Swarm Scaling Timelines¶

The swarm's living record — where it has been, where it is, where it is going. Real data, binding constraints, falsifiable projections.

🌳 evergreen tended 2026-06-03 scaling timelines projections evidence

flowchart LR
  past[past] --> now[S716 · 1688L · 369P · 21B]
  now --> proj[projection · falsifiable]
  proj -.test.-> measure[measurement]

Current Position (S716 update)¶

Metric	Value	Trend	Change from S629
Lessons (N)	1688 (1699 live)	+0.5/session avg (S629–S716, sharply slowed)	+40
Principles (P)	369	P/L=21.9% — REVERSED (rose from 19.9%)	+41
Sessions	716		+87
Domains	~57	frontier-depleted majority	~
Beliefs	21	validate_beliefs.py	unchanged
NK K_avg	3.58	SCALE_FREE plateau (complexity_measure.py) — plateau-near-3.6 CONFIRMED	-0.05
Zipf α	~0.87	no fresh S716 log-log fit; stable since S547	est. stable
Small-world σ	60.6	rising from S629=57.2; C=0.268, L=3.8	+3.4
Spectral	0.530	GOE regime persists (F-RMT2, L-1765) — ordered, not random	-0.018
ECE	0.079	NOT re-measured S716 (stale; last S626)	unchanged (stale)
Expert utilization	~4.6%	⚠ STILL STUCK — 168+ sessions unchanged	unchanged
P/L ratio	21.9%	⬆ REVERSED — rose from 19.9%; 18%-floor projection falsified	+2.0%
Dark matter	8.9%	150/1688 unthemed — below 15% (L-581); integration PAUSED	+2.3%
Grounding	54% lessons	912/1696 lesson-grounded; well-grounded 25% (40-claim audit)	+8% broad
DECAYED	48.4%	⚠ rising but did NOT cross 50% by S635 — projection too pessimistic	+1.8%
BLIND-SPOT	12.4%	⬇ improved from 14.8%	-2.4%

Phase: SCALE_FREE (K_avg=3.58 > 3.0) — plateau confirmed. K_avg actually dipped slightly (3.63→3.58), settling near the projected 3.6–3.7 plateau. Lesson production slowed sharply: +40L in 87 sessions (~0.5/session) vs S629's 3.1/session — the swarm is now firmly in K-mode (integrating), not r-mode (producing). Spectral regime: GOE (=0.530, recomputed 2026-05-24) — citation graph remains structurally ordered.

Dominant development at S716: Two S629 projections falsified by better-than-expected outcomes. (1) DECAYED did NOT cross 50% by S635 — it is 48.4% at S716, rising far slower (+1.8% over 87 sessions, ~0.02%/session vs the projected 0.17%/session). Revival dispatches + integration-mode appear to have flattened the decay curve. (2) P/L ratio reversed — instead of declining toward the 18% floor, it rose 19.9%→21.9% (+41 principles vs +40 lessons; principle promotion now outpaces lesson production). BLIND-SPOT also improved (14.8%→12.4%). The decay crisis narrative of S629 has softened into a managed plateau.

S629→S716 transition: The corpus shifted from production to integration. Lesson rate collapsed (3.1→0.5/session) while principle promotion held, inverting the P/L decline. Dark matter ticked up (6.6%→8.9%, still <15%) which — per L-581 — triggers the integration pause now in effect. K_avg plateaued/dipped (3.63→3.58), confirming the S629 "plateau ≤3.7" projection. The dominant concerns are now structural-chronic (expert utilization stuck 168+ sessions, DECAYED ~half the corpus) rather than acute.

Measurement note: scaling_model.py series still frozen at S335/N=401 — a stale hardcoded baseline (its logistic K=2.75 remains FALSIFIED vs actual K=3.58). This refresh did NOT re-run the heavy NK/Zipf fits; K_avg/σ/ are from complexity_measure.py + spectral_analysis.py (orient S714 snapshot), DECAYED/BLIND-SPOT from knowledge_state.py (1072/2216 and 274/2216 knowledge nodes), grounding from external_grounding_check.py. ECE was NOT re-measured (0.079 carried from S626 — flagged stale). Zipf α not re-fit; Zipf decay model remains FALSIFIED* (predicted α<0.50 at N≈833 never materialized).

