EigenRep, Sector Reputation: 3-Version Comparison

Index , the report is chronological; early sections are kept as history. For the June recommendation jump to §14 (all of it superseded by the 2026-07-22 rebuild, see banner).

Three states of the per-sector reputation, for the same 12 accounts:

• OLD, the 2026-05-28 run (archived today).
• NEW, the 2026-06-03 live run (10× deeper S1 crawl; 6.75M accounts).
• IMPROVED, the NEW scores with two fixes applied (edge-weighting + whitening, explained in plain terms just below). These change sector scores only, the global/overall score is unchanged. (Historical: §8 later showed the whitening half breaks at full-population scale. Kept as the baseline for §3's heatmaps; the surviving lineage is edge-weighting → §9 → §11 → §13 → §14.)

How the versions evolved after this comparison was written: the production v2 export (2026-06-08, §8) shipped edge-weighting + λ-whitening, its whitened _score broke at population scale; v2 minus whitening (§9) kept edge-weighting + log-raw; Approach 2 (§11–§13) added the famous-source down-weight and became the §13 final; v2.1 (§14, current) upgrades the edge weight with a purity factor (strength × purity) and un-pins the scale's top band. Every later variant runs on the same NEW graph, only the scoring changes. Full map with statuses: the index above.

Metrics (lower = healthier sectors):

• sector↔global correlation, how much a sector score is just "how central are you overall." High = sectors are echoes of the global rank.
• cross-sector correlation, how much the 10 sectors move together (e.g. fintech↔space). High = no real differentiation.
• % top-1% in all 10 sectors, accounts ranked elite in every sector at once (impossible if real). The flagship symptom.

The data behind each version

The graph grew sharply between runs; the seed set did not change. IMPROVED is computed on the NEW graph, so it shares NEW's size, only the sector scores are recomputed.

Seeds per sector , verified identical across OLD and NEW (same 318 seed handles & per-sector counts; the seed set was frozen between these runs, only the crawl grew)

* master = sector-agnostic core anchors (used only in the global run). 318 unique s0 seeds total, some belong to multiple sectors, so the row sums to more than 318. These hand-picked anchors are the only "ground truth" the whole system is built on, which is why seed quality matters so much. Verified: OLD and NEW share the exact same 318 seeds (0 differences). An earlier pre-promotion run (2026-04-23, deep archive) used a smaller 233-seed set under a ≤2,000-following cutoff, the jump to 318 happened before both versions shown here.

The proposed fixes, in plain terms

All tentative, none are in production. They split into two kinds of change, which matters a lot for cost and risk:

• Edge-weighting (fix 1) and the later famous-source down-weight (fix 4) = algorithm changes. They alter how trust propagates (the trust matrix itself), so they require a full re-run of the EigenTrust pipeline. They fix the scores at the source, but they're the heavy lift (pipeline owner, recompute).
• Whitening · residual · log-raw, and the later un-pinned band (fix 5) = pure post-hoc transforms. Just arithmetic on the already-computed scores (the published `scores.csv`), no re-run, no graph, runs in seconds, ships immediately. They re-describe the output; they don't change what the algorithm computed.

All three above target correlation ("high in every sector"). A separate, fourth improvement, log-raw scaling (also a post-hoc transform), fixes a different symptom (the top saturating, "everything above 950"); see §5. outcome: survived, backbone of the final scale (§12–§14)

Two layers that joined the stack later (after this section was written, they replaced whitening/residual as the decorrelation layer)

Green = healthier (more decorrelated), red = more polluted, but a low correlation alone is not health. The dimmed whitening-family rows crush correlation by destroying the reputation signal itself: §8 showed that at full-population scale they rank zero-reputation junk above real authorities, so their impressive-looking numbers are quoted for history, not as candidates. Edge-weighting is the half that survived (it decorrelates without breaking reliability) and evolved into the §13/§14 line. The two bottom rows are the variants that replaced them. Their correlations are higher than whitening's, no reputation information is destroyed buying the number down, but the flagship pathology, accounts top-1% in all 10 sectors, falls 0.26% → 0.12% (§13) → 0.02% (§14 v2.1, 13× cleaner than live, and the ~1,300 accounts that remain are the genuinely-elite-everywhere hubs). On the un-pinned scale the score-level sector↔global reaches 0.40 (§14 corr table).

The three fixes in isolation, each applied to NEW on its own

These are three separate, independent fixes, each shown alone (the combined IMPROVED above = edge-weight + whitening). Global is unchanged in all (they only touch sectors). Darker = higher. Compare the textures to see what each one does.

① Edge-weight only , algorithm change · needs a pipeline re-run

② Whitening only , post-hoc transform · no re-run

③ Residual only , post-hoc transform · no re-run

The first two fixes attack correlation ("high in every sector"). This third, separate one attacks the symptom you flagged, "so many accounts above 950", which isn't a reputation problem but a scaling one. It's an alternative scale for the published number; the §3 heatmaps all use percentile/rank scaling, and log-raw is shown here on its own, not folded into them.

