45% Threshold (P9) — If a single agent achieves greater than 45% of optimal performance, improve the agent before adding more agents. Derived from DeepMind’s multi-agent scaling research showing diminishing returns from parallelization when a single agent is already performing well.
Attention Budget (P2) — Context is a finite resource where every token competes for attention weight via n-squared self-attention relationships. Adding one token does not cost one token of attention — it creates new relationships with every existing token, making context cost quadratic.
BECAUSE Clause (P5) — The reason appended to every rule in persistent documentation that enables the model to generalize beyond the literal instruction. A rule with a BECAUSE clause becomes a living principle; without one, it becomes a dead rule.
Blueprint-Branch-Build (P4) — Three-phase disposable plan workflow: write the plan in one session, start a fresh session with a clean context, then implement against the plan. The context break between planning and building prevents sunk-cost spirals.
Cascade Pattern (P9) — Level 0–4 escalation model for agent team sizing. Level 0: single agent. Level 1: agent with tools. Level 2: worker plus reviewer. Level 3: small team of 3–4 agents. Level 4: multi-team with coordinator. Each level adds coordination overhead; escalate only when the previous level measurably fails.
Context Poisoning (P2) — Stale or contradictory context that actively misdirects the model, producing worse output than no context at all. A common result of kitchen-sink context loading combined with unmaintained documentation.
Dead Rule (P5) — A directive without explanation that covers only one literal case and cannot generalize. The model follows it when the exact situation matches, but has no basis to apply it to analogous situations. Opposite of a living principle.
Distribution Center (P5, P6) — The generic average output LLMs produce without specific constraints or expert vocabulary. Named for the tendency of neural networks to regress to the mean of their training distribution when given vague instructions.
FM-3.1 Rubber-Stamp Approval (P7, P8) — The most common quality failure in multi-agent systems, catalogued in the MAST taxonomy. A review agent approves work without meaningful scrutiny, producing the appearance of quality control with none of its benefits.
Hardening (P1) — Replacing a probabilistic LLM step with a deterministic, tested tool for any mechanical task that must behave identically every time. The first principle because it removes entire categories of failure from agentic workflows.
Kitchen Sink Context (P2) — Loading every reference document, plugin, and tool “just in case,” causing quadratic attention degradation. Named for the common pattern of cramming everything into context on the theory that more information is always better.
Living Principle (P5) — A rule with a BECAUSE clause that generalizes to novel situations the original author never anticipated. The model understands the intent behind the rule and can apply it by analogy. Opposite of a dead rule.
Lost in the Middle (P2, P10) — A 30%+ accuracy drop for information placed in the middle of the context window, demonstrated by Liu et al. (2024). Models attend strongly to the beginning and end of context but measurably underweight the middle.
MAST Taxonomy (P7) — A framework cataloguing 14 failure modes across three categories: communication (4 modes), coordination (5 modes), and quality (5 modes). Provides a checklist for anticipating and instrumenting multi-agent failure points.
Optimal Zone (P2) — 15–40% context window utilization, the sweet spot balancing between hallucination risk (too little context) and attention dilution (too much context). Below 15%, the model lacks sufficient grounding; above 40%, important details start drowning in noise.
PRISM (P6, P10) — Persona research findings demonstrating that brief identities of fewer than 50 tokens outperform elaborate backstories, and that flattery (“you are the world’s best”) actively degrades output quality. Grounds Forge’s agent identity design.
Progressive Disclosure (P2, P10) — Four-layer context loading strategy: always-loaded (project identity and critical rules), task-triggered (activated by workflow phase), on-demand (fetched when the agent needs it), and compressed (summaries of large references). The architecture behind jig’s selective loading.
Sunk-Cost Spiral (P4) — Continuing to patch failing code because of time already invested, instead of restarting from a revised plan. The primary failure mode that disposable blueprints prevent — when the plan is cheap, discarding it costs nothing.
U-Shaped Attention Curve (P2, P10) — High attention at context start and end, measurably low in the middle. Wu et al. (MIT, 2025) showed this is caused by causal masking and RoPE positional encoding — architectural features that cannot be patched away with prompting tricks.
Vocabulary Routing (P6, P10) — Expert terminology in prompts activates domain-specific knowledge clusters in the model’s embedding space (Ranjan et al., 2024). Saying “cyclomatic complexity” routes to software-engineering knowledge; saying “code quality” routes to the distribution center.