent evaluator with sole authority to terminate. It holds the result against threshold and decides whether it ships. Nothing else can call it done.
the harness H = ( E, T, C, S, L, V )
The offering is one signed tuple. We don't ship prompts. We ship a versioned, signed harness: six components that turn a vague "AI project" into a governed system with a contract. The workspace and the real data sources the agent is allowed to touch. S coped, not the open internet. An allowlist of executable procedures. Anything external or destructive (sends, deploys, payments) requires a human approval gate. The proof obligation. E very cycle must produce a ProofPacket of measured results. T he agent never self-declares "done." V alidators that block bad output before it ships, and ceilings that stop an agent from acting outside its bounds. An inner quality loop per deliverable, a weekly evolution loop, and a self-breeding loop that crystallizes repeated wins into reusable skills. An independent evaluator with sole authority to terminate. It holds the result against threshold and decides whether it ships. Nothing else can call it done. the harness H = ( E, T, C, S, L, V ) Component What it is in your system EEnvironment TTools CContract SSafety gates LLoops VVerifierThe maker is never the grader.
Does the work
Critiques it
Decides ship
Approves risk
separation of powers
The single most important rule. The thing that does the work is not allowed to decide the work is finished. That's where silent failures come from. We separate four roles, and no agent reviews its own output.
Builds the artifact: the audit, the automation, the result. Tuned to produce, not to judge.
Checks each artifact: is it grounded in real data, or plausible-but-invented? Catches the hallucination.
Holds the ProofPacket against threshold θ. Sole termination authority. Only this clears a cycle.
You approve external, irreversible, or paid actions. Everything safe runs autonomously.
generator
evaluator
verifier
human
The maker is never the grader. Does the work Critiques it Decides ship Approves risk separation of powers The single most important rule. The thing that does the work is not allowed to decide the work is finished. That's where silent failures come from. We separate four roles, and no agent reviews its own output. Builds the artifact: the audit, the automation, the result. Tuned to produce, not to judge. Checks each artifact: is it grounded in real data, or plausible-but-invented? Catches the hallucination. Holds the ProofPacket against threshold θ. Sole termination authority. Only this clears a cycle. You approve external, irreversible, or paid actions. Everything safe runs autonomously. generator evaluator verifier humanA green dashboard can't tell signal from noise.
honest statistics, not a dashboard
Most AI work is graded by eyeballing a chart. We grade on probability. Three instruments do the work.
surprisal: is this result real? S(o) = −ln P(o) Before a reported lift goes in a report, we score how surprising it is under the noise model. A "win" that's statistically unremarkable is noise wearing a suit. Counts get modeled as Poisson / Negative-Binomial and we report an upper-tail p-value, not a screenshot of a number going up.
robust drift detection z = 0.6745 · ( x − median ) / MAD A robust z-score (median + median-absolute-deviation, immune to outliers) watches every metric. A reading that crosses ±5σ (a sudden rank drop, a quality collapse, an anomalous spike) automatically triggers deep review instead of quietly riding along.
calibrated confidence before we commit promote iff Brier-validated · E open ≈ 0 We only promote a decision ("scale this," "this is done") once its confidence is Brier-calibrated (our stated probabilities match reality over time) and the deliverable's residual open energy is near z ero. Confidence is the gate, not the calendar.
A green dashboard can't tell signal from noise. honest statistics, not a dashboard Most AI work is graded by eyeballing a chart. We grade on probability. Three instruments do the work. surprisal: is this result real? S(o) = −ln P(o) Before a reported lift goes in a report, we score how surprising it is under the noise model. A "win" that's statistically unremarkable is noise wearing a suit. Counts get modeled as Poisson / Negative-Binomial and we report an upper-tail p-value, not a screenshot of a number going up. robust drift detection z = 0.6745 · ( x − median ) / MAD A robust z-score (median + median-absolute-deviation, immune to outliers) watches every metric. A reading that crosses ±5σ (a sudden rank drop, a quality collapse, an anomalous spike) automatically triggers deep review instead of quietly riding along. calibrated confidence before we commit promote iff Brier-validated · E open ≈ 0 We only promote a decision ("scale this," "this is done") once its confidence is Brier-calibrated (our stated probabilities match reality over time) and the deliverable's residual open energy is near z ero. Confidence is the gate, not the calendar.Autonomy is earned, not assumed.
when the machine acts vs when you do
Each capability gets a build-mode score (BMS). The higher the proven confidence, the more autonomy it's granted. Anything below threshold escalates to a human.
build-mode routing BMS ≥ .75 → A1, .45 → A2, .25 → A3, < .25 → A4 A1 runs deterministically, on its own. A2 is a supervised hybrid. A3 is agent-bounded with checkpoints. A4 is human-led with the AI assisting. Any public or irreversible action requires your approval regardless of mode. The system knows what it doesn't know.