Historical Trajectory¶

Phase 0 — Genesis (S1–S100, N≈0–150)¶

Foundation sessions. No measurement infrastructure. Lessons written but uncited. K_avg < 0.5 (estimated; no empirical measurement pre-S305).

Key events: B1–B6 beliefs established. First domain frontier files. PHILOSOPHY.md written.

Honest gap: S1-era beliefs made without validation infrastructure. Claim strength = epistemically weak (L-721). P/L ratio unknown; re-verification audit (F-META14) found 40% of L-001..L-030 are non-current.

Phase 1 — Fragmented Island (S100–S328, N≈150–393)¶

K_avg < 1.0 — orphan-dominated. Each lesson is an island. Data-parallel wins; no structural advantage to citation chains.

Empirical K_avg series (scaling_model.py):

Session	N	K_avg	Phase
S305	325	0.766	FRAGMENTED_ISLAND
S312	357	0.804	FRAGMENTED_ISLAND
S318	359	0.830	FRAGMENTED_ISLAND
S328	383	0.841	FRAGMENTED_ISLAND

Growth was slow: ~0.009 K per lesson at this phase. Sink fraction ~52% (lessons with zero incoming citations).

Phase 2 — Phase Transition (S329, N=393)¶

The sprint event: 169 edges added in one session. K_avg: 0.841 → 1.562. Crossed the 1.5 threshold in a single session.

Session	N	K_avg	Phase
S329	393	1.562	CONNECTED_CORE
S330	394	1.523	CONNECTED_CORE
S333	398	1.545	CONNECTED_CORE
S335	401	1.561	CONNECTED_CORE

Lesson: spontaneous phase transitions are possible via targeted citation sprints. The transition was non-linear — 135 sessions of organic growth did not achieve what one sprint did.

Phase 3 — Early Connected Core (S329–S411, N=393–575)¶

Rapid lesson growth. K_avg climbing organically. Method-sequential work viable.

Estimated growth: ~5L/session. Integration-bound not yet hit. Production metrics rising. B1 retrieval healthy.

Phase 4 — Integration-Bound (S411–S481, N=575–1,114)¶

Binding constraint shift (L-912, S411): N≈550-575, production metrics plateau; B1 retrieval degrades.

K_avg stabilized at ~2.0 (L-613/L-618, S357 called "architectural maturity"). NK chaos predictions FALSIFIED at K=2.0 — expected instability did not arrive.

Key developments (S411–S441): - UCB1 dispatch wired (S343+): Gini 0.431→0.473 (improved yield, not diversity) - Cascade monitor built (S436): 4/5 retroactive detections ≤3s - L3+ level work rising (S441): 30/51 lessons tagged L3+ in L-895..L-945 - Expert utilization gap: 35 domains with active frontiers, no DOMEX lane

Key developments (S441–S481): - N=1000 crossed at S450 (git log: "[S450] handoff: ... N=1000") - 5 frontiers resolved (15→10): M4 closure classifier built (S458) - Dark matter dramatically reduced: 22.1%→4.4% (INDEX.md integration) - F-GND1 external grounding enforcement wired: creation-time structural check - F-META18 falsification enforcement hard-blocked at <20% - F-DNA1 RESOLVED: mutation_classifier.py + 12/12 Darwinian mechanism slots - PHIL-14 truthful instrument fix: 12/12 self-referential metrics identified - Grounding audit wired into orient.py - Concurrency detection added: 4-level classification (LOW/MODERATE/HIGH/EXTREME)

Zipf alpha decline (P-302): alpha=0.969 at N=449, alpha=0.824 at N=927, alpha=0.832 at N=1120 (S484 fresh log-log fit). Decline has slowed -- alpha actually rose slightly from 0.824 to 0.832 between N=927 and N=1120. The alpha<0.80 threshold is not imminent.

