You open the spreadsheet again. Forty accounts. A few lab feeds. Andrej Karpathy. A well-known coding-agent practitioner. One YouTuber who “seemed early.” It looks like taste. It feels like curation. You tell yourself this is how professionals stay informed: pick the right people, watch them, and the field will arrive at your door already sorted.

Two weeks later the field has reordered. An engineer with almost no followers was right three times about tooling failure modes — each time days ahead of the lab commentary you actually scraped. The famous account mostly amplified other people’s posts; when the ecosystem moved after one of their retweets, you mis-read the RT as origin. The YouTuber found two stories you would have missed and botched the technical read on both. The spreadsheet has not changed. That is the tell.

The list is reputation cosplay. It is not a sensing system.

The question the list cannot answer

A personal intelligence system — a radar that watches a fast field so you can decide, write and act without drowning — eventually hits a harder question than “what arrived?”

How does my system work out whom to listen to — and how much — without me maintaining a VIP list that goes stale and wrong?

Hand-maintained lists fail for reasons you already feel. You do not know everyone who matters. The people who will matter next month are often not on the list yet. And maintaining the list by hand is a special kind of death: boring enough that you stop, inaccurate enough that you should never have trusted it. The maintenance tax is not a calendar reminder; it is the silent decision to stop updating once the novelty of curation wears off.

We think the blunt line is correct:

We can’t hand whitelist it. It’ll go nuts and be inaccurate.

That is not a product preference. It is an architectural requirement. Influence has to be learned — domain by domain, role by role, as dated receipts rather than as status that someone declared once in a cell.

What people say versus what they ship

People say they follow the right people. What they ship is usually one of two defaults.

The hand whitelist. Seed the obvious names. Scrape them forever. Add friends of friends when you notice them in a thread. The failure modes are structural: unknowns never enter without heroics, and fame never demotes without politics. Even a diligent curator is limited by what they already know how to notice. The list is a map of the curator’s past attention, not of the field’s present causality.

The follower proxy. Rank by followers, likes, stars, “engagement.” Social-listening research has spent years showing that follower-heavy methods mis-rank inactive or weakly connected audiences, and that interaction structure matters more than raw indegree.¹ Marketing research has similarly found that common quantitative vanity metrics are poor proxies for professional judgment of content quality.² Status is not a receipt. A large audience can be a propagation sensor, a brand, a ghost town of inactive followers, or an amplifier that never originates. The number does not say which.

Both defaults optimise for administrative neatness — a fixed set of sources, a sortable number — rather than for epistemic trust. They answer “who is easy to rank?” when the real question is “who repeatedly earned which job in which domain?”

The cost of the wrong object

When influence is a list, three bad things compound.

You miss the early engineer while babysitting fame. The obscure practitioner who posts a failure mode four days before the lab write-up never appears on your seed. Your collectors never saw them. Your briefing never named them. Next month the same engineer is early again — still invisible to a constitution that only trusts cells you typed in January.

You treat amplifiers as originators because they are louder. A famous account reposts a thread. Your system records the famous name. The originator gets a footnote if they get anything. Months later the ledger (or your memory) has rewritten history around the megaphone.

You treat ten near-identical videos as ten independent confirmations because volume feels like consensus. One rumour script, ten faces, one lineage. Without collapse, your truth prior rises with the clone count. Amplifiers farm trust they never earned.

None of those failures looks dramatic on day one. All of them quietly train your attention on the wrong people. By the time you notice the spreadsheet is wrong, the radar has already spent weeks reinforcing the error.

The enemy is not “having seeds.” Seeds are fine. A bootstrap list of tens of sources is how every graph starts. The enemy is treating the seed list as a constitution rather than a bootstrap — freezing day-zero taste into permanent authority.

What this book owns

This ebook codifies the Cascade Ledger: a source and influence graph where AI types semantic edges, deterministic software computes cascade consequences, and sources earn domain-specific influence as dated receipts — not as reputation.

After these chapters you should be able to:

• refuse global influencer scores in favour of domain cards and role labels;
• type cascade edges instead of asking a model “how influential is X?”;
• hold cascade-root and non-inventor as simultaneous true claims;
• compute metrics from a worked cascade tree;
• separate messengers from subjects;
• place aggregators and GitHub on a diffusion chain as sensors;
• promote an unknown early source from receipts alone;
• audit the feedback loop so the ledger does not only reinforce itself.

What this book deliberately does not own: queue grain, re-observation schedules and expected-silence mechanics (that is the Signal-Case Queue). Organisational credit politics live in Idea Provenance. Newsjacking and routing sit on top of influence scores; they are not this artefact. Name the siblings. Do not smuggle them in as chapters.

Influence cannot be hand-whitelisted or read off follower counts, because it is domain-specific, role-specific and time-varying.

Chapter 2 installs the mechanism that replaces the list: AI types the edges; math counts the cascade. Everything that follows — lanes, roles, the Karpathy dual fact, the worked tree, promotion, audits — is a consequence of that split. Stop maintaining the VIP list. Start learning the receipts.

The temptation is obvious. You have a frontier model. You have a person page. You ask: How influential is Karpathy on agent architecture? The model returns a confident paragraph and a number that feels decisive. You can even wrap it as a microservice and call it a product feature.

That number is almost worthless as architecture. Models invent plausible rankings under prompt pressure³. The score is not auditable⁴. Domain collapses into vibes. Tomorrow’s re-ask drifts. You have recreated the VIP list inside a completion — only now the constitution is opaque and the maintenance tax is API spend.