Tested by downsampling the current crawl to lower coverage and measuring edge-weighting's decorrelation (space). Can't test beyond the current ~43% directly, that needs new crawl data for the uncrawled S1, so this shows the slope and where it's heading.

Read: two things improve as crawl rises, the baseline correlation falls on its own (0.995 → 0.725; more routing nodes let trust differentiate sectors), and edge-weighting's benefit accelerates (−0.001 → −0.099, the bars). The curve is still steepening at 43%, not flattening → pushing S1 crawl toward 100% should keep helping.

Caveats: the uncrawled ~69K S1 may be less central than this random downsample implies (real gain at 100% could be smaller); crawling S1 also reveals more S2 sinks; and the 6.6M uncrawled S2 remains the larger untapped population. The deeper crawl looked bad for raw scores but is a prerequisite and ongoing amplifier for the fix.

• The deeper crawl (OLD→NEW) didn't fix sectors, it worsened them. Coverage rose 6×, but the flat-generalist pathology rose 8× (0.03% → 0.26%).
• Edge-weighting is necessary but not sufficient, it cuts sector↔global (0.699→0.608) but barely touches cross-sector bleed alone.
• The output layer is the cure. Edge-weight + whitening ≈ solves it (0.065 / 0.138 / 0% flat); edge-weight + residual kills sector↔global (0.072) but leaves more cross-sector bleed (0.433).
• Whitening's catch: it can assign generic accounts spurious specialties (Pavel→quantum). So: whitening for analysis (max decorrelation), residual for the displayed score (interpretable, no fake specialties).
• Two symptoms, two kinds of fix: "high in every sector" = correlation → edge-weighting + residual; "everything above 950" = scale saturation → log-raw scaling. They're independent and stack.
• Recommended stack: edge-weighting (pipeline re-run) → residual (decorrelate output, post-hoc) → log-raw scaling (de-saturate the scale, post-hoc). All content-free; the sector fixes are validated across all 10 sectors (space is representative, not an outlier).
• And keep crawling S1, it's a prerequisite *and* an ongoing amplifier for the edge-weighting fix (§6); the benefit was still climbing at 43% coverage.

High-impact AI accounts, where each version puts them

The six most unambiguous AI figures alive, plus two zero-reputation accounts for contrast. Percentile columns (green header) rank them correctly, all pinned at the top. The whitened score columns (red header) invert it: the titans read mid (≈400–530) while the train company maxes at 1000.

Read: Sam Altman is ai_ml_percentile = 999 in every version (correct, he's a top-0.1% AI account), but his ai_ml_score is only ~419, ranking him top-35% of all 6.75M on that column. @ouigo_es, a train operator with global_score 52, gets the maxed ai_ml_score 1000. λ (0.1→0.8) nudges the titans' score up a few points but never fixes the inversion. The percentile is the only column that behaves; the whitened score must not be ranked on.

The 12 sample accounts under v2 (λ=0.5), the same split, on familiar faces

PERCENTILE (λ0.5) , the usable rank; note it lands ≈ NEW (§3), i.e. v2's percentile reverts to ~v1 behavior, still prominence-correlated

WHITENED SCORE (λ0.5) , the broken column; same accounts, a wholly different (and much lower-magnitude) picture

Same 12 accounts, same run, two columns that don't agree. The percentile block is dark (high rank, like NEW); the score block is muted and reshuffled. That gap is the whitening, and on these reputable accounts it's merely confusing; on the full population it surfaces the junk above.

Root cause & the fix

• Why it breaks: "remove the prominence factor" was implemented as a ratio (sector ÷ overall). Ratios explode when the denominator is tiny, a no-name whose entire micro-footprint is one AI edge looks "100% AI-pure" → 1000, while a broad authority's AI share is modest. It measures purity, not reputation.
• A reputation floor alone doesn't save it: even restricted to the global top-1%, the whitened score still ranks obscure niche accounts above Altman, because purity ≠ reputation. Gating removes the global_score-52 noise but not the inversion.
• The fix, one number per sector, computed as a floored magnitude residual: residual = sector_raw − predicted(sector_raw | global_raw), kept in raw magnitude units (additive, not a ratio), with a floor that zeroes the low-signal tail. A tiny account then has a tiny residual and cannot float up; a genuine authority's real sector excess survives. Rank that one residual → no score-vs-percentile split to disagree. λ becomes the residual strength, now safe to turn up.
• Reconciling with this report: §1–7 validated whitening on the curated 12, all of whom have real reputation, so the ratio's blow-up never showed. v2 applied it to all 6.75M unfloored. Whitening isn't wrong; it needs the reputation gate + magnitude form above.

Decision: ship λ=0.5 using the percentile as the sector score and drop/quarantine the whitened _score (coherent product today); have the v2 author recompute the single sector number as a floored magnitude residual to actually reduce the cross-sector bleed without the purity artifact.