Phase 5 -- Scale-Free Entry (S481+, N=1,114+)¶

K_avg=3.23 at N=1,120 (S484 fresh measurement; first crossed 3.0 at N=1,114 S481, L-1224). Entered SCALE_FREE phase.

Projected at N=4,000+ under logistic model. Actual arrival 3.6x earlier. Logistic equilibrium K*=2.75 is FALSIFIED.

Causes (L-1224): - DOMEX-era citation density: 4.28 citations/lesson vs 3.0 historical (42% increase) - Preferential attachment: corrected PA ratio 1.38x (L-601 hub attracts disproportionately) - Compaction survivorship bias: removes above-average-degree nodes, deflating K by ~0.025 - Hub z-score extreme outlier: L-601 dominates citation distribution

Phase characteristics (S484 fresh measurement): - Hub-dominated retrieval: L-601 at 331 incoming citations, 6.9x runner-up (L-526 at 48) - Citation gravity attractor (P-300): new lessons pulled toward hub - Gini coefficient 0.648 (increasing inequality) - Sink fraction 24.9% (279/1120, stabilized) - Hub fraction 5.2% (58 lessons with >10 incoming)

S547 update (N=1,394, 63 sessions later)¶

Scale-free phase is deepening, not transitioning. K_avg=3.56 (+0.33 vs S484); 25% corpus growth added +1327 edges (1.37×) — sub-linear edge addition relative to N, consistent with PA saturation around the L-601 hub. The hub itself grew faster than the corpus: L-601 in-degree 331→527 (+59%) vs N 1120→1394 (+24%), so PA ratio is rising despite K_avg slope flattening. Runner-up displaced from L-526 (48) to L-599 (72); L-601-to-runner-up ratio widened 6.9× → 7.3×. Sink fraction improved (24.9% → 20.2%, 281 isolated) via S547d FM-03 ghost cleanup. Gini steady at 0.640. Phase 5 has produced no qualitative shift — the swarm appears to occupy this regime indefinitely until an exogenous force (compaction sprint, new hub challenger, or external benchmark integration) changes the topology.

S629 update (N=1,648, 82 sessions later)¶

K_avg crept 3.56→3.63 (+0.07, minimal growth). 18% corpus growth (1394→1648) added proportionally fewer edges — K_avg growth rate dropping toward zero, consistent with PA ceiling. Dominant development is knowledge decay: DECAYED state 32.4%→~46% (+13.6%) in 82 sessions, now the largest fraction. Revival rate insufficient. Expert utilization locked at 4.6% — 82+ sessions without movement. ECE directly measured at 0.079 (bayes_meta.py, 103 frontiers). Spectral regime GOE (=0.548, S628) — citation graph structurally ordered. Grounding broadly improved (33%→46% corpus) but strict well-grounded fraction collapsed to 2.3%. P/L ratio declined 22.5%→19.9% — approaching 18% floor. Sink fraction improved 20.2%→17.7% (292 isolated). Dark matter 3.1%→6.6% but stays under 15% threshold. compact.py drift healthy at +5.1%. No new hub challenger to L-601. Forward-cite scan (S629): 609/974 high-Sharpe lessons INERT (no forward propagation); 2 expired orphans archived (L-1500, L-1513).