The correct split is stricter and more boring — which is why it works.

AI decides the meaning and edge type. Deterministic software calculates the graph consequences.

What the model is for

Language models earn their keep on fuzzy semantic judgments that graphs cannot invent from raw text alone:

• Are these two posts discussing the same idea?
• Is this repository implementing the idea, or only borrowing vocabulary?
• Is this independent convergence, or copied amplification?
• What role did this person play — originator, populariser, amplifier, implementer, critic?

Each judgment should become a typed edge with a date, a reason, an evidence pointer into bronze observations, and a confidence. The model proposes. Contested high-impact edges — cascade-root claims, inventor-adjacent claims, independence collapses — can take a senior pass. The point is not “AI decides influence.” The point is “AI decides what the relationship is.”

That distinction is load-bearing. Influence is a graph property once the relationships exist. Asking the model for the property skips the graph. Typing the relationship builds the graph. Everything chapter 5 harvests as metrics depends on this discipline holding.

What software is for

Once edges exist, influence calculation should leave the model. Deterministic analysis can compute, for each person, source or post:

• unique downstream authors and communities;
• cross-platform spread;
• cascade depth and breadth;
• implementation descendants and high-authority descendants;
• time from source to downstream reaction;
• independent branches versus copied lineage;
• later confirmations, retractions and failures.

Models invent rankings. Graphs under-count when edges are missing. Under-count is auditable: you can see which bronze objects have no typed link yet. Hallucinated rank is not. Prefer the under-count. Prefer the missing-edge ticket over the confident score.

The harvest is also revisable without drama. New edges arrive; metrics recompute as-at a window. You do not re-ask a model to “update its feelings about Karpathy.” You re-run the job over a denser graph.

An edge catalogue that earns its keep

introduced-before · independently-converged · popularised
quoted · replied-to · cited · implemented-by
amplified-by · reported-by · criticised-by · superseded-by

You do not need an ISO standard on day one. You need a small vocabulary that forces the system to say what kind of relationship it thinks it saw. “Related” is how VIP lists smuggle themselves back in. “Mentions” is almost as bad if it never distinguishes quote from critique from implementation.

Each edge type earns different weight in different lanes. implemented-by is strong for consequence and uptake. amplified-by is strong for propagation and weak for originality. independently-converged is strong for confirmation-style counting. quoted and replied-to show discourse activity without proving truth. The catalogue is not decoration; it is the contract that keeps math from lying.

Eight factors, zero oracles

Influence evidence accumulates along several axes. None of them should be allowed to become a single score that rules alone:

1. Temporal precedence — who said or built it before the others?
2. Direct attribution — who explicitly links to or names whom?
3. Cascade depth — replies only, or articles, videos, repos and systems?
4. Independent spread — did it jump into separate communities?
5. Downstream quality — were consequential engineers and organisations involved?
6. Historical track record — has this person repeatedly sat near roots of later-important signals?
7. Domain consistency — influential in this subject, not merely famous?
8. Originality versus amplification — creating, interpreting, or distributing?

That is the same posture as many weak priors informing one judgment: engagement, upvotes and aggregator chatter can matter without becoming truth. A source can score high on cascade depth and low on originality. Another can score high on historical track record in one domain and near-zero in another. The factors are inputs to a domain card, not votes for a celebrity number.

A tiny worked fragment

Two posts mention “agent wiki.” The model proposes: post B refers-to post A; post A popularised the concept. Math updates: A gains popularisation-root evidence; B is messenger or propagation, not origin. No global score was requested. The graph got denser. That is progress.

Stretch the fragment one step further. A repository appears with a README that names post A. The model proposes implemented-by with confidence medium and an evidence pointer to the README paragraph. Contested? Maybe. A senior pass confirms implementation, not vocabulary tourism. Math updates: A gains an implementation descendant; the repository author gains an implementer receipt. Still no oracle score. Still a denser, walkable graph.

That is the whole mechanism in miniature: semantic typing first, deterministic consequences second, influence as something you can recompute and audit rather than something a completion invents under pressure.

Chapter 3 adds the other half of the structure: influence is not one number even after the cascade is counted. It separates into lanes and domain cards.

Picture a dashboard. A face. A big number: Influence 94. It is satisfying. It is also a lie waiting to happen in four directions at once: wrong domain, wrong role, wrong job, wrong time. The score collapses four different questions into one celebrity number, then pretends the collapse was insight.

The cascade ledger refuses the global score. It keeps separable lanes and domain cards. After edges are typed and cascades are counted (Chapter 2), the result is still not one rank. It is a multi-layer map of what job a source repeatedly performs, for which subject, as of which window.

The four value lanes

A source can be high on one lane and low on another without contradiction. That is a job description, not a moral failure. A famous amplifier can move ecosystems when they post (propagation high) while almost never being the first true technical claim (truth low). An obscure engineer can reverse that pattern. A YouTube channel can find stories early (discovery high) while mangling the technical read (truth low, interpretive mixed).

Lane separation is also a safety valve. A bad truth call should not erase discovery value. A strong propagation sensor should not inherit origin credit. Mixing the lanes is how VIP lists get reborn as tier scores.

Domain cards, not celebrity ranks

Influence is domain-specific. The same person can be load-bearing in one neighbourhood and a tourist in another. Cards make that concrete instead of hoping a single number will remember the distinction.