§8's recommendation made concrete: keep edge-weighting (sector reputation) and log-raw (de-saturation), drop the whitening. The v2 export omitted the sector raw, so this is a fresh edge-weighted run on the live graph, its global matches v2 bit-close (sama 1000, @pavelprata 998, @uselegion 792), so the sectors are comparable.

AI sector, the §8 inversion disappears

Dropping whitening flips §8's defect the right way round: the titans return to the top of AI (1000), the junk accounts collapse (global 88/59, off-sectors near 0, @ouigo_es AI falls 1000→738→450). log-raw additionally de-saturates: the share scoring ≥950 drops 5.0% → ~0.2% in every sector.

EDGE-WEIGHT + PERCENTILE (no whitening), 12 sample rank scale · compare to §3 NEW: hubs still high, but @uselegion's noise (ai 478, robotics 411) is gone

EDGE-WEIGHT + LOG-RAW (no whitening), 12 sample de-saturated magnitude · same accounts, a real gradient (no 999-pileup)

③ The fix, edge-weight + floored magnitude residual

Now the decorrelation done right: residual = log10(sector_raw) − [ a·log10(global_raw) + b ], "how far above the prominence-predicted baseline is your sector reputation," in log-magnitude (additive, not a ratio), computed only for accounts above a reputation floor (global pct ≥ 500; keeps 3.37M), ranked among them. Recomputed live (v13).

EDGE-WEIGHT + FLOORED RESIDUAL, 12 sample sharp, real specialties surface; junk floored out

Compare to the §3 NEW block (near-solid dark): the residual carves out genuine specialties, @pavelprata space 945 (everything else ≤235), @wazzcrypto fintech 966, @uselegion space 891 / biotech 798. No fake elites, and the two junk accounts are floored out entirely (below the reputation floor → no sector score at all).

The AI super-accounts under all three variants

The six clearest AI figures alive, full per-sector rank under each new variant. Watch the texture: the first two are solid 1000 everywhere (the bleed); the residual is the only one that differentiates them.

① edge-weight + percentile , every titan, every sector = 1000 (maximum bleed)

② edge-weight + log-raw , still 1000 everywhere (log-raw de-saturates the scale, not the correlation)

③ edge-weight + floored residual , the bleed finally breaks into a per-sector profile

So which do you actually want? If the sector score should answer "who's the biggest authority in AI" → use the percentile (Altman = top), accepting that prominent people rank high in several sectors. If it should answer "who genuinely specializes in AI beyond general fame" → use the floored residual (specialists surface, generalists flatten, junk gone). The two are different products; v2's whitened score tried to be the second but botched the math.

The trade-off, across all the options

* the residual elevates the titans (no inversion, sama AI = top-21% among reputable accounts, not bottom-35% like the whitened score) but, being a specialty lens, won't rank a famous generalist #1 in their home sector. Pick percentile for authority, residual for specialty.

Bottom line: there is no single column that is both "biggest authority" and "purest specialist", that's the real tension, now measured. Percentile = authority (Altman tops AI, with some cross-sector bleed). Floored residual = specialty (sharp real specialties, junk floored, bleed broken, but famous generalists flatten to "broadly strong"). Whitening tried to be the specialist lens and broke the math (junk on top, titans sunk). The shippable pair: percentile for the headline sector rank, and the floored residual as a "specialty/edge" signal beside it, never the whitened score.

Stronger gating: hubs unmoved · specialists sharpen · peripherals & junk break

5 sectors. @sama (≡ all 6 titans) stays 1000 everywhere, sharper gating can't move him. @pavelprata sharpens to a clean space specialist (off-sectors 989→116). The cost: @uselegion, a genuine low-profile space account, is crushed (space 892→157), and the junk floats up to a uniform ~591. Strong gating is worse, not better.

The one number that matters, @sama's raw reputation

Altman's AI reputation is 17–38× his quantum, at the current weighting already. The separation exists. But his quantum raw, though tiny for him, still beats ~99.9% of 6.75M accounts, so the percentile pins it at 1000. Ranking against the population destroys the within-account distinction.

Two questions, two scales, the resolution

The fix for the thing that looked wrong. "Altman is AI, not quantum" is a within-account statement, and it's already true in the raw (17–38×). No cross-account ranking can surface it, his absolute quantum standing genuinely is top-tier. So the per-account sector profile should be scored by within-account-normalized raw, applied only to reputable carded accounts (you never card a train company, so the junk blow-up never arises). That is exactly what the InvestorDNA radar already does, each person normalized so their top sector = 100.

• Per-account "what are they about" (radar): within-account-normalized raw → Altman on AI. ✓
• Cross-account "who's top in sector X" (leaderboard): population percentile → fine; the hub "bleed" is genuine absolute reputation.
• Do not pursue stronger edge-weighting (doesn't move hubs, crushes peripheral genuine accounts) or the whitened ratio (junk on top).

