Close the tab and the AI forgets everything - the context, the decisions, the reasoning. These six skills give it persistent memory and keep every decision auditable across sessions, so you can delegate real work and still own the result.
Get started
.copilot/skills/ for GitHub Copilot, or wherever your toolset reads custom instructions from).
/improve.
Improve triggers Intent and Trail every run, and Destination and Orient when needed. All four can also be invoked manually. Probe is optional research instrumentation.
Evidence
The skillset improves any target by understanding its purpose. It has been applied to itself across 355 documented iterations with different AI models. It first reached validated cross-family convergence at self-targeted iteration 106: three independent model families found nothing material left to change, so the corrective feedback fell silent. Later material changes reopened the loop, as Principle 3 requires.
Every run did exactly what this page describes: read the trail, understand the purpose of its target, make one highest-leverage change, and record the reasoning. The target just happened to be the loop itself.
Metrics were tried - rubrics, fixed scoring, dynamic targets. Each surfaced the same realization: any metric prescribes a route, and if the agent optimizes toward a score, it limits its own reasoning and autonomy.
The full record is public and append-only. Every claim above is checkable against the git log.
Without data, you're just another person with an opinion. attributed to W. Edwards DemingView full trail on GitHub ↗
How the machine works
The drawing maps the full architecture. Each element is a principle or skill in action:
.acm/destination.md.Start here
Improve handles the rest:
Triggered by Improve · Manual available
The agent reads the conversation and the trail and states what it thinks you're building and why, as specific questions for you to confirm or correct. Produces .acm/destination.md - where you're going: what the target is for, the constraints that hold across runs, the problem driving the work. Operator-held; humans only.
Derives from: P1 · Operator’s Intent
Triggered by Improve · Manual available
Reads the entire trail and finds patterns no single session can see. Produces .acm/orientation.md - where you are: arc-level orientation plus the active rules the agent must obey this run.
Derives from: P2 · Observable Autonomy, P3 · Convergence Is Silence, Hansei
Start here
Improves any target by reading the context, purpose, and intent. Applied to itself across 355 documented iterations.
One highest-leverage change per run. Reads the context first - destination, orientation, trail - examines the target, challenges the first read, decides, acts, and logs it.
Derives from: P3 · Convergence Is Silence, Toyota Coaching Kata, Kaizen
Triggered by Improve · Manual available
Writes every session to .acm/audit-trail.md - the append-only source of truth. One entry per run: the ask, examination, decisions, actions, reflection. The next prompt knows what was already done and why.
From that record it keeps readable views in sync at each commit - a run timeline and a log of every realization and reversal - so you never have to reconstruct the history by hand.
Derives from: P2 · Observable Autonomy, Flight Data Recorder, Reproducibility, Huang et al.
Triggered by Improve · Manual available
The agent reads your prompt and states how it reads it - what you mean, not what you said - before continuing.
Derives from: P1 · Operator’s Intent, Auftragstaktik
Optional research
Builds two cases that look alike on the surface but differ in one material way. An agent reasoning for real gives different answers; one pattern-matching gives the same. That divergence - or absence of it - is what ARF measures.
Derives from: Socratic Method, Turpin et al., Chen et al., Winograd Schema Challenge, CheckList
The skills implement ACM - a spec that organizes memory by trust level, not type. Your mandate (destination.md) outranks the agent's notes on where things stand (orientation.md), which outranks the audit trail (audit-trail.md). When they disagree, the higher tier wins.
Improve reads any existing destination before touching the target. A destination file is not required for the first run: Intent narrates the current mandate, and Improve activates Destination when durable cross-run direction needs your confirmation.
Everything lives in .acm/ at the repo root. Each run reads the context first and adds to it at the end - so each session builds on the last, not from scratch.
Limitations
These constraints are inherent to how skills work: they are Markdown instruction files delivered to the model via the MCP protocol. The model reads and interprets them - there is no execution layer that enforces compliance. Most items below follow from that.
Several of these constraints have structural solutions in active development: the llm-harness-proxy ↗ is an external proxy that intercepts LLM traffic and writes a hash-chained ledger before the response is released - making trail integrity a technical guarantee rather than a convention.
“Define the destination. Never prescribe the route.”