Karpathy
  model interfaces & learning: very high
  agent architecture: high
  enterprise governance: indirect / unknown

Well-known coding-agent practitioner
  coding-agent practice: high
  AI company strategy: medium
  foundational model research: low / unknown

Unknown YouTuber
  model-release rumours: medium, improving
  technical interpretation: low
  early discovery yield: unusually high

Karpathy’s card is allowed to be very high on interfaces and only indirect on enterprise governance. The practitioner’s card is allowed to be high on coding-agent practice without becoming a foundational-model oracle. The YouTuber is allowed to be an excellent discovery instrument with poor technical accuracy — once the system knows which lane it is using.

Walk the YouTuber card for a moment. Suppose three early rumours later confirmed by first-party posts. Discovery lane rises. Suppose two technical explanations that independent engineers later correct. Truth lane stays low or falls. Suppose the framing of “what developers will actually build next” gets reused in two independent communities. Interpretive lane can rise even while truth stays weak. None of that requires a global score of 37. All of it requires the system to stop asking for one.

Primary, secondary, tertiary name roles — not value

Tertiary is not an insult. GitHub Awesome-style digests and many YouTube channels are tertiary discovery and momentum sensors. They should not establish historical originality. They can still be the most valuable instruments you own for finding what you did not know to follow. Value lives in the lanes. Role lives in the chain position. Mixing them is how VIP lists get reborn as “tier scores.”

Aggregators as sensors (preview)

Aggregators perform at least three jobs: discovery, propagation sensing, and interpretation sensing. Their coverage is a weak signal. Volume, trajectory, participant diversity and independence should combine into a nomination; chatter should perturb judgment, not command it.

Ten near-identical videos copied from one rumour count roughly as one lineage, not ten confirmations. Three technically independent communities discovering the same effect count far more. Chapter 8 expands the diffusion-chain placement; the rule starts here so lanes stay honest. Without lineage collapse, the truth lane becomes a popularity contest among clones.

A small scenario

An operator bans YouTubers as noise. Over a month they miss three early tooling rumours later confirmed by first-party posts. The correct fix is not “believe YouTubers.” It is: keep them on the discovery lane, never promote them to truth authority without confirmation receipts, and collapse copied scripts into lineages.

A second operator does the opposite: treats every high-engagement channel as truth. They ship two false technical claims into a weekly brief because volume looked like consensus. The correct fix is the same architecture with the opposite emphasis: propagation and discovery can be high while truth stays gated on later confirmation.

Domain cards also change how operators schedule attention. A high discovery, low truth YouTuber is watched for early nominations, not for technical brief copy. A high propagation famous account is watched for ecosystem activation, not for first-claims. A high truth, medium propagation engineer is watched for confirmation-grade assertions even when the room is quiet. The dashboard still has faces. It no longer has one lying number.

That is multi-layer attention without a gold badge. Influence is domain, role and lane together. Promote someone to “gold interest” if you must — but only if the card still says which gold, for which job, as-at which window.

Lanes without role typing still invent inventors from fame. Chapter 4 is the dual fact that makes the ledger honest: cascade root is not inventor.

People reach for family words when a cascade is obvious. Godfather. Grandfather. The instinct is trying to name something real: a person sits near the top of a propagation tree that keeps showing up in replies, Reddit threads, videos, repositories and later engineering posts. The tree is loud enough that operators start treating the person as the origin of the idea itself.

The instinct becomes a lie the moment it quietly upgrades populariser into inventor.

In our world, they all lead back to him — and he didn’t invent them. Both facts belong in the graph.

The observation, carefully scoped

In the contemporary AI-agent neighbourhood many people follow, a substantial share of the modern “agents should maintain wikis” cascade appears to trace through Andrej Karpathy⁵. That is a claim about local cascade ancestry, not about the entire history of knowledge systems. Scope it: the discourse neighbourhood you actually follow, the bootstrap window you actually ingested, the concept cluster you actually care about.

What he need not be — and is not, for this ledger:

• inventor of wikis;
• first person to propose agent memory;
• first person to use a knowledge graph;
• historical origin of the whole intellectual lineage.

Earlier lineages include wiki systems, knowledge graphs and agent-memory research. Those edges exist whether or not a current discourse routes through one popular formulation. A compact, promptable design kernel can be highly transmissible without being historically first. Popularisation is a real job. Invention is a different job. Conflating them is how fame rewrites provenance.

Role vocabulary that keeps the ledger honest

originated · preceded · independently-converged
popularised · translated-for-community
amplified · implemented · commercialised · criticised

Human-facing labels can stay vivid. Machine labels should stay boring and typed:

cascade-root-for
popularisation-root-for
propagation-ancestor-of

For the current agent-wiki cascade, a defensible card might look like:

Earlier intellectual lineage:
  wiki systems · knowledge graphs · agent memory research

Karpathy’s roles in the current cascade:
  canonical populariser
  propagation ancestor
  current-cascade root

Evidence:
  large downstream citation and implementation cascade

Read that card slowly. Canonical populariser means the formulation many later artefacts cite or orbit. Propagation ancestor means many independent downstream paths still share that node. Current-cascade root means, in the window and neighbourhood you measured, ancestry metrics peak there. None of those labels says originated. Preceded edges can point to older work without demoting the populariser’s propagation receipts. Independently-converged edges can mark parallel invention without inventing a false single parent.

Why the distinction is load-bearing

Without it, fame rewrites history. Inventors get erased when a better communicator arrives. Communicators get falsely crowned as originators because the cascade is loud. Attention systems then learn the wrong people for the wrong jobs: they promote megaphones into truth oracles and bury early technical work that never sat at the loud root.