First, the algebra: why "1/global on every edge" mostly cancels

In EigenTrust each account distributes 100% of its trust as shares across who it follows (rows sum to 1). Multiply all of one account's out-edges by the same factor (1/global) and the renormalization-to-100% divides it straight back out, a no-op. So only the forms that break that assumption do anything:

• A, seed weighting (seed mass ∝ 1/global): the seed/teleport is injected separately, never normalized away → real, but turned out negligible here.
• C, target-side discount (weight edges into famous accounts less): varies across a source's edges → doesn't cancel.
• Approach 2, non-renormalized source down-weight (a famous account passes <100% of its trust, remainder returns to seed): deliberately breaks the 100% rule → the faithful realization of the idea.

Decorrelation, sector↔global correlation (lower = better)

A (seed) is a near-perfect no-op, reweighting ~30–48 curated seeds is too diluted to survive propagation. C (target) actually made it worse. Approach 2 is the only one that improves every sector (and it beats our current edge-weighting's ~0.61), its source factor correctly targets the famous (f: @sama 0.005, @pavelprata 0.61, @uselegion 0.999), lifts the niche (@uselegion space 892→920, @fabrizio 951→966) and sinks junk (@ouigo quantum 18→2). Matty's instinct is validated, as a non-renormalized source down-weight it's a real, modest pipeline improvement.

Full per-account grid, 12 sample accounts (global + 10 sectors)

baseline (B0), current edge-weighting

A · inverse-global seed weighting , identical to B0 within ±2; the seed reweighting is a no-op

C · target-side global discount , lifts niche, sharper but worse aggregate corr

Approach 2 · non-renormalized source down-weight , best aggregate decorrelation

Read: A_seed ≈ B0 (a no-op). C and Approach 2 both lift the genuine niche accounts, @uselegion ai 478 → 621 (C) / 558 (P2), space 892 → 955 / 920, and Approach 2 best decorrelates (every sector's corr drops, table above). The 12 here all clear the reputation floor; the differentiation is real, not bleed.

The 6 AI super-accounts, full grid (this is the whole point)

B0 / A_seed / Approach 2, identical: 1000 across every sector

C · target discount, the only variant that even dents them (a handful of 997–999)

Six different people, all pinned at the ceiling in all 10 sectors, under the baseline and the strongest source down-weight (Approach 2). Target-discount (C) barely scratches a few cells to 997–999. No weighting variant gives them a real per-sector profile, exactly the iron law below.

The iron law, raw separation explodes, percentile never moves

Every lever widens @sama's within-account AI-vs-quantum raw ratio, up to 2,218× under Approach 2 (his quantum raw is 0.05% of his AI). And at every step his quantum percentile stays pinned at 1000. We slashed his quantum reputation 2,200-fold and he's still top-percentile, because even that crushed value beats 99.9% of 6.75M accounts.

@sama, AI vs quantum vs nuclear, three scorings of the same raw

Percentile and clipped-log-raw both pin his quantum at 1000, his top-0.1% quantum raw sits above the 99.9th-pct clip, so it saturates. Only the unclipped magnitude scale preserves the 2,218× raw gap → AI 1000, quantum 735.

Unclipped log-raw, the 6 AI titans finally get real profiles

AI is now clearly #1 for all six, and genuine secondary strengths surface: @karpathy robotics 1000 (Tesla autopilot), @demishassabis biotech 1000 (AlphaFold), @sama robotics/space/semi ~785–800. The magnitude scale recovers what percentile and the clip hid.

…but the cost, everyone below the elite compresses

The same 12 sample accounts that read 900–1000 on percentile now read ~250–650: @pavelprata (global 998) mostly 350–650, @uselegion space 309 / ai 51. Stretching the scale to the single top squashes the merely-very-good. Fine for a per-account radar, poor for a cross-account leaderboard.

Where this leaves the scale choice

• percentile / clipped log-raw, hub saturates, no per-sector profile; but clean ranking for the non-elite.
• unclipped log-raw, the only cross-account scale that shows the hub's profile (and secondary strengths), at the cost of compressing the 99%.
• within-account normalization (radar), the clean answer: the hub gets AI≫quantum and everyone keeps their own profile, because each person is normalized to themselves.

So log-raw can bolt onto Approach 2, and unclipped it does reveal the hub, but it inherits the cross-account compression. Same split as §10: leaderboard → percentile (clip/de-saturate as you like); "what is this person about" → within-account/radar.

What "Approach 2 · source down-weight" actually is

An EigenTrust variant that stops globally-famous accounts from leaking generic prominence into every sector. It's two layers on top of the trust propagation, then a scale:

1. Edge-weighting (target relevance). Each follow is first weighted by how sector-relevant its target is, how many of that sector's seeds follow the target (ssi(target) + floor), then row-normalized so every account hands out 100% of its trust as shares. (The same edge-weighting used throughout this report.)
2. Source down-weight (the new layer). Each account's total outflow is then multiplied by f(u) = gref / (gref + global(u)), a factor that's ≈1 for normal/niche accounts and →0 for the genuinely globally-prominent (gref = the 99.9th-percentile global score). So a famous account passes only a small slice of its trust onward; a niche domain expert passes ~all of it. The unspent trust (the 1−f(u) part) is returned to the seed, so the system stays mass-conserving and converges.

