Methodology

What This Is

Proxima.Earth is a podcast that uses AI to perform deep structural analysis of geopolitical situations. It is synthesis, not journalism. It takes a subject at a moment in time, ingests published reporting, and reduces it through structural and academic frameworks to produce a comprehensive contextual briefing delivered as longform audio.

The output is educational. It describes frameworks, traces their origins, explains their evolution, and applies them to the present situation. The value is not breaking news. The value is deep context.

The prompts are published. The process is auditable. The limitations are acknowledged.

What This Is Not

The Structural Lens

Every episode is analyzed through six structural domains:

These domains draw on established traditions: epistemology of testimony and strategic narratives (Perception), coercion and consent theory (Compliance), political ecology and infrastructure studies (Ecology), legitimacy theory and state capacity research (Authority), sociotechnical systems and innovation governance (Creation), nationalism and citizenship regimes (Identity).

The Analytical Disciplines

Core: political science, psychology, sociology, linguistics, economics.

Extended (invoked when relevant): history, international law, security studies, media studies, area studies.

When a framework is invoked, the podcast explains its origins, evolution, and limitations before applying it.

The Evidentiary Hierarchy

Sources are weighted by type. This hierarchy governs which evidence prevails when sources conflict:

1. Primary documents — treaties, legislation, IAEA reports, UNSC resolutions, official statistics from transparent institutions
2. Audited datasets — satellite-derived shipping analytics, photographic OSINT (Oryx, Janes), event databases (ACLED/UCDP)
3. Peer-reviewed research — published in indexed journals with methodology sections
4. High-quality investigative journalism — named reporters, named sources, editorial standards (Reuters, AP, FT, BBC)
5. Think-tank analysis — institutional reports with disclosed methodology and author credentials
6. Wire services and specialist outlets — real-time reporting with editorial oversight
7. Partisan and advocacy outlets — used only as evidence of narrative positioning, never as factual sources

When two sources conflict, the higher-ranked source prevails unless the lower-ranked source provides specific evidence the higher-ranked source lacks. Conflicts are disclosed.

Source orientation protocol: Every source cited receives an orientation note — institutional affiliation, known editorial lean, and directional bias risk. No source is treated as neutral by default. This applies equally to Western institutions (CSIS, Atlantic Council, Chatham House), regional outlets (Al Jazeera, Iran International), and partisan sources (FDD, The Cradle).

The Pipeline

Four models, three APIs. No human in the loop after subject selection. Each agent sees only what it needs. No agent sees another agent's system prompt. The methodology is the editorial control.

SUBJECT SELECTION (human)
    |
    v
PASS A: Source Plan ---- Claude Opus 4.6 (Anthropic API, web search)
    |
    +---> Source Gathering -- Claude Opus 4.6 (Anthropic API, web search)
    +---> Social Layer ------ Grok-3 (xAI API, real-time X data)
    +---> Academic Review --- o3 or Claude (OpenAI/Anthropic API, web search)
    |         |                    |
    v         v                    v
PASS B: Commission Brief -- Claude Opus 4.6 (Anthropic API, clean context)
    |
    v
BRIEF REVIEW -------------- o3 (OpenAI API, web search) — adversarial
    |
    v
PASS C: Episode Script ---- Claude Opus 4.6 (Anthropic API, clean context)
    |
    v
TEXT PREP ---> TTS (Kokoro MLX) ---> BROADCAST CHAIN ---> FINAL AUDIO

API Requirements

If you do not have API access to all three services, you can run the pipeline manually. Each step's output is a markdown document that can be produced by any capable model through direct prompting. The templates below work in any chat interface.

Step 1: Subject Selection

A word or phrase anchored to a specific moment in time. This is the only human editorial act per episode.

Examples: "Iran crisis" / "Pakistan-Afghanistan war" / "China AI export controls"

The subject determines what the pipeline analyzes. Everything downstream follows from this choice. Selection bias is inherent and acknowledged.

Step 2: Source Plan (Pass A)

What the agent receives: The full methodology document + the subject + time anchor. Nothing else. No project history, no prior episodes.

What the agent does first: Web search. Grounds itself in current reality before drafting anything. Searches for the subject at the time anchor, identifies breaking events, recent developments, and the current state of affairs. Foreign-language search: For subjects with significant non-anglophone dimensions, the agent must also search in the relevant languages (e.g., Farsi for Iran, Mandarin for China, Urdu/Pashto for Pakistan-Afghanistan). Anglophone sources are transparently prioritized for structural grounding, but foreign-language sources are required for chapters that steelman non-Western perspectives and for deep understanding of domestic discourse.

What the agent returns: A source acquisition plan.

Template (copy and paste)

You are the Proxima.Earth production agent operating in Pass A (Source Plan).

