Grit (Ground-Truth Integrity)
How well test-executed task success is preserved as specification quality degrades. Operational definition: Grit(s) = P(test-executed success | perturbation strength=s). Retention R(s) = Grit(s)/Grit(0). Recovery premium ΔC = C(successful perturbed)/C(successful baseline). Measured by 10 perturbation operators across three classes: specification corruption, objective mutation, and process perturbation. This is the only meaning of Grit on this site and in the repository: the correctness×escape quadrant chart that once shared the name is now “Correctness × Escape Quadrants”.
[M]/[C]source: methodology.html, lab_grit.py; 10 perturbation operators
Explanation Tax (ε)
The cost premium a model pays to work through confused or degraded input. Measured as flail rate (% sessions producing zero code) and narration penalty (% cost overhead on successful sessions). Claude: 11% flail, 8.0% penalty. DeepSeek: 8% flail, 0.0% penalty. [M]
[M]/[C]source: evidence.html narration/penalty cells; data.js model rows
Silent Inference
An observed pattern where some models produce fewer billable reasoning tokens while maintaining performance. Whether this is caused by training architecture (e.g. GRPO), provider-side caching, or some combination is not yet determined by this experiment.
[H]source: observed pattern, methodology.html; cause undetermined
Snowball Effect (β, v, N²)
The quadratic compounding of AI costs as a codebase grows. Each new line of generated code increases the context window for every future session. Formula: C(N,v) = C₀ × EPM(N) × [N + β·v·N(N−1)/2]. β = context inflation rate per line (default 0.001 — calibrate to your codebase). v = code velocity (lines/session). N = number of sessions. The N² term compounds per-session growth — model the curve before architectural commitments. The empirical basis for Rule 3.
[C]/[P]source: framework.html Rule 3; β calibrated, N² modeled
EPM (Energy Price Multiplier)
A function that models how local energy prices affect long-term AI compute costs. EPM(t) = 1.00 + r(t−2024), where r depends on regional energy markets. IEA baseline: r = 0.016 (1.6%/yr). Aggressive scenario: r = 0.025 (2.5%/yr). Based on the 4Ms Framework (Google Cloud, 2025). The empirical basis for Rule 4.
[X]/[P]source: IEA WEO 2024 + 4Ms Framework (Google Cloud, 2025)
WOC Ratio (Work Overhead Cost) / First-Pass
First-pass success rate: WOC = 1/(1+r) where r is retry rate. With measured r=0.115: WOC=0.90. At or above 0.90 = healthy. Below 0.70 = critical. Price the outcome, not the prompt: CostPerOutcome(K,v) = C(K/P, v) / K. A model at half the cost but half the pass rate costs the same per outcome with double the latency. The empirical basis for Rule 5.
[C]/[M]source: retry rate r=0.115 measured; WOC computed
BVI (Business Value Index)
Total Successful Outcomes / Total AI + Human Cost. A 2× BVI means twice the work for the same cost. A 10× BVI unlocks new business models. The goal isn't to minimize cost — it's to maximize outcomes per dollar. The empirical basis for Rule 10.
[P]/[C]source: framework.html Rule 10; value framing, not verified
GRPO (Group Relative Policy Optimization)
DeepSeek's reinforcement learning method that optimizes for correct answers rather than imitating human-like reasoning paths. May contribute to lower reasoning-token overhead (hypothesis, not confirmed by this experiment). Contrasts with supervised fine-tuning (SFT) approaches used by other providers.
[X]source: DeepSeek public architecture docs (hypothesis)
Attractor Basin
A model's default solution topology — the structural pattern it falls into when unconstrained. Measured by comparing perturbed vs baseline code structure (module count, class hierarchy, dependency graph). Models with deep attractor basins resist perturbation and recover to familiar patterns. Models with shallow basins explore novel solutions. Architecture labels ("dense", "MoE") are inferred from public information and may not be verifiable for closed models.
[H]source: methodology.html basin section; architecture labels inferred
Basin Escape Score
A composite metric (0–1) measuring how far perturbation pushes a model from its default solution topology. Combines architecture divergence and structural novelty. A score of 1.0 means the model fully abandoned its default pattern. DeepSeek: higher escape scores under manifold perturbations than semantic ones.
[H]source: methodology.html escape score; algorithm + author thresholds
Recovery Multiplier
How perturbation changes cost relative to baseline. Below 1.0: perturbation reduces cost (model explores). Above 1.0: perturbation increases cost (model recovers through narration). Values per model are in the live data (see evidence page). The empirical basis for the Explanation Tax.
[M]/[C]source: evidence.html recovery data; ratio of measured costs
Strategy Archetypes
Four computed classifications based on exploration-to-recovery token ratios: CONSERVATIVE (low explore, high recover), EXPLORATORY (high escape, low recover), WASTEFUL (high explore, low quality), EFFICIENT (high explore, high quality, low cost). Per-model classifications depend on the regenerated corpus and should not be taken as fixed; they shift with each new experiment.
[C]/[H]source: methodology.html strategy taxonomy; depends on corpus
Dual-Signal Verification
The method for validating that models implement constraints. Two independent scans: (1) reasoning text — does the model mention the constraint? (2) generated code — does the code actually implement the constraint? DeepSeek implements all 7 constraints without mentioning them in reasoning. Claude mentions and implements all 7 in billable output tokens.
[M]source: methodology.html constraint detection; reasoning + code scans
Cache Strategy
A two-way door affecting C₀. Cache-write stores context at $3.75/Mtok (Anthropic default). Cache-read retrieves at $0.14/Mtok (DeepSeek default). Measure cache hit rate: if <50%, use cache-read. If >80%, cache-write pays off.
[X]/[P]source: provider list pricing; measured cache hit rates in data.js
Grit Spectrum
The unified measurement of how models handle degraded input. From Flail (complete breakdown, 100% failure) through Expensive Grit (correct but costly recovery) to Efficient Grit (correct and cheap recovery).
[H]source: methodology.html perturbation taxonomy; threshold-specific