A practical operating guide for framing, reasoning, acting, verifying, and learning
These pages are a reasoning operating system, not a poster set. Page 1 is the map. Pages 2 through 7 move from context to verified action and learning. Page 8 is the scientific learning layer that can operate throughout the cycle.
The framework is sequential where sequence matters, but deliberately recursive. Move forward only when the current page has produced what the next page needs. When evidence, state, or execution invalidates the reasoning, return to the earliest point where it became false.
Do not advance because a page has been completed. Advance because the next stage has what it needs. Completion is not the gate. Evidence quality, causal adequacy, authority, and verified state are the gates.
| Page | Purpose | Required output | Go back if... |
|---|---|---|---|
| 1 | Orient the work. | Shared map of the current reasoning state. | People are debating details before agreeing where they are. |
| 2 | Establish state, regime, boundaries, and unknowns. | Shared structured situation summary. | Context is missing, stale, disputed, or changing. |
| 3 | Set scale and scope, then qualify evidence. | Bounded evidence set with uncertainty explicit. | Evidence is weak, stale, contradictory, or at the wrong scale. |
| 4 | Build causal structure and mechanisms. | Causal model with conditions and scale limits. | You only have correlation, a black box, or a single-scale model. |
| 5 | Stress the model and compare explanations. | Risk map plus explanations, confidence, and falsifiers. | The conclusion fails plausible stress or alternatives. |
| 6 | Choose and authorize action. | Action plan with monitoring, reversibility, and triggers. | The action is not aligned, authorized, monitorable, or reversible enough. |
| 7 | Execute, verify, compare expected with actual, and learn. | Replayable execution record, verified state, and variance analysis. | The result is unverified, production differs, or reality contradicts the model. |
| 8 | Run the scientist loop and preserve learning. | Reusable learning, safe updates, and verified output. | The result is not explainable, reproducible, or safe to generalize. |
Page 1 is the navigation map; the detailed page is authoritative when the overview compresses a stage. Page 8 is cross-cutting, so invoke the Scientist Loop whenever a claim can be tested or falsified.
Use it to orient the decision, not to do the detailed work.
Stage crosswalk Page 1 = overview (all stages)
At the beginning of a decision, at the beginning of a workshop, or whenever people are solving different parts of the problem without realizing it.
Locate the work in the architecture. Identify what is already known, what stage is active, which outputs already exist, and which stage should own the next question.
A one-sentence stage declaration: “We are at Page X because Y is established, Z is still unresolved, and the next required output is ____.”
Everyone can explain why the current stage is the right stage and what the next stage requires.
The group is arguing across stages. For example, one person is proposing actions while another is still disputing the facts.
Establish the situation before the system is allowed to reason forward.
Stage crosswalk Page 2 = Stage 0 (establish the active situation)
A new problem arrives, the operating environment changes, a previous conclusion is being reopened, or the team is not confident that it is looking at the same situation.
Turn raw inputs into a shared current-state model. Separate what is happening now, what happened before, what is allowed, and what is changing over time.
A structured situation summary with current state, known facts, unknowns, assumptions, regime, boundaries, constraints, and material risks/opportunities.
The active situation is clear enough that evidence can be sought against a specific question, scope, and time horizon.
New information changes the regime; assumptions fail; context is stale; key inputs are missing; or the situation cannot be described without argument.
Define where to look and what deserves to count as evidence.
Stage crosswalk Page 3 = scale/scope/boundary/regime control + Stage 1 (find and qualify evidence)
The question is clear enough to investigate, but evidence is scattered, inconsistent, noisy, or comes from different levels of the system.
Set scale, scope, boundary, and regime first. Then find evidence that is relevant, trustworthy, current, complete enough, and appropriate to the question.
An evidence ledger: source, claim supported, scale, recency, credibility, contradictions, uncertainty, missing evidence, and applicability to the current regime.
The evidence set is good enough to support causal work, and the team can explain both what the evidence says and what it does not justify.