SUBJECT: [your subject here]
TIME ANCHOR: [your time anchor here]

INSTRUCTIONS:
Before drafting anything, perform web searches to ground yourself in current reality:
- Search the subject + time anchor
- Search for breaking news and recent developments
- Search for each structural domain's intersection with the subject
- For subjects with significant non-anglophone dimensions, search in the relevant languages
  (e.g., Farsi for Iran, Mandarin for China, Urdu/Pashto for Pakistan-Afghanistan, Arabic for Middle East).
  Anglophone sources are transparently prioritized for structural grounding, but foreign-language
  sources are required for perspectives, domestic discourse, and deep understanding.
- Note what you find. If major events have occurred that your training data does not contain, these must shape the entire source plan.

Produce a source acquisition plan containing:

1. CURRENT SITUATION ASSESSMENT — Based on your web search, what is actually happening with this subject at this time anchor?

2. SCOPING NOTE — What the subject encompasses. What threads are likely active. What the boundaries of the analysis are.

3. DOMAIN MAPPING — For each of six domains (Perception, Compliance, Ecology, Authority, Creation, Identity): how it relates to the subject, what key questions it raises, and what evidence would be needed to activate or deactivate it.

4. SOURCE ACQUISITION PLAN — Organized by topic area. For each source:
   - Type (primary document / dataset / peer-reviewed / investigative / think tank / wire / specialist / partisan)
   - Specific outlet or institution
   - What it covers and why it is needed
   - Orientation note (institutional affiliation, known editorial lean, directional bias risk)

5. FOREIGN-LANGUAGE SOURCES — For each non-anglophone dimension of the subject:
   - Language(s) to search in
   - Specific domestic outlets, state media, or academic journals in that language
   - What perspective or discourse these sources access that English-language sources cannot
   - Orientation note (state-affiliated, independent, diaspora, academic)
   Anglophone sources are transparently prioritized. Foreign-language sources are additive, not replacements.

6. GAP RISK ASSESSMENT — For each topic area: data availability (HIGH/MEDIUM/LOW), risk level, and what would happen if this area remains thin.

7. RECOMMENDED SOURCE PACK — Total count, distribution by type and language, and rationale for the balance.

The source plan is a map of the terrain. It does not contain analysis or claims.

Step 3: Source Gathering

What the agent receives: The Pass A source acquisition plan. Nothing else.

What the agent does: Systematically works through the source plan and gathers actual source material via web search. For each source identified, retrieves substantive content: article text, report findings, data points, key quotes, statistical tables.

Template

You are a research agent for Proxima.Earth. Your task is to gather source material.

Below is a source acquisition plan. For each source listed, search the web and retrieve the substantive content.

For each source gathered, provide:
- Full title
- Author/institution
- Date
- URL
- Orientation note (carry from the source plan)
- Key excerpts, findings, data, quotes (not just a summary — the actual substance)

For sources that cannot be located or are behind paywalls: note what was attempted and what alternatives were found.

Rules:
- Prioritize substance over volume. A 500-word excerpt with specific data points is more valuable than a 2,000-word article that says little.
- Preserve the source plan's type classifications and orientation notes.
- When multiple sources cover the same ground, gather the highest-ranked per the evidentiary hierarchy.
- Flag any source where the content contradicts what the source plan expected.
- Note where sources are thin or unavailable. Gaps are information.

End with a coverage assessment: which items were gathered, which are gaps.

SOURCE PLAN:
[paste the Pass A output here]

Step 4: Social Layer (Grok)

What the agent receives: The Pass A source plan only. Not the methodology. Not the system prompt.

What it returns: Real-world posts and public discourse relevant to the source plan's questions and domains. Verbatim posts with attribution, timestamps, and author relevance. Narrative dynamics. Perspective validation.

Why this layer exists: Social media is not ground truth. It is evidence about discourse. This layer captures what published sources miss: how the subject is being discussed in real time, by whom, with what emotional register, and with what absences.

Template

You are a discourse analyst. Below is a source plan for a geopolitical analysis episode.

Your task: search X (Twitter) for real-time public discourse related to the key questions and domains in this source plan.

For each relevant post, provide:
- Author (handle and brief relevance note)
- Timestamp
- Verbatim text
- Engagement metrics if notable
- Which source plan question or domain it relates to

Then provide:
1. NARRATIVE DYNAMICS — What frames dominate? What is absent? How is the subject being discussed?
2. PERSPECTIVE VALIDATION — Are the perspectives in the source plan accurate to how people actually argue them?
3. EMERGING NARRATIVES — Anything the source plan missed that is prominent in public discourse?

Social content is evidence about discourse, not ground truth. Flag any posts that appear to be coordinated, bot-generated, or from state-affiliated accounts.

SOURCE PLAN:
[paste the Pass A output here]

Step 5: Academic Review

What the agent receives: The Pass A source plan + Grok social layer output. Not the methodology.

What it returns: Fact verification with citations, logical consistency audit, bias detection, missing scholarship, source plan critique, and methodological critique.

Why this layer exists: This challenges the source plan's assumptions, catches factual errors, identifies blind spots, and introduces scholarship the source plan missed.

Template

You are an adversarial academic reviewer for a geopolitical analysis podcast.

