Agent Context Drift & Anchor Lock Lab

Anchor Locks 2/3 Locked

Locking Mechanism

Attentional Noise Rate 0.22 Simulates distracting clause noise & token degradation

Clause Constraints (Doc 9150)

REASONING STEP: 20/20 CLAMPED

Reasoning Telemetry CONVERGED

Drift Index @ Step 20

0.08

Fact Retention Score

94.2%

Hallucination Prevention

98.1%

Recommended Cadence

Every 3 chunks with weight decay threshold 0.45

Prompt Scaffolding (JSON)

Agent Context Drift & Anchor Lock Lab

How does the simulator model context drift across reasoning steps, and how do lock mechanisms alter the synthetic drift trajectory?

This laboratory simulates how an autonomous agent might accumulate attentional degradation across a 20-step document reasoning sequence. The underlying script models unanchored degradation as a linear growth formula driven by step count and an attentional noise slider. Applying anchor strategies modulates this curve by calculating a lock ratio from active clause weights, scaling down simulated drift and recomputing synthetic retention and hallucination prevention indices.

All curves, drift indices, fact retention percentages, and hallucination prevention ratings are generated by simplified programmatic heuristic equations in the client-side JavaScript. They do not represent live transformer attention measurements, real token-level perplexity, or benchmark results from an actual language model runtime.

Try a worked example

With default settings (step 20, noise rate 0.22, and constraints C1 and C2 active), switch the Locking Mechanism dropdown to 'Rolling Memory Window (No Anchor)'. The Drift Index changes from 0.08 to 0.89, Fact Retention Score drops from 94.2% to 41.3%, and Hallucination Prevention decreases from 98.1% to 45.0%, reflecting an unmitigated degradation model.

Mathematical Mechanics of the Trajectory Model

For each step s from 1 to 20, unanchored degradation is calculated in the client script as rawDegradation = min(1.0, s * 0.045 + noise_rate * (s / 10)). When constraints are toggled, a lock ratio is computed as the sum of locked constraint weights divided by total available constraint weight.

Selected locking strategies modify an anchor factor: 'strict_hard_lock' scales factor down with lock ratio, 'semantic_boundary_clamp' clamps the base drift near 0.08, 'decaying_soft_lock' introduces step-dependent decay, and 'rolling_window' directly mirrors 85% of raw degradation. The final anchored drift is then bounded between 0.02 and 1.0.

Derivation of Telemetry Metrics

When not using the default hardcoded baseline calibration, the telemetry panel evaluates fact retention as max(10, 100 - (drift * 60) - (noise_rate * 25)) and hallucination prevention as max(5, 100 - (drift * 80) * (1 - lock_ratio * 0.3)). These formulas serve as educational proxies to illustrate how constraint retention counters compounding noise.

Weighted anchor ratios and default fixture overrides

Your browser does not support video.

Read the explanation

The source stores three constraints weighted one, point eight and point six. The first two are locked by default, so locked weight is one point eight out of total two point four, a ratio of point seven five. Bars use one hundred sixty pixels per weight unit. These source weights represent authored importance assumptions; they are not independently measured attention or legal compliance. For the default semantic boundary strategy, weighted ratio point seven five and step twenty, the general expression would round drift to point one one six. But an exact branch overrides it to point zero eight whenever there are two locks, noise point two two and step twenty. Bars use three thousand pixels per drift unit. Unanchored degradation reaches its cap one at this step. The advertised benefit is therefore partly a calibrated fixture, not a model evaluation. Default retention and prevention are assigned ninety four point two and ninety eight point one whenever strategy, lock count and noise match the fixture. At drift point zero eight the ordinary retention expression would give eighty nine point seven, and the prevention expression about ninety five percent. Bars compare the retention labels only. This page simulates curves and stores settings locally; it does not call a language model, measure hallucinations, or establish actual regulatory correctness.