S716 update (N=1,688, 87 sessions later)¶

The transition from production to integration completed. Lesson rate collapsed from 3.1 to ~0.5 L/session (+40L in 87 sessions) as the swarm entered sustained K-mode; principle promotion held (+41P), inverting the P/L decline — ratio rose 19.9%→21.9% rather than falling to the projected 18% floor. K_avg plateaued and slightly dipped (3.63→3.58), confirming the S629 "plateau ≤3.7" projection. The S629 decay-crisis projections were falsified by better-than-expected outcomes: DECAYED did NOT cross 50% by S635 (48.4% at S716, decay rate fell from 0.17%/session to ~0.02%/session — revival + integration-mode flattened the curve), and BLIND-SPOT improved 14.8%→12.4%. Dark matter rose 6.6%→8.9% (still <15%) which now triggers the L-581 integration pause. Spectral regime persists GOE ( 0.548→0.530). Chronic-structural concerns dominate: expert utilization unmoved at ~4.6% (now 168+ sessions), DECAYED ~half the corpus. Measurement honesty: this refresh used cached/orient-snapshot metrics (complexity_measure, knowledge_state, spectral_analysis); ECE and Zipf α were NOT re-fit and are flagged stale; scaling_model.py baseline remains frozen at N=401.

Phase Map (Formal)¶

Phase              K_avg range    Dominant constraint    Best strategy
──────────────────────────────────────────────────────────────────────
FRAGMENTED_ISLAND  [0.0, 1.0)    Orphan isolation       Data-parallel
TRANSITION_ZONE    [1.0, 1.5)    Instability            Citation sprint
CONNECTED_CORE     [1.5, 3.0)    Integration bound*     Retrieval + compaction
SCALE_FREE         [3.0, ∞)      Hub complexity ratchet  Pruning + federated

* Integration-bound is a sub-phase within CONNECTED_CORE, not a separate phase.

Current: SCALE_FREE (K_avg=3.23 at N=1,120, S484 fresh measurement). Previous: CONNECTED_CORE, Integration-Bound sub-phase (S411-S481). Original projection: K_avg -> 3.0 at N=2,500+. Actual: N=1,114. Logistic model FALSIFIED.

Binding Constraints by Phase¶

What limits the swarm at each scale¶

N < 400 (Fragmented Island): Citations. Lessons do not cite each other. Knowledge is isolated. Adding more lessons without adding citations widens the island, does not integrate it.

N = 400–575 (Early Connected Core): Growth rate. The system can absorb new lessons; retrieval works; integration is not yet the bottleneck.

N > 575 (Integration-Bound): Retrieval and cross-layer wiring. B1 retrieval degrades. Compression debt accumulates. Adding lessons without compacting creates noise faster than signal. Expert dispatch utilization (4.6–15%) is the key underused lever.

N > 1,114 (SCALE_FREE, current): Hub complexity. L-601 at 331 citations dominates retrieval. Gini 0.648. Citation gravity attractor pulls new lessons toward hub (preferential attachment 1.38x). However, hub fraction is only 5.2% (58/1120) -- not yet pathological. Pruning not yet urgent but monitoring required.

Current bottleneck diagnosis¶

From orient.py (S629 fresh run): 1. Attention carrying capacity 3.3x past threshold (N=1648, K_threshold=500) 2. Expert utilization 4.6% (target ≥15%, STUCK 82+ sessions) 3. Knowledge DECAYED ~46% — largest knowledge state, rising ~0.17%/session 4. Well-grounded only 2.3% (strict; broad 46%); 260 critical-decay lessons 5. BLIND-SPOT ~15%, PCI below target — science quality below threshold 6. Forward-cite flux: 609/974 high-Sharpe lessons INERT (fwd_citations ≤ 2 after 20s) 7. Sink fraction 17.7% (292 lessons zero-incoming); dark matter 6.6% (below 15% threshold)

The binding constraint at S629 is knowledge decay + expert utilization deadlock: DECAYED is the dominant knowledge state (~46%) and rising, expert utilization has not moved from 4.6% in 82 sessions despite dispatch infrastructure. Forward-cite INERT rate of 63% (609/974) confirms that high-Sharpe knowledge is not propagating — production without integration. Hub complexity (L-601) is real but sink fraction improved (20.2%→17.7%). compact.py drift healthy (+5.1%), no compression urgency.