The evidence changes, contradictions are unresolved, important sources are missing, or the scope has shifted enough to make the existing evidence incomparable.
Move from correlation to an explanation that can survive scale changes.
Stage crosswalk Page 4 = Stage 2 (causal structure and mechanisms)
You have qualified evidence and need to explain what actually produces the outcome, not merely what moves with it.
Build a causal model with variables, mechanisms, conditions, interactions, feedback loops, emergent risks, and dynamics. Qualify the model by scale.
A causal map or mechanism model that names the important variables, relationships, conditions, feedbacks, time dynamics, and scale boundaries.
The team can explain how change propagates from cause to outcome and can identify where an intervention would plausibly change the result.
The model depends on unexplained correlation, contains a black-box mechanism, fails at another relevant scale, or cannot explain important observed behavior.
Try to break the model before reality does.
Stage crosswalk Page 5 = Stages 3 and 4 (what could break the outcome; cause and effect)
A causal story exists, but before acting you need to know how it can fail and whether another explanation fits the evidence as well or better.
Stress-test the causal model and reason through competing explanations. The goal is not to defend the first plausible story. The goal is to find the explanation that survives evidence, counterfactuals, and failure analysis.
A fragility and explanation record containing failure modes, boundary conditions, early-warning signals, dependencies, competing hypotheses, counterfactuals, confidence, and falsifiers.
The preferred explanation has survived serious alternatives and the team knows what evidence would lower its confidence or overturn it.
A plausible failure mode is unmanaged, an alternative explanation fits equally well, confidence is being asserted without falsifiability, or the evidence cannot distinguish among hypotheses.
Turn understanding into authorized, monitored, reversible action.
Stage crosswalk Page 6 = Stage 5 (choose actions) + permission and trust architecture
The evidence and causal reasoning are strong enough to justify intervention, and the next question is what to do.
Generate options, compare impact and trade-offs, choose an action, define implementation and monitoring, and ensure the action sits inside explicit permission and trust boundaries.
An authorized action record: selected option, expected impact, rationale, trade-offs, owner, permissions, implementation plan, monitoring plan, thresholds, reversal path, and learning objective.
The action is feasible, authorized, monitorable, sufficiently reversible for its risk, aligned with values and policy, and designed to generate useful evidence.
Authority is unclear; constraints are violated; no one owns the outcome; monitoring is missing; or the action is too irreversible for the confidence level.
Reality is the gate. No claim survives without evidence of what actually happened.
Stage crosswalk Page 7 = Stages 6, 7, 8 (6 execution truth/provenance/current-state gate, 7 control the outcome over time, 8 production matches design) + compare expected with actual
An action has been approved and execution is beginning, or a result has occurred and must be verified before anyone claims success or failure.
Capture provenance, verify state change, control the outcome over time, make sure production matches design, compare expected with actual, and decide what must update.
A replayable execution record with provenance, verified current state, outcome metrics, expected-versus-actual variance, root-cause attribution, production validation, and update decision.
The state change is verified, the result is explainable, the production system matches the intended design, and the next action or terminal state is justified by evidence.
The current state cannot be verified; production differs from design; metrics conflict; the result surprises the model; or the variance cannot yet be localized.
Use experimentation, memory, and test harnesses to compound learning safely.
Stage crosswalk Page 8 = automated scientist loop, memory and test harnesses (cross-cutting)
Throughout the process when a claim can be tested, and at the end of a cycle when results should update memory, models, policies, or future experiments.
Turn questions into hypotheses, falsification tests, measurable predictions, evidence-quality checks, updates, and observed outcomes. Preserve what was learned and prove that improvements do not create regressions.
A scientist-loop record plus durable memory: hypotheses, tests, predictions, observations, updates, test-harness results, reusable lessons, and a verified terminal output.
The result is usable, verifiable, reproducible enough for its risk, safely incorporated into memory or policy, and remains reviewable by humans.