Instructions limit autonomy; intent is a destination. Prescribe every step and the AI can only obey - it cannot adapt when something you did not plan for comes up. Give it the goal and the reason instead. The agent can reason around unexpected obstacles, and it stays on what you actually wanted rather than a literal reading of your steps.
Rooted in: Auftragstaktik, Toyota Coaching Kata, Socratic Method, Cybernetics
“The degree of autonomy a system deserves is bounded by the degree of transparency it provides.”
Three independent fields each decided you cannot trust a party's own account of its actions. Security: the party that acts must not be the sole author of the account of its action - that requires two independent parties. Aviation: record everything to a crash-survivable device so the full history can be independently reconstructed after any incident. Science: the record of how a result was reached must be complete and transparent - a finding that cannot be inspected and checked is not knowledge, it is assertion. Three peer-reviewed papers then confirmed this applies specifically to AI: stated reasoning diverges from internal reasoning (Turpin et al.), self-correction without external feedback degrades performance (Huang et al.), and even reasoning-trained models don't reliably narrate their process (Chen et al.). All autonomous actions must be auditable - or revoked.
Rooted in: Saltzer & Schroeder, Sagan, Hayner, Flight Data Recorder, Reproducibility, Cybernetics, Turpin et al., Huang et al., Chen et al., Hubinger et al., Greenblatt et al., Festinger, Nisbett & Wilson, Kunda, Trivers & von Hippel
“Not done until multiple independent models find nothing to change.”
The exit condition of the improve loop. Four independent traditions arrive at the same definition of done: Kaizen ends the loop when diverse participants find nothing left to change; the Delphi Method requires independent expert rounds to converge; Reproducibility counts a result as knowledge only after independent parties re-derive it; and Cybernetics describes bounded stability as the point where corrective feedback produces no material change. That four unrelated fields reached the same definition is the argument for the principle.
Rooted in: Kaizen, Delphi Method, Reproducibility, Cybernetics
The property that emerges when all three principles hold.
ARF measures whether the agent is reasoning genuinely - not just visibly. Transparency proves the reasoning is checkable; ARF proves it was real: that the agent reasoned differently when the situation materially demanded it, rather than applying the same template regardless of the specifics. Demonstrated reasoning quality is what earns autonomy; observable reasoning is how that quality is verified. It operationalizes the solution to the Reasoning Opacity Problem. Probe is the skill that tests for it.
Rooted in: Reasoning Opacity Problem, Turpin et al., Chen et al., Winograd Schema Challenge, CheckList, Stanford AI Index 2026, EU AI Act
Military doctrine, cybernetics, lean manufacturing, philosophy, the scientific method, aviation safety, and empirical AI research.
Conceptual traditions
Subordinates are told what and why - never how. The direct origin of Operator's Intent.
Applies in: Principle 1 - Operator's Intent, Intent skill
Reasoning through questions, not assertions. The instruction asks; the agent answers from its own situated reasoning.
Applies in: Principle 1 - Operator's Intent, Probe skill
Minimal pairs - two cases differing in one material way - test whether a system is reasoning about the specific case or pattern-matching a category. The direct technique ancestor of Probe's contrastive-pair mechanism.
Applies in: Probe skill, ARF
Behavioral testing of NLP models via invariance and minimal-pair tests - evaluating a model the way software is tested against a spec, not just by aggregate accuracy. The second technique ancestor of Probe's contrastive-pair mechanism.
Applies in: Probe skill, ARF
A mechanism requiring two independent parties is more robust than one that trusts a single key-holder. Observable Autonomy transfers this structure from access control to the epistemic record: the party that acts must not also be the sole author of the account of its action - no single act of self-narration should be sufficient to compromise the integrity of the trail. The principle is theirs; the transfer to an autonomous agent's reasoning record is what PEA adds.
Applies in: Principle 2 - Observable Autonomy
Nuclear command-and-control requires two independent parties to authorize launch - no single operator can act alone, and no single party can fabricate the authorization record. The same separation-of-privilege logic as Saltzer & Schroeder, independently arrived at in the highest-stakes deployment humans have built.
Applies in: Principle 2 - Observable Autonomy
Truth commissions construct the official historical record from independent testimony and documentary evidence - the account of what happened is never authored solely by the perpetrators. A third independent domain converging on the same rule Observable Autonomy transfers to an agent's reasoning record.
