What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.

What this project reads. This project currently only consumes data from an author's substack. Views expressed elsewhere are not currently integrated into the scoring. Only articles from 2020 and later are evaluated to avoid evaluating stale passages that no longer represent a writer's views. Specifically: a rolling 6-year window measured from whenever the tool is run (not a fixed cutoff at 2020 — that date will move forward each year), and preview-only (paywalled) passages under ~20 words are discarded, since a one-line teaser can't support a real position estimate.

How passages are evaluated. Every crux and vision is compared against a writer's full set of chunked passages using a small local embedding model (BAAI/bge-small-en-v1.5), which evaluates passages by similarity to a given crux/vision and selects the top 50. Claude Haiku then makes a yes/no relevance call on each candidate. Claude Haiku then makes a yes/no relevance call for each passage and evaluates any relevant passages for attribution (is this coming from the author or is the author quoting someone) and how much it supports the crux/vision on a scale of 1-5. Claude Haiku is then asked to give a yes/no on relevance (is it actually talking about the given crux) and attribution (is the author actually saying this or quoting someone), and a 1-5 rating on whether the passages supports or opposes the given crux.

How the average rating is calculated. Final score is calculated by weighted average according to passage recency. Passages written 5 years ago are given half the weight of recent passages.

Specifically: recency = max(0, 1 − age_in_years / 10), and final_score = Σ(recency × passage_score) / Σ(recency).

How confidence is calculated. Confidence combines two things: how much relevant, recent evidence there is, and how consistent that evidence is. More passages generally means more confidence — but if a writer's own passages disagree sharply with each other, that pulls confidence back down even when there are a lot of them. That confidence figure is adjusted to account for the measurement noise found in Haiku's scoring.

Specifically: adjusted_weight = raw_weight / (1 + max(0, author_variance − 0.128)^1.5), where 0.128 is the measurement-noise floor found by repeat-scoring the same passages, and confidence = adjusted_weight / (adjusted_weight + 2). Displayed as low (adjusted_weight ≤ 0.67), unsure (between 0.67 and 6), or high (≥ 6).

Predicted vision: building the belief vector. Each writer's raw 1–5 crux scores are re-centered around the neutral midpoint, so a neutral “3” contributes nothing and stronger views in either direction contribute more — turning the writer's scores into a picture of how strongly, and in which direction, they lean on each crux.

Predicted vision: calculating alignment. The IAPS report maps each of the 9 visions to each of the 8 cruxes as supporting, opposing, or neutral. This project compares a writer's belief vector against the IAPS map by taking a normalized dot product of the two and evaluating which vectors are best aligned. To visualize the uncertainty in this judgement the project resamples the available passages several times to produce a distribution of how consistently a given vision wins across the board. You can view the distribution in the Landing Chart to see how decisively a given view wins relative to other competing visions.

alignment = Σ(centered_score × crux_direction) / (2 × number_of_live_cruxes), displayed on the same 1–5 scale as everything else via displayed_score = 3 + 2 × alignment.

Advocated vision: the four axes. Rather than scoring a writer's real, evidenced position against each of the 9 named visions directly, this project decomposes “advocated vision” into four narrower, independent questions — axes — and locates a writer's position on each one from their own passages. Scoring visions holistically let unrelated questions bleed into each other (a passage arguing against one specific regulatory mechanism isn't by itself evidence for or against a “single dominant project” vision, for instance); splitting into axes keeps those questions separate.

The four axes, and the exact statement each one is scored against:

• Concentration — “AI development should be concentrated in a single dominant actor (company or country), rather than remaining distributed and competitive across many independent actors.”
• Control — “AI development should be controlled and run by government or public institutions, rather than by private companies.”
• Coalition breadth — “AI governance and development should involve a broad, globally inclusive coalition of all major powers, rather than being unilateral or limited to a single nation.”
• Regulatory intensity — “Private AI development should be closely and strictly regulated by government, rather than left to light-touch, minimal regulation.”

Each axis is scored using the same mechanism as a crux (same passage retrieval, same Haiku relevance/attribution/scoring call, same recency-weighted average) but on a −2 to +2 scale centered on zero, where 0 means the writer hasn't taken a clear position rather than that they hold a moderate one.

Advocated vision: matching axis positions to named visions. Each of the 9 IAPS visions is itself mapped onto the same four axes (for example, “Single private project” sits at high concentration / low control, while “Global coordinated regulator” sits at low concentration / high control / high coalition breadth / high regulatory intensity). A writer's advocated alignment with a given vision is the normalized dot product of their 4-axis position against that vision's 4-axis position — the same mechanism as predicted-vision alignment above, just computed over axes instead of cruxes. If a vision doesn't take a position on one of the four axes, that axis is excluded from the comparison for that vision entirely, rather than treated as neutral.

advocated_score = 3 + 2 × [Σ(writer_axis × vision_axis) / (4 × number_of_live_axes)], displayed on the same 1–5 scale as everything else.

Cruxes and axes with no data. If a writer hasn't written enough about a given crux or axis to score at all, that dimension defaults to neutral (3/5, or 0 on the axis scale) with the lowest possible confidence.