Why it helps. A famous account's follows are topic-blind, everyone follows them, they follow everyone, so their trust spills generic prominence into every sector. Shrinking their outflow makes sector trust flow preferentially through the niche, topic-specific accounts (the ones picked as seeds for being domain experts, not for being famous). That sharpens the sector signal: it's the best aggregate decorrelation of everything tested (§11, sector↔global 0.60→0.53), and it lifts genuine niche accounts while sinking junk hardest.

The final leaderboard, Approach 2 + log-raw (clipped) · global + 10 sectors

12 sample accounts , GLOBAL = de-saturated overall score; sector cells = log-raw graph rank

6 AI super-accounts

What you get: a graph-wide, cross-account sector rank (you can compare and rank people on it), but de-saturated, @pavelprata reads space 1000 with genuine standing elsewhere (585–972), not the percentile's flat 968–1000; @wazzcrypto fintech 941; @uselegion sits lower across the board (global 792 → cells 84–519) because his overall reputation really is lower. No 0s for strong accounts, and lower-reputation accounts correctly rank lower everywhere, that's the comparability the radar throws away.

Footnote, optional per-account "radar" (a separate card visual, not a rank)

For a per-person card you can additionally show a within-account view (each person normalized to their own top sector). It answers "what is this person about", @sama AI-dominant, @karpathy robotics, @demishassabis biotech, and is useful as a shape/visual. But it is self-relative, not a graph rank: a 1000 means "their own top sector," not "#1 on the graph," and you cannot rank people on it (it's exactly why a global-998 account like @pavelprata shows 0s in it). Use it only as a card visual alongside the overall score, never as the sector leaderboard. The leaderboard above is the final scoring.

What changes vs §13, 2 of 3 components

1. Edge weight: one question → two questions (purity edge-weighting). Both versions decide how much sector trust a follow may carry by looking at the follow's target. §13 asked a single question: "how many of this sector's seeds follow the target?", an absolute count. The purity upgrade asks a second one: "and of all the seed endorsement the target gets, across all 10 sectors, what fraction comes from this sector?", and multiplies the two. In words: endorsement strength × endorsement purity. The worked example below shows why the second question is the one that stops hub bleed.
2. Source down-weight f(u), unchanged. The Approach-2 layer from §13 stays exactly as documented there.
3. Scale: un-pin the top, two variants, same ranks. §13's log-raw clip collapses everything above the 99.9th percentile to 1000; both variants replace that clip with a 950–1000 log-magnitude band so the elite finally differentiate. The magnitude scale (LOG-KNEE) keeps §13's log-raw scale untouched below the band (scores ≈ §13 × 0.95, same de-saturated feel), and spreads the formerly-pinned top 0.1% across 950–1000. The rank scale (KNEE) uses the percentile below the band instead (mapped 0–950, so bulk scores read higher) with the top 0.5% spread 950–1000, "≥950" literally means top 0.5% of the graph. Both are monotonic in the raw → identical rank order; the choice is display semantics only.

Decorrelation, every one of the 10 sectors improves

Sector↔global correlation (rank level; lower = more sector-specific signal). purity weighting beats the §13 edge weight in all 10 sectors; purity + un-pinned scale together cut the average from 0.54 → 0.40:

Columns: §13 raw scored by percentile · purity-weighted raw by percentile · §13 raw on the un-pinned (knee) scale · purity-weighted raw on it. The §11 source down-weight took 0.60 → 0.53; purity weighting + an un-pinned scale continues to ~0.40, the largest decorrelation of anything tested in this report, with no degenerate side effects (unlike whitening, §8). The two v2.1 scale variants are rank-identical, so their correlations match (LOG-KNEE avg 0.394 · KNEE avg 0.396).

The v2.1 leaderboard, new scores (purity edge-weighting + Approach 2), shown on the magnitude scale

This is the v2.1 proposal as it would ship. The scale is LOG-KNEE, §13's log-raw (clipped) scale, kept: every score below the top band is the §13-style log-raw value (×0.95), so the numbers read exactly like the §13 final you've already seen. The only scale change is at the very top: the scores that §13's clip pinned to a flat 1000 now spread 950–1000 by magnitude.

Hubs: @sama reads AI 1000 · quantum 958, clearly elite everywhere, #1 only where he genuinely is. @karpathy's robotics 1000 and @demishassabis's biotech 1000 (Isomorphic Labs) are real and kept, and the magnitude band separates them hard where they're weakest (@demishassabis nuclear 844, @karpathy nuclear 892). Specialists: profiles get shape without losing graph-comparability, @pavelprata space 957 vs AI 392; @wazzcrypto fintech 893 vs AI 397. Junk sinks further (@ouigo_es ai/ml 267, 0 elsewhere, global 56). Even the GLOBAL column de-saturates: the hubs differentiate at 994–996 instead of all reading 1000.

