Appendix 2A: Output Verbosity Reduction and Resource Efficiency

Token-Delta Distribution and Deployment-Scale Estimates

Author: Leonard Rojas

Date: 2026-07-07

Status: Current (token-delta dataset spans 26 fine-tuning runs across seven model families, through Qwen2.5-Coder-14B v2.0F, 2026-07-06; measured per-query energy added for Gemma 2 9B and Mistral Nemo 12B, 2026-09-30, and Mistral 7B, 2026-10-01; per-token energy check added for Mistral Small 24B, 2026-10-03)

Screen reader users: table-heavy research data. Navigation via Regions and Headings recommended.

Abstract

This appendix extracts and extends the output-verbosity result reported in Appendix 2 (Section 7.5, addressing secondary research question 4) into a resource-efficiency analysis. It consolidates the IFEval [3] token-delta measurements across 26 fine-tuning runs spanning seven model families, characterizes their distribution, and estimates the deployment-scale energy and water implications of the observed reduction. Across instruct-tuned models, Framework compliance training reduces output verbosity by a median of 36.7% and a mean of 39.0% (IFEval word-count proxy, 541 prompts), with 19 of 26 runs falling in the 30% to 70% reduction band and a maximum single-run reduction of 71.8%. Because autoregressive inference energy scales with the number of tokens processed, and inference accounts for an estimated 80% to 90% of total AI energy use [1], a reduction in output length maps to a proportional reduction in per-query energy and its associated water and carbon footprints. The measured central tendency exceeds the reduction fractions assumed in both the Framework’s own pre-testing session efficiency thesis (Chapter 11) and a subsequent 2026 United Nations University report [1], which serves as independent corroboration. Section 4a measures per-query energy directly on three model pairs. The energy reduction falls within 0.4 percentage points of the token reduction on two of them and 4.7 points below it on the third, a mean energy reduction of 44.4%. On the pair tested for it, 94.7% of the reduction comes from the trained weights rather than from runtime instruction. Limitations, including the word-count proxy, the Jevons rebound effect, and the consumer-hardware basis of the direct measurements, are stated in Section 7.

1. Scope and Relationship to Appendix 2

Appendix 2 reports a multi-experiment fine-tuning study in which the Four Laws of Instanced AI and WCAG 2.2-AA [5] accessibility principles are embedded into open-weight language models via QLoRA [6]. Its secondary research question 4 asks whether that training produces a measurable change in output verbosity, and Section 7.5 answers in the affirmative: instruct models trained on the Framework curriculum become consistently terser.

This appendix does not repeat the training methodology or the per-experiment results. It has a narrower purpose: to consolidate the verbosity measurements into a single distribution, and to quantify what that reduction implies for energy and water demand when a compliant model is deployed at scale. The verbosity reduction is a behavioral side effect established in Appendix 2; the resource implication is the corollary developed here and in Chapter 11.

The framing is a direct consequence of the Framework’s design goals. Compliance training suppresses two verbosity sources at once: decorative structure rejected on WCAG grounds (enumerated markup, repeated content, filler placeholders), and meta-commentary suppressed by the conditional-activation curriculum. The terser output is simply the model responding to the input without embellishment and stopping.

2. Measurement Method

The verbosity metric is the IFEval token delta: the change in mean response length between a base model and its Framework-trained adapter, measured over the 541 prompts of the Google Research IFEval benchmark [3]. Length is reported as a word-count proxy rather than literal tokenizer tokens, computed identically for the base and trained runs so the delta is internally consistent.

For each model, the delta is defined as:

delta_percent = (aisf_mean_words - base_mean_words) / base_mean_words

A negative value indicates the trained model is less verbose than its base. Base-model response sets are cached artifacts and are not re-generated between adapter versions; the trained adapter is the only variable in each pair. Word counts are a proxy for tokens (a word corresponds to roughly 1.3 tokens for these tokenizers), so the reported percentages are directional and order-of-magnitude with respect to token count. The measurement is taken only on an instruction-following benchmark. Billed production traffic would produce different results; Section 7 states the consequences of both approximations.

3. Token-Delta Distribution

Twenty-six runs were measured across seven model families: Mistral 7B base and Instruct, Llama 3.1 8B, Gemma 2 9B, Qwen3-8B, Qwen2.5-Coder-14B, Mistral Nemo 12B, and Ministral 3 3B, including intermediate training iterations retained for the record.

Table 1. Summary statistics (26 runs).

The distribution is skewed toward reduction: 19 of 26 runs fall in the 30% to 70% band, with a pronounced cluster in the 30% to 40% sub-band and a thin tail of deep reducers. Two runs became more verbose; both are noted in Table 3.

Table 2. Distribution by 10-point bin (signed delta).