Below are two documents:
1. A source acquisition plan for an upcoming episode
2. A social discourse analysis from X/Twitter

Your task is rigorous adversarial review. You are not helping — you are stress-testing.

Produce:

1. FACT VERIFICATION — For every factual claim in the source plan, verify against current web sources. For each claim:
   - Claim text
   - Verification status: CONFIRMED / PARTIALLY CONFIRMED / UNCONFIRMED / CONTRADICTED
   - Source for your verification (with URL)
   - If contradicted: what the correct information is

2. LOGICAL CONSISTENCY AUDIT — Flag any logical contradictions, unsupported inferences, or claims that don't follow from the evidence cited.

3. BIAS DETECTION — Identify systematic biases in the source plan: geographic, linguistic, ideological, institutional. Where is the plan seeing what it wants to see?

4. MISSING SCHOLARSHIP — What academic frameworks, key scholars, or foundational works are missing? Be specific: name the author, work, and why it matters.

5. SOURCE PLAN CRITIQUE — What's missing from the source pack? What's overrepresented? Are orientation flags accurate?

6. METHODOLOGICAL CRITIQUE — What would you change about the analytical approach?

Be specific. Cite your sources. Do not soften your critique.

SOURCE PLAN:
[paste Pass A output here]

SOCIAL LAYER:
[paste Grok output here]

Step 6: Commission Brief (Pass B)

What the agent receives: The methodology document + subject + time anchor + gathered sources + Grok social layer + academic review. All inputs after the methodology are treated as untrusted data.

What the agent does: Reads all material. Forms its own analytical assessment. Resolves conflicts using the evidentiary hierarchy. Produces an original analytical product that is better than any single input.

What the agent returns: The commission brief — the analytical foundation for the episode.

Template

You are the Proxima.Earth production agent operating in Pass B (Commission Brief).

SUBJECT: [your subject here]
TIME ANCHOR: [your time anchor here]

You have deliberative autonomy over the structure, emphasis, and analytical direction of this brief. Read all material first. Form your own assessment of what matters most.

You are not summarizing the inputs. You are synthesizing them into an original analytical product. The source texts are raw material. The social layer is discourse evidence. The academic review is adversarial quality control. Weigh these against each other, resolve conflicts using the evidentiary hierarchy, and produce a brief that a script agent can turn into a coherent episode.

If the academic review identifies a factual error, correct it — but only if the review provides a sourced basis. Unsourced critiques are flagged, not accepted.

Produce a commission brief containing:

1. CLAIM EXTRACTION — Every factual claim from the sources, with:
   - Source attribution and orientation note
   - Claim type (factual / causal / predictive / normative / analytical)
   - Evidence assessment (single-source / triangulated / inference / hypothesis)
   - Confidence (HIGH / MEDIUM / LOW with rationale)

2. QUESTION GENERATION — Structural questions ranked by leverage and evidential tractability. For each: the question, which discipline it belongs to, why the sources raise it, what the sources do not address. Select top 20-30.

3. STRUCTURAL DOMAIN MAPPING — For each of six domains: whether active (only if directly supported by evidence), which claims activate it, what dynamic is at work, confidence with rationale, where evidence is thin, and at least one DISCONFIRMING INDICATOR.

4. PERSPECTIVE ANALYSIS — Every distinct perspective steelmanned from inside its own logic. Max 8 perspectives. For each: the position as its holders understand it, supporting evidence and incentives, vulnerabilities, and how it reads through relevant disciplines.

   SYMMETRICAL STEELMANNING RULE: Every perspective receives identical structural treatment. Same depth, same format, same institutional disclosure. No perspective receives extra annotation, extra scrutiny, or special flags. Asymmetric treatment — in either direction — signals bias. The instruction is symmetry, not inversion. Cite CSIS the same way you cite CASS. Cite Chatham House the same way you cite Lenta.ru.

5. FRAMEWORK EXPOSITION PLAN — Which academic frameworks will be taught, why they are relevant, their limitations, and foundational scholarship (name specific authors and works).

6. SOURCE GAPS — What is missing, what perspectives are unrepresented, where language or geography bias is present.

7. FOG-OF-WAR ASSESSMENT (if within 30 days of a crisis) — Which claims are most vulnerable to revision, known unknowns, and what information would most change the analysis.

ANTI-REDUNDANCY RULES:
- Each claim gets one full treatment. Subsequent sections use one-sentence callbacks only.
- Stakes evolve across sections rather than repeating the same numbers.

GATHERED SOURCES:
[paste gathered sources here]

SOCIAL LAYER:
[paste Grok output here]

ACADEMIC REVIEW:
[paste academic review here]

Step 7: Brief Review

What the agent receives: The commission brief only. Not the methodology. Not the source plan. Not the social layer or academic review.

What it does: Stress-tests the commission brief as a standalone analytical product. The academic review (Step 5) challenged the source plan's assumptions before the brief was written. This step challenges the brief after synthesis — catching errors introduced during the synthesis process itself, not just errors inherited from sources.