Autonomy is earned, not assumed. when the machine acts vs when you do Each capability gets a build-mode score (BMS). The higher the proven confidence, the more autonomy it's granted. Anything below threshold escalates to a human. build-mode routing BMS ≥ .75 → A1, .45 → A2, .25 → A3, < .25 → A4 A1 runs deterministically, on its own. A2 is a supervised hybrid. A3 is agent-bounded with checkpoints. A4 is human-led with the AI assisting. Any public or irreversible action requires your approval regardless of mode. The system knows what it doesn't know.It gets cheaper every time.
Grill loop
Evolution loop
Self-breeding loop
inner Grill loop Each deliverable runs ≥3 question/answer cycles and only advances when its open energy is ≈0. Kills confident-but-ungrounded output at the source.Grill loop
weekly Evolution loop We log (input, output, energy, outcome) every cycle, update reliability weights, and recalibrate thresholds. We fix the worst-performing part first.Evolution loop
on repeat Self-breeding loop On the third repeat of a pattern, the proven blocks crystallize into a sealed, reusable harness. The tenth instance is near-instant.Self-breeding loop
why it compounds
Three loops turn a one-off build into a system that improves itself.
Each deliverable runs ≥3 question/answer cycles and only advances when its open energy is ≈0. Kills confident-but-ungrounded output at the source.
We log (input, output, energy, outcome) every cycle, update reliability weights, and recalibrate thresholds. We fix the worst-performing part first.
On the third repeat of a pattern, the proven blocks crystallize into a sealed, reusable harness. The tenth instance is near-instant.
It gets cheaper every time. Grill loop Evolution loop Self-breeding loop why it compounds Three loops turn a one-off build into a system that improves itself. Each deliverable runs ≥3 question/answer cycles and only advances when its open energy is ≈0. Kills confident-but-ungrounded output at the source. We log (input, output, energy, outcome) every cycle, update reliability weights, and recalibrate thresholds. We fix the worst-performing part first. On the third repeat of a pattern, the proven blocks crystallize into a sealed, reusable harness. The tenth instance is near-instant.The repositories that back this.
rigforgeopen · MIT The engine. Catch your AI coding agent when it lies about "done." Cryptographically signed ProofPackets make agent output re-verifiable (MCP + OpenTelemetry). request access → proofpacket-sealing-templatefree Seal any build into a hash-verifiable ProofPacket. Proof, not vibes. The fastest way to feel the mechanism on your own code. request access → rig-deterministic-build-starteropen A Vite + TypeScript starter that ships with a sealed ProofPacket: deterministic CI/CD, green tests, and a signed Definition of Done. request access → rig-prediction-studioteaser Offline, Brier-calibrated 12-persona forecasting swarm. The calibration math on this page, made executable. Full source on purchase. request access → rig-operating-proceduresteaser Deterministic spec-to-app scaffold generator with packet-chain enforcement. The harness discipline as a tool. Full source on purchase. request access → RIG-Doctrine-Repoopen The canonical operating doctrine, agent packs, and proof standards. The rules every RIG build follows, in writing. request access → Get in touchdon't take our word: read the source
The core engine is open source. The free template lets you seal your own build into a verifiable ProofPacket today. This is the difference between a claim and a receipt.
Studios marked teaser are public previews; full source ships on engagement. Get in touch .
The repositories that back this. don't take our word: read the source The core engine is open source. The free template lets you seal your own build into a verifiable ProofPacket today. This is the difference between a claim and a receipt. Studios marked teaser are public previews; full source ships on engagement. Get in touch . rigforgeopen · MIT The engine. Catch your AI coding agent when it lies about "done." Cryptographically signed ProofPackets make agent output re-verifiable (MCP + OpenTelemetry). request access → rigforge open · MIT request access → proofpacket-sealing-templatefree Seal any build into a hash-verifiable ProofPacket. Proof, not vibes. The fastest way to feel the mechanism on your own code. request access → proofpacket-sealing-template free rig-deterministic-build-starteropen A Vite + TypeScript starter that ships with a sealed ProofPacket: deterministic CI/CD, green tests, and a signed Definition of Done. request access → rig-deterministic-build-starter open rig-prediction-studioteaser Offline, Brier-calibrated 12-persona forecasting swarm. The calibration math on this page, made executable. Full source on purchase. request access → rig-prediction-studio teaser rig-operating-proceduresteaser Deterministic spec-to-app scaffold generator with packet-chain enforcement. The harness discipline as a tool. Full source on purchase. request access → rig-operating-procedures RIG-Doctrine-Repoopen The canonical operating doctrine, agent packs, and proof standards. The rules every RIG build follows, in writing. request access → RIG-Doctrine-RepoWire the verifier first.