Same scores, alternative display, the rank scale

Identical v2.1 scores (purity edge-weighting + Approach 2), identical rank order, only the number semantics change. Here the score below the top band is the percentile (mapped 0–950), so bulk numbers read much higher than §13's log-raw, and "≥950" literally means top 0.5% of the graph. Pick this one if scores should feel like ranks ("top 3%") rather than magnitudes; the choice between the two displays is deferred to product calibration.

Compare any row with the table above, e.g. @matty_ quantum reads 559 on the magnitude scale and 921 here. Same account, same raw score, same rank; the magnitude scale spreads people by how much reputation they have, the rank scale by how many accounts they beat.

Interim path (no pipeline re-run): if the v2.1 pipeline re-run isn't scheduled yet, the un-pinned top band also works as a pure display swap on the unchanged §13 scores, identical ranks, ships in seconds, and already removes the flat-1000 hub rows (@sama reads AI 1000 / quantum 977), but decorrelation stays at §13's 0.54 until the re-run lands (full table: results_v21_spec_full10.txt, P2_knee).

Choosing between the two displays, pros & cons

Reminder: both sit on the same v2.1 raw scores, identical ranks, identical leaderboards, identical decorrelation. Only the printed number changes; the trade-off is what that number communicates.

Leaderboard sanity, top 10 per sector under v2.1

The decorrelation does not disturb who floats to the top, every sector's leaderboard is its genuine authorities, and crossovers appear exactly where they should (@elonmusk in robotics + space only; @karpathy in AI + robotics; @jigarshahdc in climate + nuclear):

Trade-offs & the dead ends (so nobody re-tests them)

1. Purity weighting is harsh on sector-unendorsed accounts. An account no sector seed (or seed-adjacent path) specifically endorses collapses in that sector, @uselegion's AI cell drops to 0 (effectively unscored there; its space 848 / biotech 715 stay strong). Honest, but tunable: the share exponent (shareγ, γ<1) softens it if calibration shows it's too aggressive.
2. The iron law still holds for percentiles. Even under purity weighting, @sama's quantum percentile is 1000, his diluted quantum raw still beats 99.9% of a mostly-zero population. It's the un-pinned scale (either variant) that displays the separation. The pair is the point.
3. The purity edge weight needs a pipeline re-run (same cost as Approach 2, one extra per-sector edge-weight factor). Both scale variants are post-hoc on preserved raws.

What was verified

Per-sector rank agreement

Method: tie-aware Spearman on 200k-account random samples restricted to non-zero accounts; top-N lists ranked by production score with raw-score tiebreak. Script: scripts/experiments/v22_verify_prod_v21.py.

Why displayed values differ by a few points (and why it doesn't matter)

Production's bulk scores sit +5 to +30 points above the report's tables, most at the bottom of the scale, zero at the top. This is calibration, not ranking: fitting the prod−report delta against log₁₀(raw) gives a perfect line (residual σ = 0.26, below rounding), which is the signature of the same formula with a slightly different lo anchor. The root cause is the non-zero set: production counts a few hundred thousand more accounts as non-zero per sector (different dust threshold), which moves the 0.5th-percentile floor of the log scale and the _percentile denominator. Every shift is monotone, nobody's rank moves.

Where everything lived (as of 2026-06-10)

• Live (June): data/eigen-rep-prod/, the 2026-06-03 graph (accounts.csv, follow-edges.csv, unchanged) + the v2.1 scores (eigentrust-v2.1-downweight-purity-logknee.csv).
• Archived (2026-06-10): the original 2026-06-03 scoring → data/archive/eigen-rep-prod-scores-2026-06-03/; the v2 λ-whitening exports (§8) + source zip → data/archive/eigenrep-v2-exports-2026-06-08/. Each with a README.
• Consumers to update: the old scores.csv schema (5 cols, sector_scores JSON) is gone from the live dir, cards/investor_dna/scripts/build_eigen.py and the S1-promotion scripts still point at it and need porting to the flat 24-col schema.

The finding, why no graph change can fix sector pollution

Layer 0 · the content signal r_s, what an account is, not who follows it

For the top ~40k accounts per sector (the union where all pollution lives, the 6.6M-account tail already scores ≈0), embed name + bio + tweets with all-MiniLM-L6-v2 and take the cosine to a per-sector centroid (½ seed-account text + ½ a hand-written sector description). r_s ∈ [0,1] is semantic, not keyword, the named failure mode is Jimmy Fallon's bio joking "astrophysicist" and landing in a quantum top-500; the embedding places his comedy far from the quantum centroid regardless. The signal cleanly separates real operators from famous bystanders:

Experts sit at 0.50–0.75 in their domain; celebrities peak at ~0.2–0.35 everywhere. The per-sector gate threshold, θ = median r_s of the sector's top-1000 (quantum 0.55, space/nuclear 0.41, defense 0.36), falls in the gap between them.