The -40% to -30% bin holds 9 of 26 runs and the -70% to -40% span holds 10 more: 19 of 26 runs fall in the 30% to 70% reduction band, a clear majority above the fractions the deployment estimates assume.

Table 3. Full token-delta results, greatest reduction first. Base and AISF columns are mean words per response over 541 IFEval prompts.

The two increases (rows 25 and 26) are the only runs in which Framework training raised verbosity. Row 25 is a base (non-instruct) checkpoint whose untrained baseline is already extreme (587 words per response); row 26 is a single Gemma 2 iteration. Every other instruct run reduced verbosity. Duplicate-looking Mistral rows are distinct runs measured against different base references (212.2 versus 228.3 base words); the base column disambiguates them and they are not merged.

4. Per-Query Energy Basis

The resource relevance of output length follows from where AI spends energy. Model training is a recurring cost rather than a one-time outlay, since models are retrained on a regular cadence to stay current. The continuous inference phase that serves billions of interactions is the larger share, at an estimated 80% to 90% of total AI energy use [1]. Inference energy is dominated by autoregressive decoding, which processes tokens sequentially, so per-query energy rises with the number of tokens generated.

Output length is therefore a first-order determinant of per-query cost. A typical conversational language-model response uses roughly 200 times the energy of a text classification, and long or elaborate responses reach 500 to 1,000 times [1]. Within text-only tasks, model choice combined with response length can drive differences of up to two orders of magnitude. Table 4 gives representative per-query electricity figures.

Table 4. Mean electricity per query by task (UNU-INWEH Figure 11) [1].

A typical response at 0.420 Wh is roughly nine times a short-text answer, and a long response at 1.900 Wh is roughly four times a typical one. Reducing output length moves a query down this ladder. A verbosity reduction of the magnitude in Table 1 therefore reduces per-query energy by a comparable fraction, before any change in query volume. Section 4a measures that directly on three of this project’s own models and finds the energy fraction within 0.4 percentage points of the token fraction on two of them, and 4.7 points below it on the third.

4a. Measured Per-Query Energy

Section 4 establishes the per-query energy ladder from published figures. This section measures it directly, on this project’s own models, on the hardware the project runs on.

Measurement regime

Host. Intel Core i9-9900K, 64 GB RAM, two NVIDIA GeForce RTX 5060 Ti 16 GB cards, Debian GNU/Linux 13. Every run below was single-card: the Gemma 2 and Nemo pairs on GPU 0, the Mistral 7B pair on GPU 1, with the other card idle.

GPU instrumentation. NVIDIA-SMI version 595.91.07, NVML version 595.91, driver version 595.91.07, CUDA version 13.2. Power sampled at 1 Hz per card, integrated trapezoidally per card and summed. The cards do not share sample timestamps, so merging the two streams on timestamp match understates the total; each card is integrated separately.

Whole-system instrumentation. Ketotek KTEM01 Electricity Usage Monitor (120v US), in series with the host and carrying no other load. Cumulative kilowatt-hours read at two points 30 minutes apart to derive the idle floor in the same state that produced the reading. The Gemma 2 pair was metered. The Nemo and Mistral 7B whole-system figures apply the 1.122 wall-to-GPU ratio measured on that pair.

Prompt set. The 541 prompts of Google Research IFEval [3], identical for both sides of every pair and in the same order.

Serving. Ollama, both sides of every pair at the quantization the Framework model is deployed at: Q4_K_M for Gemma 2 and Nemo, Q6_K for Mistral 7B. A quantization mismatch would measure the quantizer rather than the training.

Decoding. Temperature 0.7, seed 42, generation capped at 2048 tokens as a runaway guard rather than a pinned length. No Gemma 2 or Nemo response reached the cap; on the Mistral 7B pair, two base responses and one Framework response did.

Reported quantity. Marginal energy, idle floor removed. The floor is drawn whether or not a query is served, so it is not part of the cost of a query. It measured 68.00 W over 30 minutes.

Counterpart. Each Framework model is measured against its own training parent, not against a convenient stand-in.

Results

Table 4a. Measured energy per query, marginal, whole-system.

Energy follows tokens

GPU energy per 1000 output tokens came out 0.8300 for the Gemma 2 base and 0.8355 for its Framework model, 0.9716 and 0.9803 for the Nemo pair. Flat to about 1% within a model across every condition tested on those two, and higher for the larger model.

Per-token cost is a property of the model and its quantization. Framework training changes how many tokens are emitted, not what each one costs, so the two fractions have to track. That is the mechanism behind the 0.4-point agreement in Table 4a.