The result cannot be reproduced, the update causes regressions, the evidence is not transportable, the system cannot explain what changed, or human accountability is unclear.
The framework can be used three ways: as a human decision-room method, as an AI or agent orchestration pattern, or as a hybrid in which machines carry evidence and reasoning work while humans retain authority for consequential outcomes.
Use one page at a time as the meeting agenda. Assign a decision owner, evidence owner, red-team role, and recorder. Keep Page 1 visible as the map. Record the required output before moving forward.
Map each page to an explicit state, prompt, tool set, memory object, and transition gate. The agent is not allowed to advance when the stage output or confidence criteria are missing.
Let the system retrieve, compare, model, simulate, and maintain provenance. Reserve approval, value judgments, authority changes, exceptions, and consequential intervention for accountable humans.
| Time | Page | What happens |
|---|---|---|
| 5 min | Page 1 | State the decision, identify the active page, and name the missing output. |
| 10 min | Page 2 | Establish current state, regime, boundaries, unknowns, and constraints. |
| 10 min | Page 3 | Agree on scale and evidence quality. Mark contradictions and missing evidence. |
| 10 min | Pages 4-5 | Build the causal explanation, then attack it with failure modes and alternatives. |
| 10 min | Page 6 | Choose an action, authority path, monitoring plan, and adaptation triggers. |
| 10 min | Page 7 | Define what must be observed to prove the action worked and how variance will be handled. |
| 5 min | Page 8 | Define the test, memory update, replay requirement, and what should generalize if the result holds. |
The higher the consequence, irreversibility, uncertainty, novelty, or cross-scale impact, the more rigor the framework should demand. Low-consequence work can use a light pass. High-consequence work should preserve full evidence, provenance, alternatives, permissions, tests, and review.
The architecture becomes operational when every major step leaves a durable reasoning receipt. These are not paperwork for its own sake. They are the evidence that allows the next stage, a reviewer, or a future system to understand why the decision was made and what would change it.
| Artifact | Source | Minimum contents |
|---|---|---|
| Situation brief | Page 2 | Current state, regime, boundaries, assumptions, unknowns, constraints, time horizon. |
| Evidence ledger | Page 3 | Source, claim, scale, recency, trust, contradictions, uncertainty, applicability, missing evidence. |
| Causal model | Page 4 | Variables, mechanisms, conditions, interactions, feedback loops, dynamics, scale limits. |
| Fragility + hypothesis record | Page 5 | Failure modes, dependencies, early warnings, alternatives, counterfactuals, confidence, falsifiers. |
| Action and authority record | Page 6 | Option, expected impact, trade-offs, reversibility, permissions, owner, monitoring, adaptation triggers. |
| Execution and variance record | Page 7 | What was executed, provenance, verified state, expected vs actual, root cause, production integrity, update decision. |
| Experiment + memory record | Page 8 | Hypotheses, tests, predictions, observations, updates, regressions, reusable learning, terminal status. |
| Gate | Question |
|---|---|
| State gate | Do we know what state the world is actually in, not merely what we expected? |
| Evidence gate | Is the evidence relevant, trustworthy, current enough, complete enough, and at the right scale? |
| Causal gate | Can we explain the mechanism, conditions, and likely failure modes, not just the correlation? |
| Authority gate | Are we allowed to act, is the action aligned, and does an accountable human own the outcome? |
| Verification gate | Can we prove what was executed, what changed, why it changed, and what should update next? |
Most failures should not restart the entire architecture. Return to the earliest stage where the assumption, evidence, model, permission, or execution record first became false. That keeps correction local and makes learning cumulative.
The purpose of the framework is not to make reasoning look rigorous. It is to make reasoning correctable. Every important claim should have evidence. Every causal story should have a failure surface. Every action should have authority and a monitoring plan. Every execution should leave provenance. Every surprise should update what the system knows. And every consequential outcome should remain owned by a human.
Know the state. Qualify the evidence. Explain the cause. Try to break the explanation. Act within authority. Verify reality. Learn from the difference.