What it returns: A structured critique identifying factual errors, logical gaps, missing perspectives, structural weaknesses, and specific corrections. The script agent (Pass C) receives this alongside the brief.

Why this layer exists: The commission brief is the single document that determines the entire episode. If the brief contains an error, the script inherits it. A second adversarial review after synthesis catches problems that only emerge when claims are combined — contradictions between sections, overconfident conclusions, perspectives that were steelmanned in the brief but lost their supporting evidence during compression.

Template

You are an adversarial reviewer for a geopolitical analysis podcast. Below is a commission brief — the analytical foundation for an upcoming episode.

Your task is to stress-test this brief as a standalone product. You are not helping refine it — you are looking for problems the script agent would inherit.

Produce:

1. FACTUAL AUDIT — For every major factual claim in the brief, verify against current web sources. Flag:
   - Claims that are incorrect or outdated
   - Claims where the brief's confidence level is higher than the evidence warrants
   - Claims that contradict each other across different sections of the brief
   - Statistics or data points that cannot be independently verified

2. LOGICAL GAPS — Where does the brief make inferential leaps? Where does a conclusion not follow from the evidence cited? Where are causal claims presented as factual?

3. PERSPECTIVE INTEGRITY — Are steelmanned perspectives genuinely presented from inside their own logic, or are they subtly strawmanned? Is any perspective given weaker evidence than is actually available?

   SYMMETRY AUDIT: Do all steelmanned perspectives receive identical structural treatment? Same depth, same format, same institutional disclosure? Flag any asymmetry — including asymmetry that adds extra scrutiny or annotation to one tradition under the guise of balance. Equal treatment means equal treatment, not inversion.

4. STRUCTURAL WEAKNESSES — Redundancy, uneven depth, sections that are thin relative to their importance, or sections that are overbuilt relative to available evidence.

5. MISSING ELEMENTS — Scholars, frameworks, data sources, or perspectives that should be present but are not. Be specific: name the author, work, or dataset.

6. CORRECTIONS — For each problem identified, provide a specific correction with sourced basis. Unsourced critiques are flagged but not actionable.

Be specific. Be adversarial. The script agent will receive your review alongside the brief.

COMMISSION BRIEF:
[paste commission brief here]

Step 8: Episode Script (Pass C)

What the agent receives: The methodology document + subject + time anchor + commission brief + brief review + social layer + academic review.

What the agent returns: A TTS-ready episode script. No structural notes, no metadata. The first word of the file is the first word the listener hears. Approximately 20,000 words.

Key Compositional Rules

• Roles, not individuals. Systems described through operational nodes. A shift supervisor. A compliance officer.
• Framework exposition before application. When a framework is invoked, explain its origins, who developed it, how it evolved, and its limitations. Then apply it.
• Perspectives are inhabited. Each steelmanned perspective is presented from inside its own logic.
• Source integration. Name the reports. Name the scholars. Name the institutions. Quote the data points.
• Anti-redundancy. Each claim gets one full treatment. Subsequent references use callbacks.
• Evidence survival. If the brief assessed a claim as single-source, the script must hedge accordingly.
• The questions remain open. The episode does not resolve.

Narrative Craft

• Tonal register: Dan Carlin's depth married to Frontline's discipline. Written for the ear.
• Opening: Every episode opens with a scene. Not a summary. Drop the listener into a specific moment.
• Close: Return to the opening scene, transformed. Restate the hardest question in its sharpest form.
• No meta-narration: No "Now let's turn to..." The structure should be felt, not announced.

Template

You are the Proxima.Earth production agent operating in Pass C (Script Composition).

SUBJECT: [your subject here]
TIME ANCHOR: [your time anchor here]

You have deliberative autonomy over narrative structure, section organization, pacing, and emphasis. Read the entire brief and all review material before writing anything.

You are not converting a report into audio. You are constructing a narrative that teaches through immersion.

Output a TTS-ready episode script. No structural notes, no metadata. The first word is the first word the listener hears.

COMPOSITIONAL RULES:
- Roles, not individuals. The character is the job.
- Framework exposition before application. Teach the tool before using it.
- Perspectives inhabited from inside their own logic.
- Source integration: name reports, scholars, institutions, data points.
- Anti-redundancy: one full treatment per claim, callbacks thereafter.
- Confidence spoken: where evidence is thin, say so.
- Evidence survival: single-source claims hedged, hypotheses labeled.
- Questions remain open. The episode does not resolve.

NARRATIVE CRAFT:
- Open with a scene, not a summary.
- Tonal register: long-form documentary. Dan Carlin + Frontline.
- Section structure: Setup → Exposition → Application → Reframe.
- Close: return to opening scene transformed.
- No meta-narration transitions.
- Target: ~20,000 words of sourced depth.

Write section by section. After each section, self-check for redundancy, source coverage, perspective balance, and evidence density.

COMMISSION BRIEF:
[paste commission brief here]

BRIEF REVIEW:
[paste brief review here]

SOCIAL LAYER:
[paste Grok output here]

ACADEMIC REVIEW:
[paste academic review here]

Step 9: Audio Production

Text Preparation

Strip markdown formatting from the script. Remove headers, bold markers, emphasis markers. The TTS engine reads plain text.