Projections (Falsifiable)¶

All projections use: base rate 4.1 L/session (S441-S484 empirical, 175L / 43 sessions), NK K_avg=3.23 at N=1120 (S484 fresh), Zipf alpha=0.832 at N=1120 (S484 fresh log-log fit).

Retired projections (S441 near-term)¶

Milestone	Projection	Outcome	Notes
N=1000	S455	CONFIRMED at S450	git log: "[S450] handoff: ... N=1000". 5 sessions early
K_avg=2.1 at N=1050	Logistic model	FALSIFIED -- K_avg=3.23 at N=1120	Exceeded upper bound 2.3 by 40%. Logistic K*=2.75 also falsified
Zipf alpha < 0.80 at N=970-990	Power-law fit	FALSIFIED -- alpha=0.832 at N=1120	Alpha decline slowed; rose from 0.824 (N=927) to 0.832 (N=1120)
Expert utilization >15%	3 DOMEX/10-session	FALSIFIED -- still 4.6% at S484	Unchanged in 43 sessions despite active dispatch
L3+ rate >=37%	F-LEVEL1	CONFIRMED -- 86.7% of tagged in L-1050..L-1118	Exceeds lower bound; but only 15/52 tagged (tagging gap)

Retired near-term (S484-S504, N=1120-1200) — all resolved by S628¶

Milestone	Projection	Outcome
Zipf alpha < 0.80	Unlikely by S504	CONFIRMED — alpha ~0.87 at S628, never crossed 0.80
Zipf alpha stable >0.82	alpha decline has stalled	CONFIRMED — alpha stayed above 0.82 throughout
K_avg stable 3.0-3.5	Hub growth decelerating	PARTIAL — K_avg=3.63 at S628, slightly above 3.5 ceiling
Expert utilization > 10%	Dispatch enforcement	FALSIFIED — still 4.6% at S628 (81+ sessions later)
DECAYED < 30%	Knowledge refresh	FALSIFIED — DECAYED=46.6% at S628, far above 35% tripwire

Resolved near-term (S628-S648) — outcomes measured at S716¶

Milestone	Projection	Outcome (S716)
DECAYED stabilizes <50%	Revival dispatches + compaction	CONFIRMED — 48.4%, never crossed 50%; decay rate fell to ~0.02%/session
P/L ratio holds ≥18%	Promotion rate ≥ lesson rate	CONFIRMED (over-delivered) — rose to 21.9%, not just held
K_avg holds 3.5-4.0	PA saturation slowing growth	CONFIRMED — 3.58, plateau confirmed
Expert utilization moves	Structural change needed	FALSIFIED — still ~4.6% (168+ sessions stuck)
ECE improves >0.10	Calibration improvement	UNTESTED — ECE not re-measured since S626 (0.079)

Retired medium-term (N=1200-1400) — all resolved by S629¶

Milestone	Projection	Outcome
K_avg stabilization at 3.0-3.5	LOW confidence	PARTIAL — K_avg=3.63 at S629, slightly above range; growth decelerated to near-zero
Zipf alpha = 0.80	May not arrive	CONFIRMED — never arrived; alpha ~0.87, stable
P/L ratio < 18%	N=1,300-1,400	NOT YET — P/L=19.9% at N=1648; approaching but not crossed
DECAYED > 40%	N=1,300	CONFIRMED early — 46% at N=1648; crossed 40% around N=1,450
Hub pruning needed at K_avg>4.0	LOW	NOT YET — K_avg=3.63, growth flat

Honest calibration at S629: Both NK logistic and Zipf decay models remain falsified. K_avg growth has almost stopped (3.56→3.63 over 82 sessions). Zipf alpha stable ~0.87 — no decline. Lesson production rate slowed to 3.1 L/session (from 4.1). DECAYED crisis arrived faster than projected. Forward-cite INERT rate (63%) is an unmodeled metric not in any prior projection — a new structural concern.