Layer 3 · the content gate, earn the sector or leave it

An account holds a sector-s score only if eligible(v,s) = is_seed(v,s) OR abstain(v) OR r_s ≥ θ_s, otherwise the score is zeroed and the percentile recomputed over the surviving pool. Three hardening layers close the gaps:

Deeper coverage for thin bios, tweets + other languages

Tweets: for 1,308 accounts whose bio alone was too thin, r_s was recomputed from bio + up to 20 original tweets, e.g. @jimmyfallon space 0.47 → 0.14 once his actual (comedy) tweets are read, so he gates out. Languages: 321 non-English "abstain" accounts were scored in a separate multilingual model, 128 kept in-sector (e.g. Japanese AI/quantum/robotics researchers), 193 gated (parenting/creator accounts), instead of blanket-keeping them.

The effect is invisible at the top (the top-15 leaderboards were already clean, §15) and decisive in the 990–999 percentile band, the number a card actually shows. All scraped inputs are committed under data/eigen-rep-prod/ (SCRAPED_DATA.md) so nobody re-scrapes; total one-time cost ≈ $8.

Layer 2 · the purity-ratio ρ specialty lens, and the pollution, gone

On top of the gate, the floored purity ratio ρ = t_sector / t_global (λ=1, applied only where global rank is elite so a tiny denominator can't explode) demotes an account's off-domain sector standing, the "specialty lens." The table below is the combined payoff, gate + ρ, v2.1 → v2.3 percentiles, on the accounts the founder named. Net sector↔global percentile correlation over the surfaced pool: 0.38 → 0.16 (avg of 10 sectors).

Layer 4 · the decibel scale, the elite band is finally scarce

The log-knee magnitude scale (§14) put 13,067 accounts at ≥900, "top" meant nothing. v2.3's global headline is a decibel scale: clip(1000 + 140·log₁₀(raw / r_ref), 0, 1000), anchored to a frozen #1-account raw so it never re-inflates as the graph grows. Now 319 accounts clear 900. Sectors and the world rank also carry a fixed-headcount tier ladder.

Tier ladder, fixed headcount by world rank (per sector and global)

Our 12 sample accounts in the final v2.3 scoring

The accounts this report has tracked since OLD → NEW → IMPROVED, as they stand in v2.3, decibel headline, world rank and tier. None currently holds a vetted sector standing: all 12 sit below the top-5000-by-graph-score band the content gate vets, so under the Tier-0 rule "only surface a sector score for a content-verified account" they show global standing only. (An earlier version of this table showed sector scores of 200–889 here, those were un-vetted tail artifacts, removed by the Tier-0 fix; see eigenrep-v3-audit.md. The Tier-2 full-graph r_s embed on the roadmap will extend sector coverage to accounts like these.)

Decibel = frozen-anchor magnitude 0–1000; tier by world rank (Notable = 2,001–30,000). That the whole set shows "no sector" is the honest post-audit state: v2.3 only vouches for the ~800–4,100 content-verified accounts per sector (quantum 797 … defense 4,135); broadening that coverage is the Tier-2 build.

What we tried, and the two hard limits (the research)

ρ is the survivor of a full sweep of source-side decorrelators, each a complete EigenTrust re-run on the 39.6M-edge graph. Metric = sector↔global rank correlation on the raw sector vector (lower = less popularity leakage); v2.1 baseline = 0.40.

Where v2.3 lived (June-graph build)

• Scores (June-graph build): eigentrust-v2.3-{purity-ratio,content-gated,scored}.csv in data/eigen-rep-prod/, global identical to v2.1; sectors = ρ + gate; the -scored file adds the decibel headline + tiers.
• Build: scripts/eigenrep_content_gate/build_v23.py → apply_gate.py → scale_tiers.py; content signal in build_rs.py. Scraped inputs (account types, tweets, r_s) are committed under data/eigen-rep-prod/, see SCRAPED_DATA.md so nobody re-scrapes.
• Consumers & API: build_sample_api.py, the S1-promotion scripts, and cards/investor_dna/scripts/build_eigen.py all prefer v2.3 (v2.2 → v2.1 fallback). The read-only API now exposes global.db, global.tier, sectors.*.tier, and a score_scales block. Full write-up: docs/technical/eigenrep-v3-redesign.md + eigenrep-v3-results.md.

The quality scorecard, eval_harness.py, ~196 handle-verified cases

The single recall miss is Mira Murati (a borderline case at the rescue threshold). A broad manual sweep confirms every sector's top-15 is genuine experts. Full record: docs/technical/eigenrep-v3-audit.md.

What was fixed, three tiers of hardening

Tier 0 · honest denominators

The un-vetted tail (6.7M accounts) is zeroed in every sector, so a sector percentile is computed over the content-vetted pool, not the flood, a fly-fishing report no longer read "space Elite". Global was re-sourced verbatim from the prod v2.1 CSV (a build had regenerated it from experiment raws and drifted up to 63% for mid-tier accounts).