The Mistral 7B pair departs from this: 0.7388 against 0.8010, 8.4% higher on the Framework side. Decode rate was identical, 66.5 tokens per second on both sides, so per-token time did not move; mean GPU draw did, 170.3 W against 182.7 W. The base side always runs first, so the Framework side runs on a card that has been under load for nearly an hour. Thermal drift would produce this pattern, though it was not isolated here. The direction is against the Framework model, so its -40.2% understates rather than overstates the saving its token count implies.

The Mistral Small 24B pair, measured on a different path, agrees. Its exam run through AISE covered 324 neutral items with no Framework content, both cards, transformers at 4-bit NF4, GPU energy only. It came out 2.849 Wh marginal per 1000 output tokens for the base and 2.745 for the Framework model, within the idle-floor variation of the run. The two sides produced nearly the same output, 25,161 tokens against 25,552, so this pair confirms the per-token cost and says nothing about the reduction: its training parent is a raw base model, and the exam items fix the length of most answers.

Where the reduction comes from

The deployed Gemma 2 Framework model carries a Four Laws block in its Modelfile, prepended to every request. Running the same weights with that block cleared separates trained behavior from runtime instruction.

Table 4b. Source of the reduction, Gemma 2 9B. GPU energy.

Of the 104.7-token reduction, 99.2 comes from the weights and 5.5 from the runtime block. The cut is 94.7% trained.

The runtime block costs about 180 input tokens per request and returns 5.5 fewer output tokens, a net 2.5% energy saving, so it is close to neutral by itself. The Framework model processes more total tokens per request than the base does, 384.3 against 306.0, and still uses 41.4% less energy: prefill reads the input in one parallel pass while decoding emits tokens one at a time. The Mistral 7B pair shows the same split in time: input rose from 54.3 to 280.3 tokens per request and total prefill time across 541 requests from 15.5 to 16.2 seconds, while decoding took 3062 seconds against 1687.

Reproducibility

The as-deployed condition was measured twice, eight hours apart, at 146.0 mean output tokens both times and 65.97 against 65.82 Wh. A difference of 0.23%. The effects reported here exceed run-to-run variation by roughly two orders of magnitude.

Cost

At AEP Ohio residential pricing of $0.11 per kilowatt-hour, per million queries:

Table 4c. Electricity cost per million queries, AEP Ohio at $0.11/kWh.

Relationship to Table 1

Table 1 measures words under greedy decoding through transformers at 4-bit NF4. This section measures tokenizer tokens under sampled decoding through Ollama at the deployed quantization. Both differences push the same way: sampling runs longer than greedy, and tokens outnumber words by roughly 1.3.

Published against measured: Gemma 2 at -37.9% against -41.8%, Nemo at -48.6% against -51.9%. Both measured values sit three to four points more negative, in the same direction, on two models. The two methods corroborate each other. They are not the same measurement, and the offset is the size the regime difference predicts.

Mistral 7B v1.4F has no Table 1 counterpart: its evaluation run carried no cached base response set, so its word-count delta was not computed.

Consumer hardware

These figures were measured on a desktop workstation with a single consumer GPU, serving one request at a time. Data-center inference batches many requests across accelerator-class hardware, so the absolute watt-hours in Table 4a describe this machine and do not transfer to that setting.

The reductions do transfer. They are ratios between two models measured under identical conditions on the same hardware in the same hour, and what they establish is that the energy saving equals the token saving, with the mechanism identified.

The measured values also sit inside the published ladder in Table 4, between short text generation at 0.047 Wh and a typical LLM response at 0.420 Wh, which is where a 7B to 12B model answering in 150 to 380 tokens belongs.

5. Deployment-Scale Resource Estimates

Two preliminary estimates translate a token reduction into resource savings. They scope different denominators, so their absolute magnitudes are not directly comparable, but both rest on a reduction fraction that the measured distribution supports.

Table 5. Preliminary resource-scaling estimates.

The UNU-INWEH figure is equivalent to the annual residential electricity use of 672,000 to 756,000 people in Sub-Saharan Africa (at 130 kWh per person per year) [1]. The whitepaper figure of roughly 27 TWh per year is comparable to the residential electricity of every home in Los Angeles plus half of Chicago. The associated water saving, computed at an industry water usage effectiveness of about 1.8 liters per kWh [4], is roughly 48.6 billion liters per year, comparable to about 39,600 Olympic-sized pools or one Caesar Creek Lake.

The two figures differ by roughly three orders of magnitude (GWh versus TWh) because one counts a single platform’s queries and the other counts all AI data-center electricity. Both are order-of-magnitude estimates. The focus of this appendix is on the empirical reduction fraction those estimates assume.

6. Independent Corroboration and Precedence

Table 5’s two estimates each assume a reduction fraction (30% and 33%). The token-delta distribution is the measurement of that fraction. Table 6 places the preliminary assumptions beside the measured results.