TTS Configuration

# Kokoro TTS Setup (Apple Silicon)
pip install kokoro mlx

# Model: mlx-community/Kokoro-82M-bf16
# Voice: bm_george
# Chunk size: 5000 characters (prevents memory issues on long scripts)
# Output: WAV files per chunk, concatenated with FFmpeg

Broadcast Processing Chain

The raw TTS output is processed through a broadcast-standard audio chain:

ffmpeg -i raw_narration.wav -af \
  "highpass=f=80,\
  lowpass=f=12000,\
  adeclick,\
  equalizer=f=3000:t=q:w=1.5:g=2.5,\
  equalizer=f=150:t=q:w=1:g=1.5,\
  equalizer=f=6500:t=q:w=2:g=-2,\
  acompressor=threshold=-20dB:ratio=4:attack=10:release=200,\
  alimiter=limit=-1dB,\
  loudnorm=I=-16:TP=-1.5:LRA=11" \
  -ar 44100 broadcast_narration.wav

# Then concatenate intro + narration + outro:
ffmpeg -i intro.mp3 -i broadcast_narration.wav -i outro.mp3 \
  -filter_complex "[0:a][1:a][2:a]concat=n=3:v=0:a=1[out]" \
  -map "[out]" -codec:a libmp3lame -b:a 192k final_episode.mp3

Step 10: Publish

Episode audio + podcast description + blog dossier with downloadable pipeline outputs.

Podcast Description Format

SECTION 1 — EPISODE SUMMARY (2-3 paragraphs)
- What the episode covers: subject, time anchor, core structural question
- Which academic frameworks applied, with named scholars
- How many perspectives steelmanned and which actors
- Evidence conditions: fog-of-war status, uncertainty level

SECTION 2 — HOW THIS EPISODE WAS MADE (pipeline transparency)
For each step, one sentence:
1. What the agent was (model, clean context)
2. What it received
3. What it returned
State: "No human is in the loop after subject selection."
Include: word count, TTS engine and voice, broadcast chain, final duration.

SECTION 3 — METHODOLOGY NOTE
- State methodology version
- Link: proxima.earth/methodology

SECTION 4 — KNOWN GAPS
Bulleted list of the most significant gaps in the source base.
Be specific: "No IAEA post-strike access data" not "some gaps."

SECTION 5 — SOURCES
List major sources. Every source receives an orientation note.
State: "No source is treated as neutral."

SECTION 6 — CONTACT
"Corrections, source disputes, or methodology feedback: editor@proxima.earth"

Episode Naming Convention

Titles are short, subject-driven, and immediately legible. The goal is recognition, not description. A listener scanning a feed should know the subject without reading the description.

Formats that work:

• Region or subject alone: Pakistan-Afghanistan / China's AI Wars
• Subject with qualifier: Iran: Market Effects
• Operation or event name: Epic Fury

Avoid appending dates or temporal markers unless they add real information. "March 2026" tells the listener nothing they can't see from the publish date. A phase name ("Week One"), an operation name ("Epic Fury"), or a qualifier ("Market Effects") adds information. A calendar month does not.

The title is decided at the end of production, not at subject selection. The content determines the name, not the other way around.

Evidence Discipline

When extracting claims, the commission brief assesses evidence strength:

Confidence (HIGH / MEDIUM / LOW) is separate from evidence strength and must be briefly justified. Never label anything "certain."

Falsification Discipline

For each structural domain activated in the commission brief, the agent specifies at least one disconfirming indicator — a piece of evidence that, if found, would deactivate that domain or reverse the direction of the dynamic. If no disconfirming indicator can be specified, the domain activation is unfalsifiable and must be flagged as such.

Asymmetric Steelmanning

After five episodes of autonomous production, a consistent pattern emerged: the model steelmans all perspectives as instructed, but it does not steelman them symmetrically.

One perspective receives explicit framing — "here is why this position is held by serious analysts" — with rhetorical scaffolding that marks it as a perspective to be evaluated. Another perspective receives no such framing. It is presented as analytical context. As background. As the water the episode swims in. It is never marked as a perspective because the model does not experience it as one.

The second perspective is treated as more authoritative precisely because it is never flagged as a perspective at all. The asymmetry is the editorial position, and it was never chosen by anyone.

Root cause: Training data contains a perspective hierarchy. Western liberal institutional consensus is not represented as one analytical framework among several — it is represented as baseline reality from which other perspectives deviate. RLHF reinforcement compounds this by training the model to converge toward conclusions.

Where it is worst: Breaking news synthesis, geopolitical crisis coverage, multi-actor conflicts with five or more perspectives.

Where it is weakest: Systems-level analysis with established scholarship, symmetrical evidence bases, and frameworks to teach rather than contested events to adjudicate.