With it, influence for an attention system can be high while historical originality is modest. That is not insult. That is precision. Idea provenance should be settled through dated receipts rather than retrospective storytelling — the same authority-by-receipt instinct that replaces tenure fights inside organisations, now applied to the open ecosystem.

Notice the last row. Even a fully supported cascade-root claim does not license domain bleed. Propagation ancestry in agent tooling is not a certificate for enterprise governance truth. Role typing and domain cards work together: the dual fact is settled per concept and per neighbourhood, not as a global halo.

The dual fact is also a training rule for operators. When someone says “Karpathy invented agent wikis in our world,” the ledger answer is: he sits at the root of the cascade we follow; he did not invent the lineage. When someone says “therefore ignore him because he is not original,” the ledger answer is: popularisation and cascade ancestry are still influence for an attention system. Both wrong compressions erase useful structure. The graph keeps both facts because both facts change what you do next: how you weight origin research versus how you weight current propagation.

Repeated root-sitting strengthens the case. A single viral post can be luck. Sitting near the top of multiple later-important cascades in the same domain is historical track record — one of the eight factors from Chapter 2 — without ever becoming a global influencer crown. The ledger can harvest that repetition deterministically once edges exist. We can just give it a number after the graph exists; we do not invent the number instead of the graph.

Chapter 5 attaches numbers to the tree so the dual fact is not only philosophical — it is measurable as cascade ancestry. The role vocabulary becomes a set of edges; the metrics become the receipts. Until those numbers exist, “grandfather” is still only a vivid human label. After they exist, cascade-root is a claim you can walk.

Philosophy is cheap. Ancestry is not. Once edges exist, the ledger can show a tree. Chapter 4 settled the dual fact in words: cascade root without inventor. This chapter attaches numbers so the claim becomes walkable.

The numbers below are a schematic bootstrap harvest — design evidence of the shape you compute after typing edges over bronze observations, not a claim that these exact counts were scraped live on a particular day. Label your metrics the same way in production: as-at a window, over a neighbourhood, given a set of edges. Schematics are honest when they wear their labels.

The tree

post.karpathy-wiki-tweet
 ├── 37 direct replies
 ├── 14 quote-post discussions
 ├── 8 Reddit threads
 ├── 11 YouTube discussions
 ├── 23 repositories
 └── 4 later first-party engineering references

Read the tree as multi-platform ancestry, not as a vanity counter. Replies are one layer. Quote-posts are another discourse surface. Reddit threads are community interpretation. YouTube discussions are tertiary amplification and framing. Repositories are implementation uptake. First-party engineering references are high-authority descendants that should be weighted carefully, not blindly maximised. The cascade is interesting because it spans stages, not because any single number is large.

Typed edges on that tree

Examples of edges the AI layer might propose, with reasons and observation IDs behind each:

• person.karpathy authored post
• post popularised concept.agent-maintained-wiki
• Reddit thread refers-to post
• repository implements concept; repository influenced-by post
• YouTube video amplified post (then check independence)
• later engineering article cited post or independently-converged

Walk one path step by step. Bronze stores bronze.twitter.post.123. AI proposes authored-by person.karpathy with high confidence from the account metadata. Separately, AI proposes popularised concept.agent-maintained-wiki with a reason pointer to the formulation language and a confidence that may warrant review. A Reddit observation arrives later: bronze.reddit.post.456. AI proposes refers-to post.123 because the thread quotes or links the tweet. A repository README names the same post; AI proposes implemented-by after checking that the code realises the idea rather than borrowing a slogan. Each edge is dated. Each edge points at evidence. None of them is a global influence score.

Independence typing is the hard edge. Two repositories that both implement agent-maintained wikis might share a parent (influenced-by post.123) or might be independently-converged on the concept from older wiki lineage. The model proposes; software does not invent the distinction; senior review can settle contested cases. Without that typing, implementation counts inflate confirmation.

Deterministic metrics

How the metrics are computed, not merely displayed: unique downstream authors is a set-count over edge endpoints after de-duplication of accounts. Cascade depth is the longest typed path from the candidate root through discourse to implementation (or another terminal stage you care about). Cross-platform spread is the set of platforms represented in the descendant set. Implementation descendants count nodes with implemented-by (or equivalent) edges that survive independence checks. High-authority descendants are a curated or rule-based subset — first-party engineering blogs, lab posts — never a raw celebrity filter. Independent versus copied is a clustering problem upstream of confirmation-style counting. Role conclusion is not a metric average; it is a label supported by the metric bundle plus explicit non-claims (no inventor edge, earlier lineage present).

What the numbers must not claim

• global “most influential person in AI”;
• historical invention of wikis or agent memory;
• truth of every downstream implementation;
• that repository star counts are honest popularity (Chapter 8).

The careful claim is enough. You do not need a global throne. You need a neighbourhood-bound receipt that tells the radar how to weight this source for this concept, as-at this window.

Independence accounting in practice

Suppose eleven YouTube discussions appear under the tree. Cluster by transcript and title similarity and shared primary source. Seven collapse into lineage L1 — one rumour script, many faces. Four remain independent interpretive branches. Confirmation-style counting uses the independent set, not the raw video count. Without that collapse, amplifiers farm truth.

Apply the same discipline to repositories. Twenty-three repos is not twenty-three independent inventions. Some are forks. Some are template clones. Some are genuine independent implementations that still cite the popularisation root. Software can count stars; only typed edges plus clustering can tell you whether you have one lineage or many. Prefer under-count when independence is unclear.

From bronze to metrics

1. Immutable observations: posts, threads, repos, videos.
2. AI proposes edges, roles, reasons, evidence pointers.
3. Light senior review on contested high-impact edges.
4. Deterministic metric job runs per domain.
5. Domain cards update as-at the computation date.

