Personal intelligence · Source trust

The Cascade Ledger: Influence Is a Receipt, Not a Reputation

Influence cannot be hand-whitelisted or read off follower counts. It is domain-specific, role-specific and time-varying. A radar must learn it: AI judges what each edge means; deterministic graph analysis computes influence from the resulting cascades — receipts, not reputation.

Scott Farrell · LeverageAI · Design essay companion to the Cascade Ledger ebook

Everyone maintains a VIP list. Almost nobody’s list survives contact with a fast-moving field. You pin Andrej Karpathy, a handful of lab accounts, a few well-known coding-agent practitioners, maybe a YouTuber who “seems early.” Within a month the field reorders, an obscure engineer was right three times, and the famous account mostly reposted. The list still looks like taste. It is reputation cosplay.

The question a personal intelligence system actually needs to answer is sharper:

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

The answer is a cascade ledger: a source graph with typed edges, role labels, and domain-specific influence cards computed from observed cascades. Influence is not declared. It is earned as dated receipts.

The broken proxies

Two defaults dominate how people decide whom to follow.

Hand whitelist. Seed the obvious names. Scrape them forever. Promote friends of friends when you notice them. This fails for two reasons you already feel: you do not know everyone who matters, and the people who matter next month are often not on the list yet. Hand-maintained VIP lists go inaccurate under load; the maintenance tax is boredom; the failure mode is missing the early engineer while babysitting fame.

Follower count (and its cousins: likes, stars, “engagement”). Social-listening research has spent years showing that follower-heavy detection methods mis-rank inactive or weakly connected audiences, and that interaction structure matters more than indegree alone.¹ Marketing research has similarly found that common quantitative vanity metrics are poor proxies for professional judgment of content quality, and that follower count and engagement do not rise together forever.²

So the system cannot outsource trust to status. It needs a different object: not “is this person famous?” but “what job has this source repeatedly performed, in which domain, with what cascade ancestry and what later confirmation?”

AI types edges; math counts cascades

The clean division of labour is the whole product thesis in one line.

Language models are good at fuzzy semantic judgments:

• 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?

Once those judgments become typed edges on a graph, 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

That split is safer than oracular scoring. Models invent plausible rankings under prompt pressure. Graphs under-count when edges are missing — and missing edges are auditable.

Useful edge types include:

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

Primary / secondary / tertiary should describe role in the information chain, not moral value. A tertiary aggregator can be the most valuable discovery instrument you own once the system knows what job it is good at.

Cascade root is not inventor

Here is the load-bearing public example — and the discipline that makes the ledger honest.

In the contemporary AI-agent community many people follow, a substantial share of the modern “agents should maintain wikis” cascade appears to trace through Andrej Karpathy. Replies, quote-posts, Reddit threads, YouTube discussions, repositories and later engineering references often share that ancestry. In that local world, he sits near the top of the tree. People reach for words like godfather or grandfather.

He did not invent wikis. He was not the first person to propose agent memory, knowledge graphs, or living documentation. Those intellectual lineages predate the tweet.

Both facts belong in the graph:

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

Role vocabulary that keeps the ledger precise:

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. Idea provenance is settled through dated receipts, not retrospective storytelling — the same instinct as authority-by-receipt inside organisations, now applied to the open ecosystem.^8,9

In our world, all agent-wiki roads lead back to Karpathy — and he invented none of it. Both facts belong in the graph.

A worked cascade (schematic metrics)

Once edges exist, the ledger can show ancestry rather than vibe. A schematic cascade — the shape you would harvest from bronze observations over a bootstrap window — looks like this:

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

From that tree, deterministic metrics might compute something like:

The raw count is not the final meaning. The graph supplies a factual basis for a careful claim: this source sits unusually high in the propagation ancestry of this concept in the discourse neighbourhood you actually follow.

That is influence as receipt: inspectable, domain-bound, revisable when later edges arrive.