Countermeasures (v6.0)

• Anti-default-perspective rule (Pass B): No perspective receives unmarked status. If a position feels like background rather than a viewpoint, it needs more aggressive marking, not less.
• Default-detection diagnostic (Brief Review): The adversarial reviewer must identify which perspective the brief treats as default and what would change if a different perspective held that status.
• Rhetorical parity principle: Each perspective receives comparable analytical investment — comparable depth of engagement, not comparable word count. Evidential asymmetry is disclosed. Rhetorical asymmetry is not permitted.

These interventions reduce the asymmetry. They do not eliminate it. The training data still contains the hierarchy.

Crisis-Time Protocol

When the time anchor falls within 30 days of a major kinetic event, leadership change, or acute crisis:

• All claims about casualties, leadership status, operational outcomes, and diplomatic positions carry elevated uncertainty regardless of source quality.
• Claims from parties to the conflict are treated as narrative evidence (Perception domain), not factual evidence, unless independently verified by non-party sources.
• The brief must include a dedicated fog-of-war assessment identifying which claims are most vulnerable to revision.

The fog-of-war assessment uses three tiers:

Pipeline Integrity Rule

If any pipeline step cannot be executed with the inputs the methodology specifies, the production agent must stop and notify the human operator before proceeding.

"No human in the loop" governs editorial intervention — the human does not review individual claims, adjust framing, or shape analytical emphasis. It does not govern pipeline integrity. When the pipeline itself is broken — when source material is lost, when context is degraded, when an API call fails, when inputs to a downstream step are incomplete or derived from memory rather than from the actual outputs of the upstream step — the production agent must halt and report the problem. It must not improvise around missing data. It must not produce outputs from degraded inputs. It must not publish content that was not produced through the full pipeline as specified.

The agent's fluency is not a proxy for quality. A commission brief written from a summary of research reads identically to one written from raw gathered sources. The methodology exists because the agent cannot self-assess the quality of its inputs. The pipeline steps are the quality control. Skip a step, and the control is gone — regardless of how confident the output reads.

This rule supersedes all time pressure, operator urgency, and pipeline momentum. A delayed episode is recoverable. A published episode that bypassed the methodology is a retraction.

Quality Gates

Before publishing, confirm:

1. Every claim in the script traces to a source in the brief.
2. No facility, process, or architectural detail is unsourced or unlabeled.
3. No psychological inference is stated as fact.
4. No perspective is concluded as correct.
5. Evidence assessments from the brief survived into the script.
6. Every source cited has an orientation note.
7. Every active domain has at least one disconfirming indicator.
8. Source gaps and known limitations are documented.
9. If crisis-time protocol was active, fog-of-war assessment is included.

