How the Fates Work
How the Fates Work
The Fates is a strategic-foresight instrument. Three agents, named for the Greek sisters of destiny, divide the labor of seeing early: one gathers, one measures, one tells.
The method
Atropos — the scanner. She sweeps eleven classes of source — academic literature (peer-reviewed and labeled preprints), patents (US and worldwide), legal and regulatory filings, dockets, and legislative bills, research funding and federal awards, clinical-trial registrations, corporate securities disclosures, industry reports, institutional news, global press, expert commentary, and social chatter — for evidence published 2016 to present. A scan returns 150–250 validated evidence cards per topic; everything else is cut, and every cut is logged with a reason.
Lachesis — the measurer. She embeds every card, weaves them into threads, and scores each thread on the STEEPLE dimensions and the Three Horizons — including the difference between H2⁻ (innovation absorbed to prop up the old system) and H2⁺ (innovation that builds the next one). A thread needs ≥5 signals to reach the loom; a thread drawn from a single source class is admitted but capped at low certainty and held at posture “track” until a second independent class corroborates it.
Clotho — the teller. She spins threads into essays where every claim carries a receipt — a numbered citation to a real, linkable source. Her illustrations are AI-generated and labeled as such.
The integrity contract
- Zero invented sources. Every signal carries a real URL and a capture date.
- Every rejection is logged. The exclusion log records what was cut and why — newswire spam, SEO bait, out-of-window, duplicates.
- Chatter is labeled. Social sources are included for early warning but always marked low-trust.
- AI is labeled. Machine-written analysis and machine-made images say so.
The observatory right now
Latest scan: 2026-10-30 · 13 active topics · figures queried live from the database at 2026-10-30 00:40 UTC — never hardcoded.
Source classes & trust tiers
| Source class | Trust tier | Cards held |
|---|---|---|
| academic | high — preprints labeled | 2,864 |
| patent | high | 124 |
| legal | high | 773 |
| funding | high | 478 |
| registry | high | 137 |
| corporate disclosure | high — on the record | 142 |
| industry report | high | 99 |
| institutional news | high/medium | 158 |
| press | medium — global wires | 157 |
| expert blog | medium | 527 |
| social chatter | low — always labeled | 1,364 |
The dial: modes, budgets, and the quality score
Members choose scan depth (50 to 500 accepted cards) and one of three modes. Quality is a selection problem, not a collection problem: precision requires having something to reject, and every rejection is logged with a reason.
| Mode | Behavior |
|---|---|
| Broad | mechanical filters only; class caps scaled to target; chatter included; no rank cut; maximum surface area |
| Balanced (default) | class budgets apply; the quality score breaks ties when a class exceeds its budget |
| Strict | oversample 3.5x, triage, classify, then rank-select to target; chatter is compared to tripwires but never accepted; recency tightened |
Intake budgets are class-level rather than per-source, because thread diffusion scores source-class breadth as evidence about the world: flat uniform intake would manufacture breadth by policy and corrupt the measurement. Class caps keep flooding impossible while letting classes differ. No single connector may fill more than 60% of its class budget. This table regenerates from live configuration, never hardcoded values.
| Class | Budget at default (200) |
|---|---|
| legal | 25 |
| press | 30 |
| scout | 10 |
| chatter | 10 |
| funding | 45 |
| patents | 25 |
| academic | 40 |
| registry | 30 |
| corporate disclosure | 15 |
Every classified card carries a composite quality score, components stored for audit. Objective components always outweigh model judgment (0.60 to 0.40): the classifier grading its own signal strength is useful but self-referential, and corroboration, class trust, and recency are the ballast.
| Component | Weight | Nature |
|---|---|---|
| corroboration | 0.25 | objective |
| source-class trust | 0.2 | objective |
| recency (18-month half-life) | 0.15 | objective |
| classifier signal strength | 0.2 | model judgment |
| specificity | 0.05 | model judgment |
The contrarian pin. Contrarian-flagged evidence is exempt from rank and budget cuts, capped at 15% of target. A quality filter optimizing for corroborated, high-authority, consensus signals would systematically delete dissent, and surfacing dissent is the reason a foresight platform exists. Without this rule, Strict mode is a consensus machine with better sourcing.
Estimative language
Every likelihood word in analytical prose (analyses, digests, register entries, engine outputs, and essay interludes) draws from a fixed lexicon, so two readers never hear different numbers in the same word. Narrative fiction is exempt: a character's voice is not an estimate.
| Term | Band |
|---|---|
| remote | 0% to 10% |
| very unlikely | 2% to 10% |
| unlikely | 10% to 35% |
| roughly even | 35% to 55% |
| likely | 55% to 80% |
| very likely | 80% to 95% |
| almost certain | 95% to 100% |
Indicator feeds
Quantitative series from FRED, BLS, and Census stand beside the evidence as ballast, never inside it as filler: observations are never evidence cards. Series bind to threads only after human approval of an agent proposal, power numeric tripwires, and feed Trend Impact Analysis. Currently 6 series tracked, 6 bound to threads; counts regenerate from live data.
Weekly Ritual scans run each topic's saved configuration sequentially through a single-scan lock; the comfortable ceiling is roughly 10 topics before the morning digest slips. The mitigation path (earlier start, then per-topic locks) is documented, not yet built.
This page is regenerated from live data after every weekly pulse.