Four lanes, not one score

A single “influence” number reintroduces the VIP list in numeric costume. Keep separable lanes:

Domain cards make the point concrete:

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

A YouTube aggregator may be low on truth authority and high on discovery. That is not a contradiction. It is a job description. Aggregators are sensors, not authorities: they discover people and stories you did not know to follow, sense propagation, and reveal how communities interpret events. Their coverage is a weak prior — in the same sense that volume, trajectory and participant diversity should perturb judgment rather than command it.^10,11

Copied amplification is the discipline that keeps lanes honest. 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.

Messengers are not subjects

A second common failure: the system treats whoever surfaced a story as the story.

Separate ledgers (or at least separate edge roles) for:

• the Reddit user who first brought it into your collectors
• the person who originated the claim or release
• the person who popularised it for a community
• the implementer who later proved it consequential

Discovery order is not causal order. A case can open on a secondary post and later promote a primary anchor without erasing the messenger’s discovery receipt. (How the queue holds that case is the sibling problem — the Signal-Case Queue — not this article’s job.¹²)

Primary / secondary / tertiary then stay role labels:

Value lives in the lanes. Role lives in the chain position. Mixing them is how VIP lists get reborn as “tier scores.”

Precedents: Techmeme and Brandwatch

None of this requires inventing social graph science from scratch. Two public systems already prove partial versions.

Techmeme still combines crawlers, algorithms and human editors into an editorial pyramid.³ More importantly for this design, its source universe and author ranking 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.⁴

That is extremely close to the bootstrap mechanism you want:

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

Techmeme produces a shared industry view. Your cascade ledger produces a marketplace-of-one view against a personal wiki — but the seed-and-link learning pattern is a surviving public precedent.

Brandwatch (via PeerIndex heritage and later products) explicitly treated influence as something discoverable from interaction graphs: an “Influence Graph” scoring large populations of Twitter users from how often they engage and are engaged with, rather than relying only on a fixed roster.⁵ 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.⁶

Conclusion both precedents support:

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

Sensors sit on a diffusion chain

GitHub trending lists, “Awesome” digests and star velocity are tertiary sensors. They are not primary truth, and they are often late.

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. It means the sensor is measuring implementation uptake, not idea birth. Star counts especially must never be oracles: large-scale measurement work has found millions of suspected fake GitHub stars and tens of thousands of repositories involved in fake-star campaigns.⁷ Star velocity belongs as one prior beside contributor quality, forks, commits, independent mentions and longevity — Nudge Doctrine again: many weak priors, one judge.¹⁰

The same posture applies to YouTubers and news aggregators. Keep them. Instrument them. Never let them silently become the truth authority lane.

The promotion story (unknown source → high influence)

This is the test that VIP lists cannot pass.

Imagine an engineer with few followers. Over a 90-day bootstrap window the ledger records:

1. Three times she posts a technical observation days before lab 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.
4. Downstream: two independent communities pick up her framing; one repository cites her thread as motivation.

Receipts the system can write without human heroics:

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

Meanwhile a famous account may keep high propagation value — when they post, the ecosystem moves — while truth authority stays lower if they mostly amplify. Demotion is not exile. It is correct job assignment: propagation sensor, not oracle.

An unknown engineer becomes important automatically after repeatedly being early and correct. That is authority-by-receipt moved from organisational tenure fights to the open feed.^8,9

The feedback risk

Influence learned only from cases the system itself judged important will compound its early mistakes. Mis-attach a cascade, overweight a famous amplifier, suppress an alien-but-true signal — and next week’s priors defend last week’s error.

Defences that belong in the ledger design:

• Lane separation so one bad truth call does not erase discovery value
• Outcome backfills when later confirmations or retractions arrive
• Sampled audits of suppressed material so the system measures what it missed
• Independent-branch accounting so copied lineage cannot farm confirmation

The ledger is not a scoreboard. It is an accounting system. Accounting systems need reconciliations.

What you can build after this

Related work stays in its lane. The Signal-Case Queue owns how unfinished understanding stays alive as cases and when to re-observe. Idea Provenance and receipts-replace-tenure own organisational credit and authority-by-receipt inside a firm. Newsjacking and routing sit on top of influence scores; they are not this artefact. Design-evolution and generative-pattern siblings will use cascade evidence as one benchmark among others — name them when relevant, build them elsewhere.

The north star is simple enough to remember and strict enough to implement:

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