sAIfe Hands

Library methodology

Perspective Map Framework

This framework helps readers see where a podcast or article sits on the AI safety spectrum, then find in-site items that counterbalance the same topic with different assumptions.

Back to LibraryTransparent and repeatableEditorial oversight required

1) Position on the spectrum

Each item receives a score from -100 to +100 based on language in title, summary, why-it-matters, editor notes where present, shelf, angle, and tags. When a stored transcript exists and is long enough to form several scored slices, the headline score blends that metadata with the median lexicon score across all sequential transcript segments (full episode walk — see below). The median is the fair cross-item anchor: one radical minute in a two-hour conversation does not silently become “the most extreme item in the library,” but spikes still show up in rationale hints, in the trail’s spread (min / max / percentiles), and in bar height. If the transcript is too short to slice meaningfully, we fall back to metadata plus a stratified excerpt (beginning / middle / end) so signal is not lost. The UI shows “Transcript-informed” when transcript language influenced scoring. We also render a transcript spectrum trail (how segments and percentages are computed). The lexicon weights governance and safety language, and also labor, jobs, and societal-harm framing that listeners often treat as high stakes. The mini-map places the headline score on a left-to-right axis so you can scan many items quickly. Library shelf cards, related picks, counterbalance cards, Signal Room priority rows, and Queue / TED detail pages also show a compact trail whenever a transcript file exists for that slug.

  • - Risk-forward: score <= -18
  • - Mixed: -17 to +17
  • - Opportunity-forward: score >= +18

Plain-language gloss

  • Risk-forward — the piece leans toward warnings, safety, governance, and downside scenarios; it does not mean “wrong,” only that caution and harms are foregrounded.
  • Mixed — risk and opportunity signals are both present without a strong automatic lean; still read the summary for the author’s actual argument.
  • Opportunity-forward — upside, productivity, growth, or efficiency framing is stronger than catastrophic-risk framing; it does not mean “naive,” only that benefits are foregrounded.

2) Primary lens

We classify the dominant analytical lens and up to two supporting lenses.

SocietyGovernanceTechnicalBusinessEnvironment

3) Evidence mode and confidence

We flag whether the item is primarily research, journalistic, or interview-led, then assign confidence based on signal density (tags, lens clarity, and rationale coverage).

  • - High confidence: strong multi-signal consistency
  • - Medium confidence: partial but coherent signals
  • - Low confidence: sparse or ambiguous signals

4) Counterbalance recommendations

For each item, we surface alternatives from the same library that score closer to the mirror of your position on the axis (the point opposite your score), so a strongly risk-forward pick tends to pair with opportunity-leaning material and vice versa — when those items exist in the catalog.

How picks are ranked (transparent rule). For each candidate we compute a distance: start with how far the candidate’s score is from −(your score). If the candidate’s primary lens differs from yours, we add a fixed penalty so same-topic pairs stay analytically comparable when possible. Lower distance wins; we show the top few results.

On item pages, the dual “spectrum vs this page” strip plots you and each counterbalance pick on the same bar so you can see the contrast visually (for example, a caution-first TED talk vs a more upside-leaning interview).

Related conversations are different: they follow editorial links or the same display shelf for continuity, not the mirror-distance rule. Their mini-maps show where each neighbor sits so you can tell editorial “what’s next” from algorithmic counterbalance.

Goal: help users compare assumptions, not force false equivalence.

Transcript spectrum trail: how it is calculated

The trail visualises the same Risk-forward / Mixed / Opportunity-forward scale as the headline score, but applied to ordered slices of the episode transcript so you can see where language leans early vs late. It does not replace the headline score; it adds a time-ordered shape when transcript text is available.

Representative vs outlier (design choice)

Cross-library comparison needs a stable headline: listeners care about the typical framing of an episode, not a single sentence cherry-picked for virality. So the headline blends metadata with the median slice score after scanning the full transcript. Outliers still matter editorially: we surface unusually extreme slices in written rationale when they diverge sharply from that median, and the trail summary reports mean, median, min–max, and p10–p90 so you can see concentration vs tails at a glance.

  1. Input. We use the stored plain-text transcript (whitespace normalised into a single stream).
  2. Chunking. The text is split into sequential segments. Each segment targets about 1,100 characters, extending or shortening slightly to break at sentence boundaries (full stop, question mark, exclamation, or newline) when that yields a chunk of at least 72 characters. The implementation walks the entire transcript in order until the text ends. A very high safety ceiling (thousands of slices) exists only to guard against pathological files; ordinary episodes never hit it. Segment count still does not map linearly to clock minutes — it tracks text mass and chunking.
  3. Per-segment score. Each segment is scored with the same weighted lexicon rules as the overall perspective profile (precaution-oriented terms pull toward risk-forward; opportunity- or acceleration-oriented terms pull toward opportunity-forward). Scores are clamped to −100 to +100.
  4. Per-segment band (aligned with the headline bands): Risk-forward if score ≤ −18; Mixed if −17 to +17; Opportunity-forward if score ≥ +18.
  5. Order vs width. Each segment records where its midpoint sits in the full transcript (0 = start, 1 = end). Bars are drawn left to right in that order. In the UI, bar columns share the row evenly across the full width—column width is not proportional to minutes of audio or to the character length of that slice.
  6. Height and colour. Bar height maps to spectrum position for that segment; colour maps to its band (Risk-forward, Mixed, Opportunity-forward).
  7. Summary percentages. The line that reports “Across n segments” and percentages for risk-forward, mixed, and opportunity-forward is computed from the fraction of segments in each band (simple counts), not from wall-clock duration or word-count weighting.
  8. Compact views. Shelf cards and some inline trails downsample (evenly spaced picks along the segment list) so roughly 18–20 bars stay readable; full detail panels show every scored segment from the full pass (very dense episodes can produce a wide bar field; the summary line still reports exact segment counts and distribution stats).

Headline score with transcript. Current blend: roughly 38% metadata lexicon score + 62% transcript-wide median slice score (both on the −100…+100 scale), then the usual band thresholds. Weights are code constants so they can be tuned as we gather feedback.

Perspective map UI. When enough slices exist, the map adds two readouts under the headline marker: a lean strip (amber / slate / cyan) = what fraction of slices landed risk-forward, mixed, or opportunity-forward; and a spread strip on the same −100…+100 axis — shaded band = middle 80% of slices (p10–p90), amber and cyan ticks = single strongest risk- and opportunity-leaning slices, white bar = median typical slice, violet dashed line = headline blend (metadata + transcript) so you can see when the catalog pulled the headline away from the transcript median. On item pages, tap those controls to open the full transcript and scroll near that moment (library / queue / TED). From grids or Signal Room, the same taps navigate to the item with a #t= deep link so the transcript opens in the right place after load.

Transcripts and assessments

For YouTube-backed entries, maintainers can fetch publisher captions into content/resources/transcripts/<slug>.json using npm run resources:fetch-transcripts (supports --slug, --limit, --force). Item pages then show an excerpt and use transcript language in perspective scoring when present. Counterbalance picks use each neighbor’s metadata-only score, but the anchor uses your transcript (when present) for mirror distance so the pairing reacts to what you are actually listening to on this page.

Optional listen-based write-ups live in content/resources/transcript-assessments/<slug>.json. Generate a starter file with npm run resources:transcript-assessment-template -- --slug <slug>, then replace the placeholder summary after editorial review.

Editorial guardrails

  • - This is decision support, not an objective truth label.
  • - Scores should be reviewed when source metadata changes.
  • - Edge cases require manual override by editorial leads.
  • - Users should read across at least two counterbalanced items before concluding.