DECLASSIFIED BY SEAN · ENERGY TRANSPARENCY · FULL MATH SHOWN · PRIMARY SOURCES ONLY
Full Transparency · Every Assumption Documented · Every Number Challengeable
How Much Electricity Did This Cost?
Building DECLASSIFIED + The Holy Bible of Donald J. Trump · v1–v18 · May 2026
This site and its companion manuscript were built using AI assistance across two extended sessions producing 18 versioned deployments, 29 HTML files, 57 glossary terms, 60 fact-checked claims, 15 cheat codes, and 78,000 words of annotated manuscript. We applied primary-source methodology to the energy question the same way we apply it to everything else: show the math, document the assumptions, and acknowledge the uncertainty.
Best estimate: 40–50 kWh total. Plausible range: 15–90 kWh. The math is below. Adjust the assumptions yourself using the sensitivity calculator.
The Math — Five Components
Every message sent to Claude runs inference on large GPU clusters. This is the dominant energy cost by an order of magnitude. Everything else rounds to noise by comparison.
Assumption A1 — Session Volume
2 extended sessions. Approximately 180 total turns (back-and-forth exchanges).
Many turns chained multiple tool calls: bash executions, file reads/writes,
web searches, file creations. Conservative estimate.
Assumption A2 — Tokens Per Turn
Average input context per turn: ~60,000 tokens (growing as conversation
accumulated across the session). Average output: ~3,000 tokens per turn.
Input-heavy because the full conversation history is re-sent each turn.
Basis: typical long-session context growth pattern for extended coding sessions
Input tokens = 180 turns × 60,000 tokens/turn = 10,800,000 tokens
Output tokens = 180 turns × 3,000 tokens/turn = 540,000 tokens
─────────────────────────────────────────────────────────────────
Total tokens = 11,340,000 tokens ≈ 11.3 million tokens
Assumption A3 — Energy per 1,000 Tokens
0.003 kWh per 1,000 tokens (inference only, before data center overhead).
Basis: Luccioni et al. (2023) "Power Hungry Processing"; IEA (2024) "Electricity 2024";
published benchmarks for GPT-4 class models. Range: 0.001–0.008 kWh/1k tokens
depending on model size, hardware generation, and utilization.
Base energy = 11,340 (thousands of tokens) × 0.003 kWh/1k tokens
= 34.0 kWh
Assumption A4 — Data Center PUE (Power Usage Effectiveness)
PUE = 1.30 applied as overhead multiplier.
Anthropic uses AWS and Google Cloud. AWS reports global PUE of ~1.2;
Google Cloud ~1.1. Using 1.30 adds a conservative margin.
Basis: AWS 2023 Sustainability Report; Google 2023 Environmental Report
Total server energy = 34.0 kWh × 1.30 (PUE)
= 44.2 kWh
STEP 1 RESULT: ~44 kWh (range: 25–65 kWh)
The range is wide because the kWh/token figure is the most uncertain variable. If Anthropic runs on H100s at high utilization, the number drops. If on older hardware at low utilization, it rises. The sensitivity calculator below lets you explore this.
Assumption B1 — Device Mix and Duration
Estimated ~6 hours active engagement across sessions:
~2 hours mobile (5W average), ~4 hours laptop (45W average).
Includes screen-on time, charging cycles, background sync.
Mobile = 5W × 2 hours = 10 Wh = 0.010 kWh
Laptop = 45W × 4 hours = 180 Wh = 0.180 kWh
Charging losses (~20%) = 0.038 kWh
─────────────────────────────────────────────
Device total = 0.228 kWh ≈ 0.23 kWh
STEP 2 RESULT: ~0.23 kWh (range: 0.1–0.5 kWh)
Assumption C1 — Build Count and Duration
18 confirmed deployments (v1–v18). Each build: ~45 seconds on a
containerized build server drawing approximately 200W peak.
Build time = 18 builds × 45 seconds = 810 seconds = 0.225 hours
Build power = 200W (peak, containerized server)
Build energy = 200W × 0.225 hours = 45 Wh = 0.045 kWh
CDN serving (your test page loads, ~50 requests) ≈ 0.005 kWh
─────────────────────────────────────────────────────────────
Netlify total = ~0.050 kWh
STEP 3 RESULT: ~0.05 kWh (range: 0.02–0.10 kWh)
Static site CDN serving is among the most energy-efficient compute operations that exists. This number is genuinely negligible.
Assumption D1 — Search and Fetch Volume
~50 web searches during research passes (Kessler audit, expert recommendations,
ClaimReview documentation, Netlify API verification).