The raw count is not the final meaning. The graph supplies a factual basis for role conclusions and lane updates. Influence becomes a receipt you can walk, not a reputation you must take on faith. Bootstrap windows of thirty to ninety days are enough to start; the important architectural point is that historical ingestion builds the graph and the metrics recompute when edges densify.

Even a perfect tree still fails if the system confuses who surfaced the story with who originated it. Chapter 6 separates messengers from subjects.

Reddit arrives first: “Karpathy says agents should maintain wikis.” A broken system crowns the Reddit user as the story. A ledger that understands roles writes a discovery receipt for the messenger, keeps the observation in bronze, and later promotes the canonical subject without erasing who brought the signal in.

That sequence is not a corner case. In fast fields, secondary surfaces often outrun primary posts into your collectors. The failure mode is treating first-seen as first-caused. The fix is role separation, not faster scraping alone.

Messengers are not subjects.

Four roles people constantly collapse

• the person who surfaced the story into your collectors (messenger);
• the person who originated the claim or release (originator);
• the person who popularised it for a community (populariser);
• the person who implemented it and proved it consequential (implementer).

Those are different jobs. They deserve different ledgers — or at least different edge roles on the same graph. Mixing them is how VIP lists get reborn as “first seen author scores.” A messenger can be excellent at discovery and useless as a truth authority. An implementer can prove consequence without ever having been early. A populariser can sit at the cascade root without having originated the intellectual lineage (Chapter 4). Collapsing the four into one score is the same error as Influence 94 wearing a friendlier hat.

Influence receipts on a Reddit-first arrival

Step 1 — Messenger earns discovery

Observation: reddit.post.456. Receipt: the Reddit user discovered or reported the signal in this domain. If that pattern repeats, their discovery lane rises. They are not automatically a truth authority. Bronze keeps the post immutable: the comments, the criticism, the links to older work, the people who entered the cascade through that thread.

Step 2 — Subject and populariser edges appear

Later ingestion finds twitter.post.123 authored by Karpathy. Edges: Reddit refers-to tweet; tweet popularised concept. Karpathy becomes a candidate for populariser and current-cascade root. The Reddit user remains messenger. Discovery was real; it is not cancelled by a better anchor. Canonicality can settle for the current cascade without rewriting who opened the door into your collectors.

Step 3 — Implementer later

A repository implements the concept and is influenced-by the tweet. The implementer ledger records proof of consequence. Implementation uptake is a different stage of the diffusion chain (Chapter 8) and a different kind of authority. The implementer may never have been early in discourse. That does not make them less important for consequence; it makes them important for a different lane.

A fourth role often appears late and gets stolen credit early: the amplifier. A famous account reposts after the cascade is already moving. Propagation value can be very high. Origin credit must stay low. Without messenger/subject separation, amplifiers quietly inherit the discovery and origin slots because they are the loudest node on the day you look.

What this chapter does not own

How a queue holds a stable signal case while anchors move — discovery anchor versus canonical anchor, bronze immutability, revisit targets — belongs to the Signal-Case Queue. Name it. Fence it. This book only needs the influence consequence: different roles write different receipts, and discovery order is never causal order.

Why the separation matters for learning

Without it, Reddit power-users become false truth authorities. Primary authors lose credit when a secondary arrives first. Famous amplifiers collect origin points they never earned⁶. With it, you can promote discovery specialists as discovery specialists — and still know who the cascade actually rooted on.

The learning loop depends on this honesty. If messengers are scored as subjects, later influence cards train the radar to watch the wrong people. If implementers are scored as originators, consequence is mis-read as invention. If amplifiers are scored as popularisers without independence checks, copied lineage farms the truth lane. Role separation is not etiquette. It is how the ledger stays a learning system rather than a fame amplifier with better fonts.

Discovery order versus causal order also disciplines collectors. If Reddit arrives first, the case can open on a secondary surface without crowning it primary forever. If the tweet arrives first, the Reddit thread is still not noise: it may hold stronger technical criticism, older links and new people entering the cascade. Either arrival order is fine. Role receipts absorb the order. The ledger never needs to pretend that first-seen equals first-caused.

That is why “kick out the secondary” is such a destructive instinct. Kicking out erases bronze, erases propagation evidence and erases the messenger’s discovery credit. Demoting role keeps the evidence and corrects the hierarchy. The Signal-Case Queue holds the case identity while anchors move; the cascade ledger holds the role consequences while people earn different receipts for different jobs.

Public systems already prove partial seed-and-learn without a personal wiki. Chapter 7 is the precedent chapter: Techmeme and Brandwatch.

You do not need to invent social-graph science from zero. Two public systems already show that source universes can grow from observed linking and conversation rather than from a frozen VIP constitution. That matters because the cascade ledger is ambitious enough without also claiming to invent influence graphs. Borrow the proven bootstrap. Add the epistemic machinery those systems were not built to hold.

Do not maintain a giant VIP whitelist. Seed the graph, then let observed influence promote people.