Tier 1 · the global ladder is now investors

Org filter → "Org" tier: 672 brands / media / agencies / VC-firms (OpenAI, NASA, Tesla, Sequoia, YC, CNBC) pulled out of Legend/Titan/Elite. ρ_global → "Fame" tier: 298 generic-fame individuals (CNN, Bernie, Pelosi, Tom Hanks) demoted by a seeded÷uniform-PageRank purity ratio. Seed-badge: the 318 curated anchors are flagged (global_is_seed) so the elite tier isn't misread as organically discovered. ρ demote, not delete: a floored purity multiplier restored the content-verified cross-sector bridges the earlier ρ had zeroed (Preskill nuclear 0→369).

Tier 2 · the signal, measured

ALF follower-composition backstop asks: is an account's follower set over-represented in sector s vs the base rate? Thin-bio experts followed by sector people are rescued, Peter Shor, whose bio is song lyrics (r_s 0.29), has followers 54× concentrated in quantum → quantum Legend; celebrities followed by everyone stay at lift ≈ 1 and are not rescued. Recall 73% → 99%. ALF corroboration closes the bio-stuffing attack, a stuffed bio can't move the follower signal, so "@realdonaldtrump + defense keyword" is blocked. Media/brand detector catches the unpaid outlets X's verifiedType misses (The Robot Report, Endpoints News, EE Times, AnandTech), lifting hard-negative precision 38% → 100%.

Honest limits, measured, not hidden

Why corr was the wrong ruler, four proofs

On ground truth, every decorrelator is the same, and why

North-star metric = AUC: per sector, the fraction of (real expert, non-expert) pairs the ranking orders correctly. Rank-based, un-gameable by rescaling, lives at the top where the product is. Measured on the raw decorrelator vectors, ~12 experts × sector media/celebrities:

The corr spread across these was huge (0.60 → −0.21); the ground-truth difference is zero. The reason is a hard ceiling: a graph-only decorrelator can only flatten uniform fame, never demote it. Elon in quantum sits at percentile 999.9 → 998.1 (ρ) → 999.5 (matched-null), no method, however principled, pushes a universally-followed celebrity below the real experts. Only the content gate (what an account posts) and the org/media filter crack that, which is exactly what they already do (§16–§17).

The one robustness idea, tested and declined, the matched-null denominator

ρ = t_sector/t_global is a known, sound quantity (1 − Gyöngyi spam-mass, a likelihood ratio). Theory says the denominator should be uniform-popularity, not the investor-seeded global, a version that needs no bridge-protecting floor. It was built and measured against the shipped floored ρ on 29,076 content-eligible cross-sector standings:

Declined. The robustness claim doesn't hold empirically, the simple floor protects genuine cross-sector bridges better than the principled null (3× fewer crushed), with no offsetting AUC gain. The elegant theory loses to the measurement.

What changed, the ruler, not the scores

Kept: the shipped decorrelator (ρ + demote-floor). Nothing beats it on ground truth, and the low-corr variants over-whiten below the real signal floor. Retired: corr as an objective, it's now a one-line tripwire that fires only on inversion or fame-leak. New north star in eval_harness.py, on the live scored CSV:

θ-lowering, the coverage/precision trade, measured

The gold-set scorecard rated θ-lowering a free win (recall ↑, AUC → 1.000, pollution/hard-neg clean), falsely: the leakers aren't in the gold negatives, and the worst (Karpathy, Nadella, Lex) are s0 seeds of a different sector, invisible to a celebrity/media list. Only a hand-built adversarial leakage probe caught them. That probe is now a permanent harness axis (cross-sector leakage: 8/8 on production).

What this means

Corrections the audit forced (intellectual honesty)

The real hole, found, measured, and fixed for free

The follower-composition backstop admits accounts on who-follows-them alone. A random sample showed it is excellent in big sectors (AI: real researchers/VCs) but content-blind in small ones, quantum surfaced a DJ, a phone company (OnePlus), and a gaming exec at top-percentile scores. Fix (no scraping, the bios were already in our data): embed the follower-rescued band and content-check it too. Result:

Technical mechanism of each change (with a concrete before→after)

What each change impacted, at scale, with examples

The sector scores are a spectrum, not all-or-nothing

fintech kept-scores: p10 250 / median 434 / p90 951, 63% below 500. A single account is a mix: @lisasu semi 1000 + defense 824; @pavelprata space 955 + fintech 399. A sector 0 = not a content-verified participant there, not "ranked low." The mid-tier VC accounts that zero out (e.g. @murphcapital) weren't content-verifiable deep-tech operators; their old sector scores were follow-graph fame. (@wazzcrypto looked identical to the cosine gate but the v2.5 LLM recovers his real fintech.)

How to read every cell: two numbers, June score → Now score. When they look identical (e.g. 999→999) the score was kept unchanged, not a duplicate. Green kept (content-verified), amber endorsed middle tier (v2.4, §22), red strike removed, · never scored. The June columns are dense (graph-only put everyone in every sector); the Now columns are sparse and honest (content gate + ρ + org/fame filters + endorsed recovery).