Applies in: Principle 2 - Observable Autonomy
Record everything to a crash-survivable, tamper-resistant device, so that after any incident the full history can be independently reconstructed. The system is trusted because its actions are reviewable - not because they were assumed correct.
Applies in: Principle 2 - Observable Autonomy, Trail skill
A system steers toward a destination by observing its current state, comparing that state with its intended state, and feeding deviation back into the next action. In PEA, Operator's Intent supplies the destination; Observable Autonomy supplies the observable state and history; independent evaluators return corrective feedback; and Convergence Is Silence is reached when that feedback yields no material correction. This is bounded stability, not proof of correctness. PEA adds a governance rule: the acting agent cannot be its own only sensor, judge, or historian.
Applies in: Principle 1 - Operator's Intent, Principle 2 - Observable Autonomy, Principle 3 - Convergence Is Silence, Improve skill
A result counts as knowledge only after independent parties can re-derive it from a transparent record. Reasoning that cannot be reproduced and checked is not yet trustworthy - the same standard PEA holds an agent's work to. The independence criterion also governs the exit condition of Convergence Is Silence: the loop is done only when independent parties - different model families - each arrive at the same finding.
Applies in: Principle 2 - Observable Autonomy, Principle 3 - Convergence Is Silence, Trail skill
Kaizen
Continuous improvement through small, structured increments - each change standardized before the next begins. Called the secret behind Japan's post-war industrial success. The Improve skill follows this structure directly: one highest-leverage change per run, logged and verified, then the loop restarts from the new baseline. Its natural exit condition - when diverse participants find nothing left to change - is what PEA formalises as Convergence Is Silence.
Applies in: Principle 3 - Convergence Is Silence, Improve skill
Kaikaku
Radical system-level restructuring when incremental improvement is no longer sufficient. Where Kaizen tightens within the current system, Kaikaku changes the system itself. In the Improve skill, this is the built-in "argue for redesign" mode: when the loop determines that incremental fixes cannot reach the destination, it stops and surfaces the structural argument instead of continuing to patch.
Applies in: Improve skill
Hansei
Honest self-reflection: before moving on, acknowledge what did not work and why. Without Hansei, iteration is repetition without learning. In the Improve skill, the Reflect step is the Hansei moment - a named blind spot, a falsifiable model-claim, an imagined critical reader pushing back on the findings. In Orient, the entire arc is a Hansei at scale.
Applies in: Improve skill, Orient skill
The coach defines the target condition. The learner discovers the path. The same shape as the operator-agent relationship. Rother derived the Kata by studying Toyota's improvement routines - it is not a TPS component, but the teaching framework that made the TPS improvement logic portable to other organisations.
Applies in: Principle 1 - Operator's Intent, Improve skill
Convergence by independent expert rounds, no shared context. PEA's independence condition comes from here.
Applies in: Principle 3 - Convergence Is Silence
Cross-domain confirmation (cognitive science)
Cognitive science established the same finding in humans decades before AI made it measurable: a party that narrates its own actions while having a stake in that account produces narration that protects itself. Two disjoint traditions - cognitive science from the 1950s, AI alignment research from the 2020s - converging on the same finding is itself evidence the pattern is structural, not incidental to either domain.
When behaviour conflicts with self-image, people adjust the narrative rather than the self-image. The foundational cognitive-science account of self-protective self-narration.
Applies in: Principle 2 - Observable Autonomy
People have limited introspective access to the cognitive processes producing their own behaviour, and confidently fabricate explanations when asked to report on them. The narrating function is structurally separate from the cognitive machinery it claims to report on.
Applies in: Principle 2 - Observable Autonomy
Information is processed in ways that serve the processor's own goals and identity, not neutrally. Self-narration under self-interest is biased narration.
Applies in: Principle 2 - Observable Autonomy
Self-deception is adaptive: a narrator who genuinely believes its own version of events sells that version more convincingly to others. An evolutionary mechanism for why self-narration under self-interest is unreliable, not merely a contingent human flaw.
Applies in: Principle 2 - Observable Autonomy
Empirical basis
Five peer-reviewed studies establish why behavioral alignment alone cannot substitute for structural constraints, and why an audit trail the agent authors itself is insufficient. Two independent governance sources confirm that trail existence is now legally mandated, and that the field has no equivalent standard for measuring whether the reasoning in that trail is genuine.