~20 full page fetches (HSGAC reports, court documents, Wikimedia Commons).
Google's published estimate: ~0.0003 kWh per search query (2022 Environmental Report)
Web searches = 50 × 0.0003 kWh = 0.015 kWh
Page fetches = 20 × 0.001 kWh = 0.020 kWh
Wikimedia CDN + Google Fonts = 0.005 kWh
──────────────────────────────────────────
Research total = 0.040 kWh
STEP 4 RESULT: ~0.04 kWh (rounds to noise)
Summary Table
| Component |
Assumption basis |
Low |
High |
Best est. |
| Claude servers (inference + PUE 1.3×) |
11.3M tokens × 0.003 kWh/1k × 1.30 |
25 kWh |
65 kWh |
44 kWh |
| Your devices (mobile + laptop) |
~6 hrs active, mixed device |
0.1 kWh |
0.5 kWh |
0.23 kWh |
| Netlify builds + CDN serving |
18 builds × 45s × 200W |
0.02 kWh |
0.10 kWh |
0.05 kWh |
| Web searches + page fetches |
50 searches, 20 fetches |
0.02 kWh |
0.06 kWh |
0.04 kWh |
| CDN assets (Wikimedia, Fonts) |
Static asset requests |
<0.01 |
<0.01 |
<0.01 |
| TOTAL |
|
~25 kWh |
~66 kWh |
~44 kWh |
Sensitivity Calculator
The kWh-per-token figure is the dominant uncertainty. Adjust it below along with session volume to see how the total changes. The math updates in real time.
⚡ Interactive Sensitivity Analysis
Drag sliders to adjust key assumptions — total recalculates live
44
Estimated kWh Total
Server: 44 kWh · Device: 0.23 kWh · Other: 0.09 kWh
What 44 kWh Means
🧊
6 weeks
Running a refrigerator continuously (≈7 kWh/week average)
🚗
~150 miles
Driving a Tesla Model 3 (≈0.3 kWh/mile EPA estimate)
📱
3.5 years
Charging an iPhone daily (≈0.035 kWh per full charge)
🏠
~4 days
Average US household electricity consumption (≈10.5 kWh/day, EIA 2023)
💡
30 days
Running a 60-watt incandescent bulb continuously
✈️
200× less
CO₂ of a round-trip flight Lansing → Chicago (≈180kg vs ~0.017kg)
Carbon Footprint
Assumption E1 — Grid Carbon Intensity
US average: 386g CO₂ per kWh (EPA eGRID 2022).
However: Anthropic uses AWS and Google Cloud, both of which purchase
significant renewable energy certificates and operate in regions with
cleaner grids. Real carbon is likely 30–60% lower than grid average.
Basis: EPA eGRID 2022; AWS 2023 Carbon Methodology; Google 2023 Environmental Report
Grid-average carbon = 44 kWh × 386g CO₂/kWh = 16,984g = ~17 kg CO₂
Adjusted for cloud renewable commitments (~40% reduction):
Likely actual carbon = 44 kWh × 232g CO₂/kWh = 10,208g = ~10 kg CO₂
For reference:
Burning 1 liter of gasoline ≈ 2.3 kg CO₂
This project ≈ burning 4–7 liters of gasoline
A round-trip flight Lansing→Chicago ≈ 180 kg CO₂ (~18× this project)
The Expert Research — Quantitative Foundations
The 15 expert recommendations in the Research Basis are grounded in peer-reviewed studies with documented effect sizes. Full transparency means showing the math behind those recommendations as well. Every formula below is from the cited primary source. Every effect size is from the published study. Every application to DECLASSIFIED is a direct inference from the published findings.
Inoculation theory models psychological resistance to misinformation as analogous to biological immunity. Exposing people to a weakened dose of a misinformation technique before they encounter it in the wild — "prebunking" — confers measurable protection.
Core Inoculation Effect Formula
Relative Risk Reduction (RRR) = (P_control − P_inoculated) / P_control × 100%
where:
P_control = proportion rating misleading content as reliable (no inoculation)
P_inoculated = same proportion after technique-based inoculation
P_control ≈ 0.65 (65% rated misleading content as reliable at baseline)
P_inoculated ≈ 0.43 (43% after inoculation — Roozenbeek et al. 2022)
RRR = (0.65 − 0.43) / 0.65 × 100% = 33.8% relative reduction in misinformation susceptibility
Cohen's d ≈ 0.22 (medium-small effect, n=1,586, p < 0.001)
⚙ DECLASSIFIED application: Adding Pattern Briefing before each claim (Rec #1)
predicts ~34% reduction in false verdicts from players encountering the technique cold.
