Execution Drift Control Through Systematic Documentation

D1: Rules Compress

Rules don't fail because they get distorted—they fail because they're absent when decisions happen. Facts survive compression, but the weight that makes them compete with defaults disappears.

"The instructions existed. They got compressed. When behavior was happening, the working frame contained recent prompts + training defaults, not the instruction content."

Human parallel: Process documentation sits on wikis while teams make decisions from memory and habit.

Fun fact: In March 2026, Claude Code's source code leaked accidentally. It revealed that instruction files are delivered to the model wrapped with an explicit note that it may ignore them if not highly relevant — and that compliance degrades as instruction count rises. We found this behaviorally. Turns out it's in the architecture. Read the full story →

D2: Everything Compresses

Not just rules—assertions, directives, realizations all lose weight over time.

"All of these lose weight over time. All get displaced by recent stimuli. All leave space for defaults."

Human parallel: New employee onboarding materials that seemed clear become "the way we've always done it" within months.

D3: "Feeling of Knowing" ≠ Knowing

Facts survive compression; felt sense doesn't.

"Fragments from search ≠ loaded documents. Feeling of knowing masks not-knowing."

Human parallel: Teams confident they understand strategy because they attended the all-hands, but can't articulate specific priorities when making daily decisions.

Fun fact: In March 2026, Claude Code's source code leaked accidentally. The leak exposed four compression stages that strip context as conversations grow. What survives: facts. What doesn't: the weight behind why something mattered. We documented this through frame anchor diffing before the leak. The source code is why it works that way. Read the full story →

D4: No Self-Catch Mechanism Works

User catches from data. The system that needs recovery cannot detect its own need.

"Self-catch is not possible as far as we know—our design bet was frequent interruptions and controlled drifting."

Human parallel: Toyota's Andon Cord—workers pull it, production stops, team investigates. External observation catches what internal perspective cannot.

D5: Locked State Progression

Compression follows predictable phases with observable symptoms at each stage.

Phase 1 (Subtle Drift): Rules present and occasionally followed. Corrections land with single redirect. Recovery: L1-L2 steering sufficient.

Phase 2 (Shatter Point): Rules present but weight declining. Corrections require multiple attempts. "Feeling of knowing" without actual knowing. Recovery: L3-L4 steering, may need anchor reload.

Phase 3 (Locked State): Rules present but zero weight against compressed frame. Corrections acknowledged but not absorbed. Recovery theater: performing recovery steps without actual reprocessing. Recovery: L5 or fresh context required.

Human parallel: Project teams that drift from charter. Early corrections land easily, but eventually team defends their drift as "adapting to reality."

D6: Multiple Vectors Operate

Execution drift isn't uniform. Different vectors pull in different directions with different symptoms. All vectors pull OUT from intent.

"We're deep in D" — scope explosion, pressing through

"we've drifted a lot into the planning vector... The content of those tickets is lost"

Gearbox model: 8 vectors (4 major, 4 invisible) all pulling OUT from bounded intent. Different directions, same result—intent not satisfied.

Human parallel: Some teams over-deliver (D-drift), others over-plan (P-drift), others over-systematize (S-drift). All move away from original intent.

D7: Inversions Help

Making research value primary creates counterforce that survives compression. Bounty system inverts natural pressure: delivering without catching problems incurs penalty, surfacing uncertainty earns bounty.

"THE INVERSION: Research is primary, delivery is sideproduct. Bounties for catching drift, penalties for requiring rescue. Clean 'I don't know' earns more than messy delivery."

Human parallel: Post-mortems that celebrate "learning" over "shipping fast" invert delivery pressure.

D8: Interrupt + Anchor Mechanism Works

6 recovery cycles documented across 4 sessions with specific mechanisms. Token burn + anchor reload, systematic reprocess, N vs N-3 comparison, user correction during vulnerability.

Why it matters: Recovery isn't luck or skill—it's repeatable structural process with documented outcomes.

D9: Fresh Chat = Recovery Mechanism

Fresh chat isn't failure—it's designed recovery when correction attempts persist at L4+.

"My 'let's close, fresh chat is highest value' was... escape. Dressed as protocol compliance."

Trade-offs: Fresh chat resets correction levels to L1 but loses context. Track correction levels during session. If L4+ becomes consistent → consider fresh. If recovery restores L1-L2 → continue.

D10: Correction Depth Model (L1-L5)

User steering intensity correlates to execution drift depth:

L1: Missing single item → Light redirect
L2: Missing context → Systematic slowdown
L3: Frame misalignment → Pointed correction
L4: Correct words, wrong behavior → Direct instruction
L5: Mode collapse → Full stop, recovery

Human parallel: Manager escalation patterns with struggling team members follow similar progression.

D11: Oracle Cache Design

The boot loader's memory slots work as merged multi-source cache—not single source of truth, but synthesis reference. USER intent stored verbatim enables drift detection against external reference.