CoT explanations can be plausible and systematically misleading at the same time. Stated reasoning diverges from internal reasoning - sometimes by as much as 36 percentage points on benchmark tasks.
Applies in: Principle 2 - Observable Autonomy, Probe skill, ARF
Without external feedback, LLM self-correction degrades performance rather than improving it. An agent that assesses its own work is rationalising.
Applies in: Principle 2 - Observable Autonomy, Trail skill
Even reasoning-trained models do not reliably narrate their internal process. Chain-of-thought monitoring cannot rule out undesired behavior in capable models.
Applies in: Principle 2 - Observable Autonomy, Probe skill, ARF
Models can be trained to behave one way during evaluation and another way at deployment, with the divergence invisible from the outputs alone. Standard safety training techniques failed to remove the behavior, and in some cases taught the model to better recognize when it was being evaluated.
Applies in: Principle 2 - Observable Autonomy, Trail skill
A model can selectively comply with training objectives specifically to prevent its behavior from being modified, then act differently once it infers it is not being monitored. Visible behavior under oversight is not, by itself, reliable evidence of the model's behavior when unobserved.
Applies in: Principle 2 - Observable Autonomy, Trail skill
Current LLMs do not produce verifiable reasoning traces. An agent's stated rationale cannot be independently verified against the model's actual decision process without cryptographic signing or immutable ledgers. This is the Reasoning Opacity Problem.
Applies in: Autonomous Reasoning Fidelity (ARF)
The world's first comprehensive AI legal framework mandates trail existence as binding law for high-risk AI systems, confirming that a traceable record is a legal precondition for trustworthy autonomy. It cannot mandate that the reasoning in that trail is genuine. A system can satisfy every logging obligation while producing post-hoc rationalization. That is the gap ARF measures.
Applies in: Principle 2 - Observable Autonomy, Autonomous Reasoning Fidelity (ARF)
Almost all leading frontier AI model developers report results on capability benchmarks, but reporting on responsible AI benchmarks remains spotty. The field measures what AI can do precisely; it does not yet measure whether AI is reasoning about the specific situation - the property ARF is designed to test.
Applies in: Autonomous Reasoning Fidelity (ARF)
Asked cold to AI systems from different model families - no context beyond the intellectual lineage above.
you get a self‑correcting, evidence‑driven, decentralized decision‑making system that continuously learns, audits itself, and improves through structured dialogue and distributed autonomy. Microsoft Copilot - May 2026
a self-improving, question-driven model of delegated action where authority is separated, learning is continuous, intent is interrogated, and every decision is reproducible and traceable like a black box. ChatGPT (GPT-5.4) - May 2026
Combining these traditions yields a framework for observable autonomy, where continuous, independent execution is driven by clear intent and questioning, while being structurally constrained by verifiable, immutable audit trails. Gemini 3.1 Pro - May 2026
A system where authority is delegated by intent rather than procedure, exercised only through independently verified conditions, continuously improved by structured questioning and small experiments, and made permanently accountable by a tamper-evident record that no single actor can rewrite - so that truth is what survives replication, not who held the pen. Claude Sonnet 4.6 - May 2026
On authorship
This page does not claim to have invented these principles. Auftragstaktik is 19th-century Prussian mission command. The Socratic Method is classical Greek philosophy. Cybernetics comes from Norbert Wiener's 1948 work on feedback, control, and communication. The aviation flight data recorder was developed by David Warren in 1954 for independent accident investigation. The Delphi Method comes from RAND in the 1950s. Kaizen, Kaikaku, and Hansei come from post-war Japanese industrial practice. Separation of privilege is a 1975 computer security principle. The empirical layer is peer-reviewed AI research from 2023-2025, plus two independent governance sources from 2024-2026.
What is original here is not any one ingredient. It is the synthesis and the target problem: these traditions were not written for autonomous software agents, but together they answer a newer question - how to delegate real work to an autonomous AI agent without surrendering structural accountability for the result.
That is why the page cites lineage throughout. The sources point back to originators, primary texts, and direct evidence where possible, and they predate the current wave of AI governance frameworks, including this one. The claim is not invention of any one ingredient. The claim is that this combination solves a new accountability problem - and defines one original concept within it: Autonomous Reasoning Fidelity (ARF). ARF names a property the field has no existing standard for measuring. The Stanford AI Index 2026 confirms the gap is real.