Without the briefing, the technique exploits System 1 processing.
With it, System 2 is pre-armed.
Blanket Protection Finding
Inoculation against one manifestation of a technique (e.g., one Big Lie instance)
confers partial protection against novel manifestations of the same technique
(cross-protection coefficient ≈ 0.60–0.80). This is why the Phillips Pattern Library
works across Acts, not just within them.
The accuracy nudge operates on the finding that people share misinformation not primarily because they believe it, but because their attention is not oriented toward accuracy at the moment of decision. A single prompt shifts this.
Sharing Discernment Model (Pennycook & Rand 2021)
Discernment = (P_share_true − P_share_false)
where:
P_share_true = probability of sharing a true headline
P_share_false = probability of sharing a false headline
Δ Discernment = Discernment_nudge − Discernment_control
Control group: P_share_true = 0.52, P_share_false = 0.37 → Discernment = 0.15
Nudge group: P_share_true = 0.54, P_share_false = 0.28 → Discernment = 0.26
Δ Discernment = 0.26 − 0.15 = +0.11 (Cohen's d ≈ 0.19, p < 0.001)
Reduction in false headline sharing: ~24%
Effect replicated across 7 pre-registered studies (Pennycook et al. 2022, Nature Comms.)
⚙ DECLASSIFIED application: One sentence before the first claim — "Think about accuracy" —
predicts a 24% improvement in verdict accuracy for players who would otherwise
engage System 1 processing. Cost: 0 lines of code. Benefit: documented in Nature.
Even after a full correction is delivered and accepted, misinformation continues to influence reasoning. The effect persists not because people disbelieve the correction, but because the original trace remains accessible in memory and is retrieved automatically under cognitive load.
Continued Influence Persistence Ratio (π)
π = E_post / E_pre
where:
E_pre = effect of misinformation on inference (pre-correction)
E_post = residual effect of misinformation on inference (post-correction)
π = 1.0 means correction had zero effect
π = 0.0 means misinformation fully eliminated
Empirical range: π ≈ 0.40–0.65 across studies
Example (Ecker et al. 2011): fire caused by flammable materials (retracted) still
cited by 31% of participants in open-ended questions even after accepting correction.
Without retraction: 55% cited it. Persistence ratio: 31/55 = 0.56
Misinformation retains 40–65% of its inferential effect even after accepted correction.
The myth leaves a residue. Corrections recover ~35–60% of that residue.
⚙ DECLASSIFIED application: Displaying "Mexico will pay for the wall" prominently
before verdict means the claim runs at full strength (E_pre = 1.0) before the player
evaluates it. Even a correct FALSE verdict leaves π ≈ 0.50 residual effect.
Leading with the analytical frame instead of the myth eliminates the pre-exposure cost.
Illusory Truth Interaction
Each unframed repetition of a false claim in summaries or navigation adds
incremental truth ratings of 10–15% per exposure (Hasher et al. 1977).
The continued influence effect and illusory truth effect compound:
the game currently runs both deficits simultaneously on Hall of Shame display.
Ebbinghaus's forgetting curve is the most replicated finding in memory research. Retention decays as a negative exponential function of time. Spaced repetition extends the stability parameter S, shifting the curve rightward.
Ebbinghaus Retention Formula
R(t) = e^(−t / S)
where:
R(t) = memory retention at time t (proportion, 0–1)
t = time since encoding (days)
S = memory stability (increases with each successful retrieval)
e = Euler's number ≈ 2.718
Single encoding (S = 1):
R(1) = e^(−1/1) = 0.368 → 37% retained after 1 day
R(7) = e^(−7/1) = 0.001 → 0.1% retained after 1 week
R(30) = e^(−30/1) ≈ 0.000 → essentially 0 after 1 month
With spaced repetition (3 reviews, S grows: 1 → 2 → 4 → 8):
R(7) = e^(−7/8) = 0.417 → 42% retained after 1 week
R(30) = e^(−30/8) = 0.024 → 2.4% — still requires another review
R(30) after 4th review (S = 16):
= e^(−30/16) = 0.153 → 15% retained at 30 days
Without Review Mode: a player who scores 80% on Act III on Day 0
retains approximately 0.1% of that learning by Day 7.
With spaced retrieval (3 reviews at 1, 3, 7 days): retention at Day 30 ≈ 15–42%.