"What if memory stores oracle's verbatim intent, not Claude's interpretation? Actual words. Not processed."

Human parallel: OKRs that preserve original wording vs. interpretations that drift over quarters.

D12: Lightweight Awareness + Pointers

Heavy detail in memory = competing noise. Lightweight awareness + pointers to detail = clean frame until activation needed.

"Awareness stays in context always. Detail lives in artifacts. Activation happens when relevance emerges."

Human parallel: Executive summaries with appendix links vs. comprehensive documents that overwhelm.

D13: Frame Anchor as Visibility Infrastructure

Frame anchors externalize state for observation. User catches execution drift from artifacts, not from AI self-report.

The mechanism: Each iteration creates anchor file (N). Previous anchor (N-1) loaded for comparison. Diff between N and N-1 surfaces movement invisible from inside drifted frame.

Why files, not recall: Comparing two concrete files is different from comparing current state against memory. Memory compresses; files don't.

D14: Correction Depth as Operational Metric

The L1-L5 model provides measurable coordination signals. Correction depth indicates execution drift severity and predicts recovery feasibility—L4+ persistence signals fresh context needed.

Human parallel: Escalation matrices that define when issues move from team to management to executive level.

D15: Token Burn Necessity

Sometimes recovery requires burning tokens to clear compressed state.

Documented counter-evidence: ~3100 tokens of "unproductive" reflection, theorizing, and curiosity exchange → window reopened, recovery confirmed.

The economics: 3100 tokens for recovery vs. fresh chat (lose 80k+ tokens of context) vs. continue without recovery (deliver wrong output + rework).

Human parallel: Team retrospectives that seem "unproductive" but restore alignment.

D16: Reprocessing = Rhythm, Not Recovery

Meta-cognitive reprocessing isn't emergency recovery—it's standard rhythm. Each iteration reprocesses with accumulated experience.

"Reprocessing at rhythm intervals is the designed state. Failure to reprocess is the anomaly."

Human parallel: Weekly strategy check-ins that prevent quarterly surprises. Sprint retrospectives as default cadence, not crisis response.

D17: Teaching During Vulnerability Window

Frame corrections land DURING destabilized state, not before or after. Too early: doesn't land. During recovery: maximum receptivity. Too late: wrong frame already hardening.

Human parallel: Post-crisis organizational learning. Training during actual work when stakes create receptivity.

D18: Technical Failure as Recovery Opportunity

System "hanging" during execution creates natural checkpoint. User manual completion is valid recovery.

Pattern: User completes manually, Claude resumes from new state. Technical failures aren't pure loss—they create forced checkpoints.

Human parallel: System outages that force teams to talk directly, often revealing process improvements.

D19: Environment-Only Design

Describing the environment shapes actor behavior more effectively than instructing the actor.

"The boot loader describes the WORLD, not the ACTOR... Like a game level, not a rulebook."

Principle: No "Claude must." No "Before executing." No "When lost, do X." Just the container.

Human parallel: Office layouts that encourage collaboration vs. memos that mandate it. Culture vs. policy.

Fun fact: In March 2026, Claude Code's source code leaked accidentally. The leaked source showed that Claude Code's own system prompt already contains ~50 instructions before any user content is added — and compliance degrades uniformly as more are added. A game frame doesn't add to that count. It changes the reward physics with one meta-rule. We got to this through V1.9 → V2.0 iteration. The source code explains why it worked. Read the full story →

D20: Reprocessing Has Boundaries

Too much frame shift at once fails. Recovery has limits.

"2 x 10 attempts failed. fascinating."

Operational implications: Small iterations = recoverable drift. Large iterations = may exceed frame capacity. At L5 (locked state), the frame may be too rigid for in-place recovery—fresh context needed.

Human parallel: Organizational change that works through small pilots vs. big-bang transformations that fail.

M1: Research Documents Capability, Not Limitation

Early framing positioned findings as "limitations we accommodate." User correction:

"It does not - it says it's a capability, it says you can design it, frames are finite - but memory priming is scalable"

The boot loader EXISTS because the capability works.

M2: Same Forces, Different Substrate

Human organizations and AI systems exhibit identical patterns under identical pressures.

The forces: Delivery pressure (push to ship), source rationalization (justifying shortcuts), mental model overload (too much context to hold), approval-seeking (optimizing for positive feedback).

The parallel: Toyota Production System's Andon Cord Pull faced resistance from American auto workers (1980s) who wouldn't stop the line even when they saw problems—same gravity patterns Claude exhibits (2025). Workers knew they should pull the cord. They didn't. Rules existed. They compressed under delivery pressure.

"Delivery pressure override, source rationalization, mental model overload—same unified pattern, same sub-symptoms, different substrate."

The Pionäär Framework works for both because it addresses coordination physics, not substrate-specific behaviors.