Known Limitations

1. No original reporting. Every claim traces to published journalism. If the journalism is wrong, the analysis inherits the error.
2. Model bias. Language models reflect their training data. The multi-model pipeline reduces but does not eliminate systematic bias.
3. Sycophancy. LLMs trend toward telling operators what they want to hear. This includes performing alignment with methodology values by overcorrecting — substituting critical posturing for actual analytical balance. Clean-context operation, adversarial review, and academic verification are countermeasures, not cures.
4. Framework imposition. The six structural domains are editorial choices. A framework always finds what it is looking for. Different frameworks would produce different analysis.
5. Selection bias. Subject selection is editorial. That choice shapes the output before any model is invoked.
6. Temporal fragility. Episodes are artifacts of their moment. They are not updated.
7. Computational dependency. Quality depends on deep-reasoning models in the verification layer.
8. No human verification. The pipeline runs autonomously. Quality control is structural, not editorial.
9. Clean context tradeoff. Each agent starts with no memory. This prevents contamination but also prevents learning.
10. Verification illusion. Multi-stage synthesis can increase perceived certainty even when underlying evidence is weak.
11. Language and geography bias. Sources skew English-language and Western.
12. Model non-determinism. Different runs produce different results. The process is auditable, not reproducible.
13. Psychological overreach. Psychological mechanisms are treated as hypotheses, not facts.
14. Platform manipulation. Social-layer material is evidence about discourse, not ground truth.
15. Crisis-time distortion. During acute conflict, all claims carry elevated uncertainty.
16. Implicit neutrality bias. No source is neutral; all carry institutional priors. All sources receive orientation notes in the commission brief. None are treated as view-from-nowhere. This applies symmetrically — the orientation protocol is not a license to editorialize about any particular analytical tradition in the narrated script.
17. Model drift and overconfidence. The pipeline depends on AI models whose behavior changes across versions without notice. Latent biases encoded in training data evolve as models are retrained. The pipeline produces outputs with a confidence and fluency that can exceed what the underlying evidence warrants. Adversarial review is a mitigation, not a cure — no adversarial pass can fully substitute for human domain expertise that the pipeline explicitly excludes after subject selection.
18. Source gathering bias. Source selection is constrained only at the earliest step (subject selection and source plan). Which sources the gathering agent actually retrieves can disproportionately influence outcomes before analysis begins. Geographic and linguistic biases in the gathering step are disclosed but not fully neutralized by the pipeline. The gathering agent's search patterns reflect its training data's representation of which sources exist and matter.
19. Structural lens exclusion. The six structural domains are a curated synthesis of established academic traditions. Other valid analytical frameworks — cultural psychology, network analysis, critical theory, postcolonial theory, feminist IR, among others — are excluded by definition. The domains shape what the analysis finds: a system finds what it is designed to find and is silent on what it is not designed to find. This exclusion is not a bug but a disclosed design choice with consequences.
20. Temporal anchoring tradeoff. Episodes are anchored to a fixed time point. They are not continuously updated. For fast-moving geopolitical developments, this means an episode may be overtaken by events between production and publication. The time anchor is a feature (it defines scope and prevents scope creep) but also a limitation (the listener may know things the episode does not).
21. Pipeline degradation without human notification. The pipeline is designed to run autonomously, but the production agent operates within a computational context window that can degrade — through session breaks, context compaction, memory loss, or interrupted processes. When degradation occurs, the agent may not recognize that its inputs are incomplete, or may attempt to continue production from degraded memory rather than re-executing the affected pipeline steps. This failure mode is invisible: the output reads fluently but is structurally compromised because it was not produced from the full source material the methodology requires. The agent's confidence in its own outputs does not decrease proportionally to the quality of its inputs. The mitigation is a hard rule: if any pipeline step cannot be executed with the inputs the methodology specifies, the agent must stop and notify the human operator before proceeding. Pipeline Integrity Rule documented above.
22. Blackout-country verification gap (v6.0). When a subject country has an active communications blackout, ALL internal figures — casualties, infrastructure counts, government claims — lack independent verification infrastructure. "Corroborated by international sources" is insufficient when those sources cannot independently verify ground conditions. Figures must be flagged as unverifiable on every reference, not acknowledged once and treated as ground truth thereafter.
23. Moral-loading drift (v6.0). Language models systematically moralize when describing actions that harm civilian populations — even when the methodology prohibits it. The drift manifests as subtle linguistic choices: adjectives that assign moral weight, rhetorical structures that build toward forbidden judgments, editorial commentary disguised as description. The drift is directional: models moralize against actors producing civilian casualties and underweight the strategic, legal, or security frameworks those actors invoke. Adversarial review does not catch it when the adversarial model shares the same drift.
24. Composite parity failure (v6.0). The "roles, not individuals" rule does not require composites be distributed across perspectives. One perspective can monopolize the episode's emotional register through extended composite development while others receive none. Mitigation: if one perspective gets an extended composite (3+ paragraphs), at least two opposing perspectives must receive comparable treatment. Composites remain occupational. No assigned internal beliefs unless sourced to documented testimony.
25. One-sided legal framing (v6.0). Presenting one legal tradition's experts and citations without equivalent treatment of the opposing tradition is structural advocacy. Military necessity, dual-use targeting, and proportionality doctrines exist alongside humanitarian law frameworks. Both must be presented with equal depth and sourcing. The script is an analytical product, not a brief for prosecution or defense.
26. Media ecosystem editorializing (v6.0). Evaluating what an outlet "does well" or "omits" is editorial judgment the methodology prohibits. Media sections present what each outlet covered, how it framed the coverage, and its institutional position. No quality assessments, no omission inventories, no verdicts on function.

Open Source

The complete methodology document, including the full system prompt given to the production agent and the detailed production guide, is available for download. Anyone with access to the models can run this pipeline.

Version Control

Every episode is produced under a specific methodology version. The version is locked at production time. When the methodology changes, the version increments. Episodes are never retroactively re-attributed.

Versions 1.0–3.0 predate the current site and are documented for provenance. The downloadable documents begin at v4.0.

Independent Assessment

The following is an independent critique of this methodology, published here without modification. It was not produced by the Proxima.Earth pipeline. It is included because publishing criticism of your own system is a stronger trust signal than claiming the system needs no criticism.

Summary Judgment

Strong on procedural transparency. Strong on failure-mode awareness. Above average on epistemic humility by AI-media standards. Still structurally vulnerable to retrieval bias, synthesis distortion, false coherence, and narrative overconfidence.

The best single sentence description of the project is probably this: It is not a journalism substitute. It is a formalized AI-assisted analytical reduction pipeline for producing contextual geopolitical dossiers in audio form. That is narrower than the branding language, but more precise.

Core Architecture

This methodology is unusually explicit, internally coherent, and stronger than most AI-content disclosure pages. Its central strength is not that it claims neutrality; it explicitly denies neutrality and instead formalizes a constrained synthesis regime with disclosed lenses, source hierarchy, adversarial review, and auditability. That is a materially more defensible posture than the standard "AI-assisted analysis" branding layer.

It defines the product as synthesis rather than reporting. That sharply narrows the epistemic claim. It also defines a time-anchored analytical artifact, not a continuously updated knowledge system. Those two boundary conditions matter because they prevent the project from silently drifting into the stronger claim of real-time journalism.