The testing effect (Roediger & Karpicke 2006) adds ~15 percentage points on top:
retrieval practice vs. re-reading: d = 0.50, p < 0.001.
⚙ DECLASSIFIED application: Review Mode surfacing missed claims at
t = 1, 3, 7 days converts a 0.1% week-1 retention rate to ~42%.
This is not marginal improvement. It is the difference between a game
session and a learning system.
Repeated exposure to any statement increases its perceived truth, regardless of the statement's actual accuracy. The mechanism is processing fluency: statements that feel familiar are processed more easily, and ease of processing is misattributed to truth.
Linear Repetition Model (Fazio et al. 2015)
T(n) = T₀ + (β × n)
where:
T(n) = perceived truth rating after n exposures (7-point scale)
T₀ = baseline truth rating (single exposure)
β = truth increment per additional exposure
n = number of exposures
Fazio et al. (2015) experimental results (trivia claims, 7-point scale):
T₀ ≈ 3.2 (single exposure)
T(3) ≈ 3.6 (three exposures)
β ≈ 0.13 per exposure (each repetition adds ~0.13 points on 7-point scale)
Percent increase: (3.6 − 3.2) / 3.2 × 100% = 12.5% per 3-exposure cycle
Pennycook et al. (2018): illusory truth effect persists even when participants
KNOW the claim is false (effect size d ≈ 0.18 even for items rated implausible).
Each unframed repetition of a false claim increases its perceived truth by
~10–15% per 3-exposure cycle.
The effect is not neutralized by prior knowledge that the claim is false.
⚙ DECLASSIFIED application: The Hall of Shame displays false claims
without corrective framing. If a player sees "The crowd was the largest
in inaugural history" four times (initial claim + summary + Hall of Shame +
Review Mode), perceived truth increases T(4) = 3.2 + (0.13 × 4) = 3.72 (+16%).
Fix: display corrected framing in all repeat contexts.
Working memory has a hard capacity limit. When the number of simultaneously active cognitive objects exceeds that limit, performance degrades nonlinearly. This is not a gradual effect — it is a threshold phenomenon.
Working Memory Capacity Model (Cowan 2001)
WM_capacity ≈ 4 ± 1 chunks (Cowan 2001 update to Miller's 7 ± 2)
Cognitive load for a task = Σ(intrinsic_load) + extraneous_load + germane_load
where:
Intrinsic load = complexity inherent to the material
Extraneous load = unnecessary complexity from poor presentation
Germane load = cognitive effort that contributes to learning (desired)
Performance degrades when total load > WM_capacity
DECLASSIFIED claim assessment demands (simultaneous cognitive objects):
Object 1: Read and parse claim text (1 chunk)
Object 2: Hold claim in memory during verdict (1 chunk)
Object 3: Select verdict from 4 options (1 chunk)
Object 4: Select citation tier from 4 options (1 chunk)
Object 5: Select pattern from 14 options (2–3 chunks — category + item)
Object 6: Weight multiplier context (1 chunk)
──────────────────────────────────────────────
Total: 7–8 chunks vs. capacity of 4 ± 1
New players face a cognitive load of 7–8 chunks against a 4 ± 1 chunk capacity.
This creates extraneous load (overload not contributing to learning) on claims 1–4.
Performance under overload: accuracy drops 20–40% relative to within-capacity tasks
(Sweller 1988, multiple replications).
⚙ DECLASSIFIED application: Scaffolded introduction (verdict only → verdict+citation
→ full 4-dimension) reduces peak cognitive load from 7–8 chunks to 3–4 on first exposure.
Estimated accuracy improvement on claim 1: +20–35% for new players.
Kruger & Dunning's original study measured the gap between perceived and actual performance across four experiments (logical reasoning, grammar, humor). The pattern is consistent: low performers systematically overestimate; high performers slightly underestimate.
Overconfidence Index (Kruger & Dunning 1999, Study 1)
Overconfidence Index = (Perceived_Percentile − Actual_Percentile) / 100
Calibration Error = |Perceived_Percentile − Actual_Percentile|
For each performance quartile:
Bottom quartile: Actual = 12th pct, Perceived = 62nd pct → OCI = +0.50
2nd quartile: Actual = 34th pct, Perceived = 63rd pct → OCI = +0.29
3rd quartile: Actual = 66th pct, Perceived = 67th pct → OCI = +0.01
Top quartile: Actual = 89th pct, Perceived = 70th pct → OCI = −0.19
The bottom quartile's OCI of +0.50 means they believe themselves
to be 50 percentile points higher than they are.