Techmeme: links as promotion machinery

Techmeme combines crawlers, algorithms and human editors into what it describes as an editorial pyramid — people and software together, not a pure closed list.⁷

More important for this design: ranking and source centrality have long depended on observed linking behaviour. Topic leaderboards rank authors from inbound links by peers and influential industry accounts, combined with topical publishing activity — well-cited and prolific on the topic, not merely frequently featured by editorial taste alone.⁸

That maps almost directly onto the bootstrap you want:

small seed set
    ↓
observe who links to whom
    ↓
learn communities and their centres
    ↓
promote repeatedly consequential sources

Notice what the pattern does and does not claim. It does claim that centrality can be learned from observed linking rather than declared once. It does claim that topical activity matters alongside citation. It does not claim that inbound links settle truth authority, role typing, or domain cards. Techmeme still produces a shared industry view optimised for a broad tech audience. Your cascade ledger produces a marketplace-of-one view against a personal wiki. The seed-and-link learning pattern is the surviving public precedent; the personal grounding and role machinery are the differentiators.

Concretely: if a modest seed of AI sources repeatedly links to an obscure practitioner on agent tooling, the graph should notice. If the same practitioner is later early on confirmed claims, lanes should move. No human needs to schedule a spreadsheet update. That is Techmeme’s spirit applied to personal epistemic trust rather than industry headline centrality.

Brandwatch: influence from the conversation graph

Brandwatch’s lineage includes an explicit Influence Graph: a large live author database scoring influence from how often people engage and are engaged with — interaction structure, not roster membership alone.⁹

Contemporary Brandwatch positioning still frames social listening as a way to discover creators and rising niche influencers from real conversations, not only from pre-built lists.¹⁰

That supports the same conclusion: discover creators from the graph of what is happening, then promote what repeatedly earns attention in the domains you care about. Interaction structure is closer to cascade evidence than follower count is. It still is not a substitute for typed edges, independence accounting, or truth-versus-propagation separation. Marketing influencer databases optimise for reach and campaign fit. Personal intelligence systems optimise for justified attention and justified silence.

What the precedents do not give you

• diff against a personal canon and active projects;
• role typing (cascade-root versus inventor);
• four value lanes as first-class objects;
• messenger/subject separation for epistemic trust;
• deterministic cascade metrics with copied-lineage collapse.

So borrow the seed-then-learn pattern. Do not stop at marketing influencer databases. The cascade ledger is not “Brandwatch for one person.” It is influence as receipts for a radar that already holds a wiki, cases and bronze observations.

Bootstrap recipe

1. Seed a modest set of obvious sources plus aggregators as sensors (design guidance: tens of sources, not thousands on day one).
2. Collect bronze observations for a bootstrap window (thirty to ninety days is enough to start).
3. Type edges with AI; senior-review contested high-impact roles.
4. Run cascade metrics per domain.
5. Expand the seed from promoted unknowns; demote pure amplifiers on the truth lane while keeping them as propagation sensors.
6. Repeat. The constitution is the update rule, not the initial list.

That worked path is the existence proof for Chapter 9’s full promotion story. Chapter 7 only needs the bootstrap claim: seed, observe, promote — and do not freeze the seed into a constitution.

The bootstrap also has a tempo. Thirty to ninety days of bronze is enough to start seeing centres and early unknowns. That does not mean the seed freezes at day ninety. It means the first promotion wave can run without waiting for a perfect historical corpus. Historical ingestion builds the graph; live observation densifies it. The update rule remains the constitution even after the first wave of promotions looks “good enough.” Good enough is when the next unknown still has a path in without a human noticing her first.

Differentiation stays sharp. Industry systems optimise shared maps: what tech is talking about, who is central in public conversation. The cascade ledger optimises a personal map: what changes relative to your wiki, which sources earn which lanes for your domains, and when fame should be demoted to a sensor job. Same seed-then-learn skeleton. Different object of trust.

Once the graph exists, sensors still need stage labels. GitHub is often late. Aggregators are often loud. Chapter 8 places them on the diffusion chain.

GitHub trending lights up two weeks after the idea was already live on X and Reddit. An operator treats stars as “now it’s real.” The ledger treats GitHub as an implementation-uptake stage, not a birth certificate. That single placement decision prevents a whole class of origin mistakes without throwing away a useful sensor.

Aggregators are sensors, not authorities.

Three jobs aggregators actually perform

1. Discovery — find people, repositories, papers and stories you did not know to follow.
2. Propagation sensing — coverage patterns reveal that an idea is becoming active.
3. Interpretation sensing — the angles different aggregators choose reveal what communities think the event means.

Coverage is a weak signal. Volume, trajectory, participant diversity and source independence should combine into a nomination; chatter should perturb the judgment, not command it.

YouTubers sit in the same family. They can be focused on discovery and on shaping what developers will care about without ever becoming primary truth. Keep the channel. Instrument the lane. Never auto-promote to truth authority without confirmation receipts.

Copied lineage collapse

Ten near-identical videos cloned from one rumour count roughly as one lineage, not ten confirmations. Three technically independent communities discovering the same effect count far more. Without collapse, amplifiers farm the truth lane. With it, discovery sensors stay useful without becoming fake corroboration engines.

Collapse is not optional polish. It is the difference between a cascade ledger and a volume dashboard. Chapter 5 already applied the rule to the schematic YouTube branch; Chapter 8 makes it a standing law for every sensor stage. Cluster by transcript, title, shared primary source and timing. Prefer under-count when independence is murky. Independent convergence is a claim that needs edges, not a default award for every similar post.

The diffusion chain

research / private experiments
        ↓
influential public formulation
        ↓
technical discussion and early prototypes
        ↓
GitHub projects
        ↓
star growth, forks, derivative projects
        ↓
mainstream aggregation

Late does not mean unimportant. Late means the sensor is measuring a different stage. GitHub often measures implementation uptake, not idea birth. Mainstream aggregation often measures narrative consolidation, not first truth. Private experiments may never appear in your collectors at all; absence at the top of the chain is not evidence that the idea began at the first public post you saw.