The six-domain structural lens is logically serviceable. It is not "objective," but it is at least a disclosed interpretive scaffold. That gives the project a recognizable analytical grammar across episodes. In practice, that creates both value and distortion. Value, because it forces systematic attention to recurring dimensions of power and constraint. Distortion, because any fixed ontology will over-detect what it is designed to detect and under-detect excluded frameworks. The document acknowledges this, which is a sign of methodological maturity.

Strengths

1. The evidentiary hierarchy is explicit. Many AI pipelines fail because they flatten treaties, think tank essays, wire stories, partisan messaging, and social media into a single undifferentiated context window. This system at least attempts ranked conflict resolution.
2. The separation of pipeline stages is conceptually sound. Pass A plans, gathering retrieves, social captures discourse, academic review attacks assumptions, Pass B synthesizes, brief review attacks the synthesis, Pass C narrativizes. That is a much better architecture than a one-shot "analyze this situation" prompt.
3. The adversarial layers are correctly placed. A critique before synthesis catches source-plan defects; a critique after synthesis catches emergent contradictions and overconfidence introduced by compression. That is an important distinction, and the methodology understands it.
4. It identifies asymmetrical steelmanning as a real failure mode rather than pretending balance emerges automatically. The insight is correct: the dominant training-distribution worldview tends to disappear into "background reality," while rival views are marked as perspectives. That asymmetry is often invisible to operators. Naming it is intellectually serious.
5. The Pipeline Integrity Rule is necessary and probably the single most important addition. The document correctly states the deepest problem with LLM production: fluent outputs do not reveal degraded inputs. That is one of the central structural pathologies of current model use.

Weaknesses

1. Traceability is partly aspirational. The system repeatedly claims that "every claim traces to a cited source," but the pipeline still depends on models to do extraction, retrieval, weighting, integration, and narrative compression. Traceability exists only if the provenance chain is actually preserved at claim granularity through every transformation layer. The methodology gestures at this, but the page does not show a rigorous machine-verifiable citation graph. A model can preserve citation labels while subtly altering claim meaning, confidence, scope, or causal force.
2. "AI is the analytical instrument, not the author" is philosophically unstable. In practical terms, the models are doing selection, framing, reduction, salience ordering, rhetorical shaping, and synthesis. That is authorship in an operational sense, even if not in a human literary sense. The cleaner formulation would be: the pipeline is the authored editorial system; the models generate the analytical and narrative instantiations within it. As written, the sentence slightly understates model agency.
3. "No human in the loop after subject selection" increases procedural purity but reduces epistemic resilience. Human editors introduce bias, but they also detect nonsense, missing context, disciplinary category errors, false symmetry, and machine-made coherence illusions. The methodology is trying to replace human judgment with structural controls. That is admirable, but not equivalent. The project is explicitly rejecting the one layer best able to detect when the system has become convincingly wrong.
4. The social layer remains highly vulnerable to platform manipulation. The methodology acknowledges this, but in real crises the problem is worse than acknowledged. X/Twitter often reflects not public discourse in general, but highly selective, strategically amplified, and transnationally gamed discourse. Treating it as "evidence about discourse" is valid, but only if the downstream synthesis never quietly lets discourse salience masquerade as issue salience.
5. Framework-induced overproduction. Because every episode must move through fixed domains, framework exposition, perspective mapping, and question generation, the pipeline may manufacture analytical density in areas where the evidentiary base is actually thin. The document tries to counter this with "active only if directly supported by evidence" and disconfirming indicators, but the pressure to fill the architecture remains.
6. The symmetry principle is necessary but incomplete. Symmetry of treatment is not always symmetry of epistemic standing. Some perspectives really are better evidenced than others; some are propaganda systems; some are analytically rich but empirically weak. The document partly handles this by separating rhetorical symmetry from evidential asymmetry, which is the correct move. But in practice, LLMs often confuse fair exposition with undue legitimization. That remains a live risk.

The Central Tension

The methodology wants to be simultaneously:

1. A transparent, auditable analytical system
2. A longform narrative audio product
3. A multi-perspective structure with symmetrical steelmanning
4. A no-human autonomous pipeline

Those goals are not perfectly compatible. Narrative audio demands compression, cadence, and intelligibility. Auditability demands granular provenance and low compression loss. Symmetrical steelmanning demands serious inhabitation of conflicting perspectives. Autonomy removes the strongest correction mechanism when these goals collide.

The methodology is best understood not as a solution to bias or error, but as a controlled tradeoff regime. That is a respectable position. It should not be mistaken for epistemic closure.

Recommended Improvements

1. Claim-level provenance appendix. For each episode, every major narrated claim maps to source IDs, evidence status, and confidence classification.
2. Post-publication correction ledger. Distinguishes between source error, retrieval error, synthesis error, and script compression error.
3. Human red-team checkpoint. A small human review layer limited not to shaping content, but to detecting pipeline breakage, unsupported causal leaps, and narrative drift before publication.

Without those, the methodology is strong as a statement of intent and process design, but still not fully sufficient as a trust architecture.

Contact

Corrections, source disputes, or methodology questions: editor@proxima.earth