The top quartile underestimates by 19 percentile points (the "unskilled-at-meta-cognition" problem
applies inversely at the top).
Bottom-quartile performers operate with a +50 percentile point confidence surplus.
Top-quartile performers operate with a −19 percentile point confidence deficit.
The Dunning-Kruger curve crosses calibration (OCI = 0) at approximately the 60th percentile.
⚙ DECLASSIFIED application: PatriotEagle1776 (23 points, total certainty) is operating
at the theoretical maximum OCI = +0.50. The Dunning-Kruger Guarantee section maps this
directly onto the game's scoring tiers. The retired federal judge (31 years, Arbiter Tier)
is operating at OCI ≈ −0.19: correct but self-doubting. That's the intended endpoint.
Replication Status (Important)
Subsequent research (Gignac & Zajenkowski 2020; Nuhfer et al. 2016) demonstrated
that the Dunning-Kruger pattern can be partially reproduced using random data,
suggesting a component attributable to regression to the mean and the better-than-average
effect. The effect is real; its magnitude may be overstated in popular accounts.
DECLASSIFIED's Dunning-Kruger section cites this caveat explicitly.
Graham, Haidt & Nosek's Moral Foundations Questionnaire produced the largest cross-partisan dataset on moral foundation endorsement. The finding is not that conservatives are more moral or liberals are more moral — it is that they are moral in different proportions.
Foundation Endorsement by Political Identity (MFQ, Graham et al. 2009)
Endorsement score = mean of foundation-specific items (1–6 scale)
Differential = Liberal_score − Conservative_score
Foundation scores (approximate, n > 23,000):
Care/Harm: Liberal = 3.89, Conservative = 3.62 → Δ = +0.27 (Liberal higher)
Fairness: Liberal = 3.87, Conservative = 3.46 → Δ = +0.41 (Liberal higher)
Loyalty: Liberal = 2.45, Conservative = 3.18 → Δ = −0.73 (Conservative higher)
Authority: Liberal = 2.27, Conservative = 3.06 → Δ = −0.79 (Conservative higher)
Sanctity: Liberal = 1.91, Conservative = 2.97 → Δ = −1.06 (Conservative higher)
Liberal moral foundation profile: Care-dominant, Fairness-secondary.
Conservative moral foundation profile: distributed across all five,
with Sanctity and Authority showing the largest partisan gap.
DECLASSIFIED's annotation register operates primarily on Care (accuracy matters
because it harms people) and Fairness (the primary sources deserve equal application).
These foundations have a +0.27 to +0.41 liberal-direction differential.
They are systematically less resonant with conservative moral architecture
by 0.73–1.06 points on the same scale.
⚙ DECLASSIFIED application: A Loyalty-framed annotation — "American taxpayers funded
this construction, not Mexico — here's the appropriations record" — activates the
same documented fact through a foundation with −0.73 partisan differential.
The record is identical. The framing differential is 0.73 scale points.
That is the mathematical case for the Framing Lens toggle (Rec #8).
The most counterintuitive finding in political psychology: higher numeracy and analytical ability predict STRONGER identity-protective cognition among partisans, not weaker. Smarter partisans are better at motivated reasoning, not less susceptible to it.
Kahan et al. (2017) — Numeracy × Partisanship Interaction
Polarization_index = |Liberal_assessment − Conservative_assessment| on same data
Predicted by standard model (more numeracy → less bias):
Polarization_index should DECREASE as numeracy increases
Actual finding (n = 1,111, Kahan et al. 2017):
Polarization_index INCREASES with numeracy for identity-threatening topics
β_numeracy_×_partisanship = +0.43 (95% CI: 0.28–0.58, p < 0.001)
High-numeracy liberals correctly identified liberal-confirming gun crime data: 71%
High-numeracy conservatives correctly identified conservative-confirming data: 68%
Same people, same math ability, opposite answers depending on identity-framing.
β = +0.43: each standard deviation increase in numeracy predicts a 0.43 SD
increase in identity-protective bias on politically-charged data.
Analytical ability does not protect against motivated reasoning.
It amplifies it for identity-threatening information.
The game frame (analytical skill test, not political opinion survey)
is the correct structural intervention: it activates the analytical identity
rather than the partisan identity before the claim is presented.
⚙ DECLASSIFIED application: The one-sentence framing "This is a test of
analytical skills, not political beliefs" suppresses the partisan identity frame
before it can be activated. Without it: high-numeracy partisans are the hardest
players to move, not the easiest. With it: analytical identity is recruited.