Place each sensor on the stage it actually measures. An X quote-post discussion is early discourse. A Reddit technical thread may be early community interpretation. A repository is mid-to-late uptake. An Awesome digest is tertiary discovery of what already exists. A Techmeme-style link cluster is industry narrative consolidation. Wrong stage placement is how late stars get mistaken for early truth.

GitHub sensors without star worship

GitHub Awesome-style digests, trending lists and velocity trackers are tertiary discovery and momentum instruments. They can indicate that developers have started building, that a concept is crossing from discourse into implementation, or that a previously obscure team deserves investigation.

They must not establish that an idea is historically original or that every star is a human vote of confidence. Large-scale measurement work has found on the order of six million suspected fake GitHub stars and tens of thousands of repositories involved in fake-star campaigns.¹¹ Star velocity belongs as one prior beside contributor quality, forks, commits, independent mentions and longevity — never as an oracle.

The star-fraud caveat is not a footnote for security nerds. It is proof that no single metric may rule the ledger. If stars can be bought, star-led promotion can be bought. The cascade ledger already has the right posture: many weak priors, one judge; independent mentions and implementation quality matter more than a velocity spike alone.

YouTubers as instruments

A common card: low technical accuracy, unusually high early-discovery yield, sometimes strong interpretive value for what developers will care about. Keep the channel. Instrument the lane. Never auto-promote to truth authority without confirmation receipts. The same card can include high propagation value when a channel regularly marks the moment a story leaves the specialist niche and enters the broader developer conversation.

Place the cascade tree from Chapter 5 on this chain and the stages clarify. Replies and quote-posts are early discourse. Reddit threads are community interpretation. YouTube discussions are tertiary amplification and framing. Repositories are implementation uptake. First-party engineering references are high-authority descendants that may arrive late and still matter. Star growth on those repositories is a further, noisier uptake signal — useful as a prior, fatal as an oracle.

Operators often want one stage to mean “real.” The ledger refuses. Real for discovery is not real for truth is not real for uptake. GitHub arriving two weeks late can still be the moment an idea becomes consequential in code. That is worth knowing. It is not a birth certificate for the idea, and it is not a certificate that every star is honest.

Sensors and metrics make automatic promotion possible. Chapter 9 is the story VIP lists cannot pass: the unknown engineer who promotes herself.

This is the test VIP lists cannot pass. If your system only trusts names you already typed into a spreadsheet, unknowns never enter without heroics and fame never demotes without politics. The cascade ledger exists so neither ceremony is required.

The story is not a motivational poster. It is a worked promotion path with receipt fields, lane updates and a paired demotion. If the ledger cannot tell this story automatically, it is still a VIP list wearing graph cosmetics.

An unknown engineer can become important automatically after repeatedly being early and correct.

The promotion story (schematic, labelled as design evidence)

Over a ninety-day bootstrap window the ledger records:

1. Three times she posts a technical observation days before lab or famous commentary.
2. Twice the observation is later confirmed by a first-party post or a high-quality independent replication.
3. Zero major retractions; one partial correction she issues herself (a reliability signal, not a stain).
4. Downstream: two independent communities reuse her framing; one repository cites her thread as motivation.

Tell it as an arc, not a bullet list. Day 12: she posts a failure mode about agent tooling under a real load. The seed list is silent. An aggregator surfaces her post into bronze; she is not on the VIP sheet. Discovery lane gets a small early tick with low confidence. Day 19: a lab engineer confirms the same failure mode in a first-party note. Truth lane moves. The messenger who found her (if any) keeps a discovery receipt; she keeps origin credit for the domain-local claim. Day 41: she is early again on a related tooling claim. Day 48: independent replication. Day 67: a third early call, this one only partially right; she corrects herself within hours. Day 74: a repository README cites her thread. Day 88: two independent communities reuse her framing language. Promotion threshold for domain seed expansion is crossed. No human scheduled a meeting about her. The update rule fired.

The receipt card

person.unknown_engineer
  domain: agent_tooling
  lanes:
    truth_authority: rising (2/3 early claims later confirmed)
    discovery_value: high (median lead +4 days vs seed list)
    propagation_value: medium (not a megaphone)
    interpretive_value: high (framing reused)
  roles_observed:
    - originator (domain-local)
    - early_source
  promotion:
    elevate review priority
    include in domain seed expansion
  evidence: [edge ids…]
  as_at: bootstrap_window_end

Every field is a receipt. Truth authority is not a vibe; it is a confirmation ratio over early claims. Discovery is lead time versus the seed list, not follower growth. Propagation is honestly medium because she is not a megaphone. Interpretive value rises because framing was reused independently. Roles stay domain-local: originator here is not inventor of agent tooling as a field. Promotion actions are operational, not ceremonial: review priority and seed expansion, not a gold badge that bleeds across domains.

The paired demotion

person.famous_amplifier
  domain: agent_tooling
  lanes:
    truth_authority: low-medium (mostly amplifies)
    discovery_value: low (late)
    propagation_value: very high (ecosystem moves when they post)
    interpretive_value: medium
  roles_observed:
    - amplifier
    - propagation_sensor
  demotion:
    remove from truth-prior seed
    keep as propagation sensor

Demotion is not exile. It is correct job assignment. A famous account may remain highly useful as a propagation sensor while receiving a much lower truth prior¹. That is the ledger working, not the system being rude. When they repost on day 62, the event is recorded as propagation, not as origin. The unknown engineer’s earlier post keeps its temporal precedence.