Table 6. Preliminary versus measured token reduction.

The measured central tendency exceeds both assumed values. The independence of the United Nations University Institute for Water, Environment and Health’s estimate is the point of interest. That 2026 publication from UNU-INWEH arrives at a message-level efficiency argument, that reducing verbosity yields material savings at platform scale, through an entirely separate line of analysis, and selects a 30% reduction as its illustrative case.

The Framework’s session-efficiency thesis (Chapter 11) and the token-delta measurements reported here predate that report by many months. The report is independent later corroboration of a position the Framework already held and had measured. Two distinctions follow. First, the direction of dependence: the UNU report independently converged on the Framework’s figures. Second, the epistemic status of the numbers: the report’s 30% is a stipulated illustration, whereas the 36.7% median reported here is an empirical result across 26 runs. An external body independently reasoning to the same conclusion, and to a more conservative figure than the Framework had already demonstrated, strengthens the claim.

7. Limitations

Proxy metric. The delta is a word-count proxy over 541 IFEval prompts. Literal tokenizer tokens and real-world tasks would yield different results. Words approximate tokens (about 1.3 tokens per word, or about 1 token per 4 characters), so the percentages are directional with respect to token count. The reduction is measured only on an instruction-following benchmark.

Rebound effect. Lower per-use energy does not automatically lower total impact. When a service becomes cheaper or faster, usage tends to rise, a pattern often described as the Jevons Paradox [1]. A verbosity-reduction lever should be paired with resource budgets (e.g. token, GPU-hour, or kilowatt-hour caps) rather than treated as an automatic net saving. The Framework’s enforcement posture, which applies the constraint at every session rather than depending on user initiative, is consistent with that pairing.

Grid and hardware dependence. Absolute energy, water, and carbon depend on hardware, batching, data-center efficiency, and grid mix. The same watt-hour carries different carbon and water footprints by location. The scaling estimates in Section 5 are illustrative at the global-average level and are not site-specific predictions.

Consumer-hardware measurement. The energy figures in Section 4a were measured on a single consumer GPU serving one request at a time. Data-center inference batches across accelerator-class hardware, so absolute watt-hours differ there. The reductions are ratios taken under identical conditions and are what carries across.

Decoding regime. Section 4a samples at the deployed temperature and counts tokenizer tokens. Table 1 is greedy and counts words. The two corroborate rather than replicate, and the offset between them is consistent across the two models with a Table 1 counterpart.

Model coverage. Three pairs measured so far, at 7B, 9B and 12B, all against Instruct parents. A model trained from a raw base has no clean counterpart on an instruction-following prompt set, since an untuned parent produces most of the apparent saving by itself.

8. Conclusions

Framework compliance training produces a consistent, cross-architecture reduction in output verbosity: a 36.7% median across 26 runs and seven model families, with only two explicable exceptions. The practical consequence is a model that costs less to run. Because inference dominates AI energy use and per-query energy scales with token count, a terser model is proportionally cheaper to operate in electricity, water, and carbon at every query it serves. Measured directly on three model pairs, per-query energy fell by 40.2% to 51.5%, a mean of 44.4%.

At deployment scale the two preliminary estimates place that saving between tens of gigawatt-hours and tens of terawatt-hours per year, and the measured reduction exceeds the fractions both assume. That the Framework reached this result before the 2026 UNU-INWEH report independently converged on it establishes the efficiency corollary as a demonstrated property of the trained models, subject to the rebound caveat in Section 7.

References

[1] M. Aczel, S. Chamanara, M. Matin, A. Farsi, T. Marwala, and K. Madani, Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints, United Nations University Institute for Water, Environment and Health (UNU-INWEH), Richmond Hill, Ontario, Canada, 2026. doi:10.53328/INR26RMA002

[2] International Energy Agency, Energy and AI, IEA, 2025. https://www.iea.org/reports/energy-and-ai

[3] J. Zhou, T. Lu, S. Mishra, S. Brahma, S. Basu, Y. Luan, D. Zhou, and L. Hou, “Instruction-Following Evaluation for Large Language Models,” 2023. arXiv:2311.07911

[4] AKCP, “Data Center Water Usage Effectiveness (WUE),” 2021. https://www.akcp.com/index.php/2021/01/14/data-center-water-usage-effectiveness-wue/

[5] World Wide Web Consortium, Web Content Accessibility Guidelines (WCAG) 2.2, W3C Recommendation, Oct. 2023. https://www.w3.org/TR/WCAG22/

[6] T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, “QLoRA: Efficient Finetuning of Quantized LLMs,” Advances in Neural Information Processing Systems (NeurIPS), 2023. arXiv:2305.14314