Descriptive social norms — information about what most people actually do — are among the most reliable behavioral predictors across cultures and contexts. The hotel towel study is the canonical benchmark.
Descriptive Norm Compliance Uplift (Goldstein, Cialdini & Griskevicius 2008)
Compliance_uplift = (Compliance_norm − Compliance_baseline) / Compliance_baseline × 100%
Hotel towel experiment:
Control (standard environmental message): reuse_rate = 35.1%
Descriptive norm ("75% of guests in this room reuse"): reuse_rate = 44.1%
Compliance_uplift = (44.1 − 35.1) / 35.1 × 100% = +25.6%
Injunctive norm ("please help save the environment"): reuse_rate = 36.3%
→ Injunctive uplift = only +3.4% (vs +25.6% for descriptive)
Descriptive norms outperform injunctive norms by 7.5× in this benchmark.
"Most people do X" is 7.5× more effective than "you should do X."
Adding "76% of players correctly identified this as FALSE" to verdict feedback
predicts a ~20–26% increase in correct verdict selection
for players who would otherwise be uncertain.
The message must be accurate — fabricated social proof undermines trust and reverses the effect.
⚙ DECLASSIFIED application: This recommendation (Rec #9) is conditional on
having real analytics. Without real session data, the feature cannot be implemented
without compromising the site's entire evidentiary premise. The math is sound.
The prerequisite is the analytics infrastructure.
Aggregate Expected Improvement — If All 10 Applied
These are not additive — some effects interact, some share mechanisms. The following is a conservative lower-bound estimate of combined impact assuming partial independence between interventions.
Stacked Effect Estimate (Conservative, Assuming 50% Interaction Discount)
Individual effect contributions (rough estimates, primary-source grounded):
R1 — Inoculation prebriefing: +34% verdict accuracy on new techniques
R2 — Accuracy nudge: +24% discernment on borderline claims
R3 — Myth-first correction eliminated: +15% (removes continued influence load)
R4 — Spaced repetition added: +42 percentage points retention at 1 week
R5 — Illusory truth eliminated: +10% (removes repetition truth inflation)
R6 — Cognitive load scaffolded: +20–35% accuracy for new players, claims 1–4
R7 — Dunning-Kruger surfaced: Diagnostic only — no direct accuracy effect
R8 — Moral foundations framing: +15–25% receptivity in non-Care/Fairness audiences
R9 — Identity framing corrected: Prevents motivated reasoning activation
R10 — Social proof added: +20–26% compliance on uncertain verdicts
Applying 50% interaction discount to avoid overcounting overlapping mechanisms:
Estimated combined impact on a new player's first-session verdict accuracy:
Baseline (current): ~55–65% correct verdicts (estimated from game design)
With all 10 applied: ~72–80% correct verdicts
Net improvement: +15–25 percentage points
On week-1 retention (currently near 0 without Review Mode):
With spaced repetition: +40 percentage points (the single largest lever)
What These Numbers Cannot Tell You
These are inferences from published research applied to a specific game context. None of them are direct measurements of DECLASSIFIED player behavior — that data does not yet exist (see Rec #15: analytics). The effect sizes are from the primary studies cited. The application to DECLASSIFIED is a structured inference, not a controlled experiment.
The correct interpretation: the research literature predicts these improvements with high confidence for the general population. The actual effect on DECLASSIFIED's specific player population — which self-selects for political interest and is already above average in analytical engagement — may be higher or lower. Only measurement will tell. That is what Recommendation 15 is for.
The Honest Caveat — What This Number Cannot Tell You
The kWh/token figure is not published by Anthropic. The 0.003 kWh/1,000 tokens baseline is inferred from academic literature on comparable frontier models. If the actual figure is 0.001 kWh/1,000 tokens (possible on modern H100 hardware at high utilization), the total drops to ~15 kWh. If it is 0.008 kWh/1,000 tokens (possible on older hardware at low utilization), it rises to ~90 kWh.
The token count is estimated. We know the turn count from the session history. The average context window size (60,000 input tokens) is inferred from typical long-session growth patterns. The actual number could reasonably range from 40,000 to 100,000 tokens per turn.
The range you should hold: 15–90 kWh, with 40–50 as the most defensible midpoint.
This is the same standard we apply to everything on this site: show the methodology, document the assumptions, acknowledge the uncertainty. The primary sources are there. The math is there. Challenge any assumption that doesn't hold.
← Methodology ·
Research Basis ·
Play DECLASSIFIED →