Wire E + seal V
Crystallize generators
Ship under GEV
days 1-3 Wire E + seal V Connect the data sources and stand up the independent verifier and ProofPacket schema before any build. Measurement first.Wire E + seal V
days 4-10 Crystallize generators Install the work as governed procedures with allowlists and approval gates. Grill-loop on a test target until open energy ≈ 0.Crystallize generators
days 11+ Ship under GEV First real deliverable runs end-to-end under verifier gating, with a human approval on every external action. Then the evolution loop goes live.Ship under GEV
how an engagement runs
The sequence is deliberate: we stand up measurement and the independent verifier before a single deliverable, so nothing can ship unproven. Intent → Questions → Research → Spec → Quality → Plan → Implement. The loop runs until confidence, not until the clock runs out.
Connect the data sources and stand up the independent verifier and ProofPacket schema before any build. Measurement first.
Install the work as governed procedures with allowlists and approval gates. Grill-loop on a test target until open energy ≈ 0.
First real deliverable runs end-to-end under verifier gating, with a human approval on every external action. Then the evolution loop goes live.
Wire the verifier first. Wire E + seal V Crystallize generators Ship under GEV how an engagement runs The sequence is deliberate: we stand up measurement and the independent verifier before a single deliverable, so nothing can ship unproven. Intent → Questions → Research → Spec → Quality → Plan → Implement. The loop runs until confidence, not until the clock runs out. Connect the data sources and stand up the independent verifier and ProofPacket schema before any build. Measurement first. Install the work as governed procedures with allowlists and approval gates. Grill-loop on a test target until open energy ≈ 0. First real deliverable runs end-to-end under verifier gating, with a human approval on every external action. Then the evolution loop goes live.What engineers ask first.
Yes. The core engine, rigforge, is open source under MIT, and the ProofPacket sealing template is free. Both are linked above. Several studios are public teasers with full source available on purchase.
A cryptographically signed, hash-bound record of a result: the evidence behind every claim, the gates it passed, and the verifier that signed off. It makes agent output re-verifiable instead of taken on faith.
A green dashboard can't tell signal from noise. We model results as count distributions and report tail p-values, use a robust z-score to catch drift at ±5σ, and only promote a decision once its confidence is Brier-calibrated.
Only inside the allowlist and only on actions scored safe. Anything external, irreversible, or paid stops for your typed approval. The human is always the decision-maker on risk.
Can I see the source code? Yes. The core engine, rigforge, is open source under MIT, and the ProofPacket sealing template is free. Both are linked above. Several studios are public teasers with full source available on purchase. What exactly is a ProofPacket? A cryptographically signed, hash-bound record of a result: the evidence behind every claim, the gates it passed, and the verifier that signed off. It makes agent output re-verifiable instead of taken on faith. Why statistics instead of a dashboard? A green dashboard can't tell signal from noise. We model results as count distributions and report tail p-values, use a robust z-score to catch drift at ±5σ, and only promote a decision once its confidence is Brier-calibrated. Does the AI ever act on its own against my systems? Only inside the allowlist and only on actions scored safe. Anything external, irreversible, or paid stops for your typed approval. The human is always the decision-maker on risk.Can I see the source code? Yes. The core engine, rigforge, is open source under MIT, and the ProofPacket sealing template is free. Both are linked above. Several studios are public teasers with full source available on purchase.
What exactly is a ProofPacket? A cryptographically signed, hash-bound record of a result: the evidence behind every claim, the gates it passed, and the verifier that signed off. It makes agent output re-verifiable instead of taken on faith.
Why statistics instead of a dashboard? A green dashboard can't tell signal from noise. We model results as count distributions and report tail p-values, use a robust z-score to catch drift at ±5σ, and only promote a decision once its confidence is Brier-calibrated.
Does the AI ever act on its own against my systems? Only inside the allowlist and only on actions scored safe. Anything external, irreversible, or paid stops for your typed approval. The human is always the decision-maker on risk.
straight answers
What engineers ask first. straight answersBring me a result you can't currently trust.
Book a free 30-minute call → See governed agents → mike@rodgersintelligence.com (262) 343-5680the next move
Tell me where your AI output is taken on faith. I'll show you exactly how the verifier, the math, and the proof packet would close that gap. It goes straight to my inbox and I answer it myself.
Prefer email or text? mike@rodgersintelligence.com · (262) 343-5680 . I reply within 24h, personally.
Bring me a result you can't currently trust. the next move Tell me where your AI output is taken on faith. I'll show you exactly how the verifier, the math, and the proof packet would close that gap. It goes straight to my inbox and I answer it myself. Prefer email or text? mike@rodgersintelligence.com · (262) 343-5680 . I reply within 24h, personally. Book a free 30-minute call → See governed agents →vfoot
Real problem, real budget? Email or text me. I answer it myself.
(262) 343-5680 · Mike@RodgersIntelligence.com
Real problem, real budget? Email or text me. I answer it myself. vfoot (262) 343-5680 · Mike@RodgersIntelligence.com Email Mike → Text Mike →