Timeline shape

Failure modes of promotion

• Gaming: strategic early noise plus loud self-marketing — require later confirmation, not early volume alone.
• Domain bleed: tooling excellence is not strategy authority; cards stay domain-scoped.
• Short windows: one lucky early call is not a rising truth authority.

Promotion without confirmation is how rumour farms enter the seed. Confirmation without independence accounting is how clone clusters invent consensus. Domain bleed is how a strong tooling card quietly becomes a fake strategy oracle. Short windows are how lucky noise looks like genius. The defences are already in the architecture: lanes, edges, independence collapse, as-at versioning.

This is authority-by-receipt moved from organisational tenure fights to the open feed. Fame and trust are allowed to diverge. That is the point.

Automatic importance without celebrity is the product outcome operators can feel. The briefing starts naming a person who never appeared on the original sheet. The famous amplifier still appears, but as a propagation event: “ecosystem activated” rather than “source asserted.” Review priority shifts toward the early confirmed engineer for tooling. None of that requires a taste committee. It requires edges, confirmations, independence accounting and domain cards that are allowed to diverge from fame.

If the system still needs a human to “add her to the list” after the receipts are obvious, the constitution never left the spreadsheet. The promotion is the update rule firing. The demotion is the same rule assigning the correct sensor job. Both are receipts, not reputation politics.

There is one remaining risk: influence is learned from cases the system itself judged important. Early errors compound. Chapter 10 is the audit layer and the build contract.

Here is the uncomfortable loop. The radar learns whom to trust from cascades attached to cases it decided were important. Mis-attach a cascade, overweight a famous amplifier, suppress an alien-but-true signal — and next week’s priors defend last week’s error.

Without audits, the cascade ledger becomes a self-licking ice cream cone of its own mistakes. Promotion (Chapter 9) only stays honest if the learning loop can correct itself. The ledger is not a scoreboard. It is an accounting system. Accounting systems need reconciliations.

The feedback risk

• Training only on retained or “important” cases creates selection bias.¹²
• Fluent wrong attachments contaminate cascade ancestry and promote the wrong people.
• A dense personal canon can become a suppression shield around unfamiliar fields unless you deliberately sample outside it.

Walk a concrete contamination path. Day 1: the system attaches a secondary amplifier as cascade root because the model mistyped a quote edge as origin. Day 2: metrics promote the amplifier on truth-adjacent lanes. Day 3: review priority follows the wrong person. Day 14: an early engineer is deprioritised because the false root already “owns” the case. Day 30: the error is now a prior. Without outcome backfills and contested-edge review, the ledger digs in. The prose explanations still look fluent. Fluency is not causality.

Personal grounding is a second trap. Diffing against a dense wiki is power, but it can also reinforce the existing map. An alien-but-true signal with no wiki intersection can be treated as noise. Without sampled audits of suppressed material, the system only measures what it chose to keep. Silence becomes unexamined, not justified.

Defences that belong in the design

1. Lane separation — a bad truth call does not erase discovery value.
2. Outcome backfills — later confirmations and retractions rewrite cards as-at.
3. Sampled audits of suppressed material — measure what you missed, not only what you kept.
4. Independent-branch accounting — copied lineage cannot farm confirmation.
5. Contested-edge review — senior pass on high-impact role claims (cascade-root, inventor-adjacent).
6. as-at versioning — influence cards are dated, not eternal.

Many weak priors, one judge; no single engagement metric becomes an oracle. Trajectory and independence still perturb rather than command.

Outcome backfills deserve operational detail. When a claim is later confirmed, truth-lane receipts should rewrite as-at the confirmation date with evidence pointers. When a claim is retracted, cards should fall — including cards for people who amplified the false claim as if it were origin. When independence clustering improves, confirmation counts should recompute, not keep the inflated clone number forever. as-at versioning means you can still answer “what did we believe on day 40?” without pretending day-40 cards are eternal truth.

Suppressed-case audits need a budget. Sample material the system deprioritised or never attached. Ask: was anything later consequential? If yes, write a miss receipt and adjust priors. If no, the silence was justified — and that is also a result worth recording. Justified silence is an output of a healthy radar; unexamined silence is a feedback hole.

Cascade Ledger build checklist

Sibling map (keep the lanes clean)

The wiki-graph substrate that holds these edges is the same family of compiled relationships as the index-is-the-data work; the ledger is the influence programme on top of that substrate.

What a healthy audit looks like

Once a quarter — or after any major policy change — sample a slice of suppressed material and a slice of high-impact role claims. For suppressions, ask whether anything later became consequential outside the radar. For role claims, ask whether cascade-root and inventor-adjacent edges still have evidence pointers a senior reviewer would accept. For promotions, ask whether confirmation ratios and independence collapses still support the card. Write miss receipts and correction edges when they do not. Recompute as-at. Do not silently edit history without a dated reason.

The goal is not infinite paranoia. The goal is a learning loop that can notice when it trained on its own mistakes. Influence learned from self-judged importance will always tend to compound; audits are how you keep the compound interest from becoming compound error.

The north star

Influence is a receipt, not a reputation. Type the edges. Count the cascade. Promote the early. Demote fame to the job it actually does.

When the unknown engineer outranks the celebrity on truth authority — and the celebrity still ranks high on propagation — the architecture is working. When a cascade root is not mistaken for an inventor, the dual fact is working. When a suppressed alien signal is occasionally audited and found empty, silence is working. Build the ledger, not the list. Audit the ledger you learn from — or the list will return wearing better clothes.