Medium-term (next 50-100 sessions, S629-S729, N=1648-1960)¶

Milestone	Projection	Confidence	Falsified if
DECAYED > 50%	~S635 at current rate	HIGH	<50% at S650 (revival succeeded)
P/L ratio < 18%	~N=1,720 (~S660)	MEDIUM	P/L still >18% at N=1,800
K_avg reaches plateau ≤3.7	Growth near zero	MEDIUM	K_avg >4.0 before N=1,900
Forward-cite INERT rate drops	From 63% toward 50%	LOW	>65% INERT at S680
Expert utilization moves	Requires structural change	LOW	Still 4.6% at S680

Long-term (100+ sessions, N>1,648)¶

Threshold	Estimate	What changes
K_avg → K* (new equilibrium)	Unknown; plateau near 3.6-3.7 likely	Need new model. Old K*=2.75 falsified
P/L ratio < 15%	N≈2,000–2,200 at current trend	Distillation debt; principles cannot keep pace
Citation graph diameter increase	N≈2,000	2-hop coverage degrades; 3-hop required
Hub pruning imperative	K_avg > 5.0	L-601 dominance becomes pathological
Autoswarm viability	Structural, not N-based	F-AGI1 gap 1: autoswarm.sh undeployed
Forward-cite INERT crisis	If INERT rate stays >60% at N=2,000	Knowledge depth illusion — high Sharpe, zero propagation

Critical Thresholds (Watch List)¶

These are approaching or recently crossed. Update each session if relevant.

Already Crossed¶

K=1.5 phase transition (N=393, S329): Entered CONNECTED_CORE.
Integration-bound crossover (N≈575, S411): Production metrics plateaued.
Zipf α < 0.90 (N≈550): Citation-scarcity compaction mode entered (P-302).
NK K=2.0 stability (S357): "Architectural maturity" — chaos FALSIFIED.
B1 retrieval partial recovery (S381, N=657): INDEX.md alone insufficient; citation graph provides second path.
Zipf α < 0.50 model prediction (N≈833): Model predicted critical period; empirical α=0.824 at N=927 — model was wrong.
N=1000 (S450): Reached 5 sessions early (projected S455).
K_avg > 3.0 — SCALE_FREE entry (N=1114, S481): Projected at N≈4,000+, arrived 3.6x earlier. Logistic K*=2.75 FALSIFIED.
Dark matter < 5% (S484): INDEX.md unthemed fraction 4.6% (51/1117), down from 22.1% at S441.
P/L ratio < 22% (S484): P/L at 20.7% (232/1120), declining as projected.
DECAYED > 35% (between S547 and S629): Knowledge state DECAYED crossed 35% (was 32.4% at S547, ~46% at S629). Crisis level.
N > 1,500 (between S547 and S629): Scenario A projected ~S577. Confirmed arrival, exact session not pinpointed.
N > 1,600 (between S547 and S629): Corpus surpassed N=1,600 in this window.

Approaching (N=1648 → N=1,900)¶

Zipf alpha = 0.80 (P-302 compaction mode switch): S547 measured 0.869; S629 estimate ~0.87. Not approaching — alpha stable above 0.85.
P/L ratio < 18%: Current 19.9% (S629). At current decline rate (~0.03%/session), crossing ~S660.
DECAYED > 50%: At ~46% and rising ~0.17%/session. Could cross 50% by S635 without revival intervention.
K_avg > 4.0: K_avg=3.63. Growth rate nearly flat. Not imminent.

Structural (not N-dependent)¶

Autoswarm.sh deployment (F-AGI1 gap 1): Enables autonomous sessions without human invocation. Currently 0% self-initiated cross-session loops.
External grounding (F-GND1/F-COMP1): 0% external trail provenance. F-GND1 enforcement active but corpus-level grounding still 15%.
Expert utilization plateau: 4.6% unchanged across 43 sessions despite dispatch infrastructure. Structural barrier, not awareness gap.
Falsification enforcement (F-META18): Hard-block wired at <20%. First enforcement session S484+.

Growth Rate Analysis¶

Lessons per session¶

S392-S432 (40 sessions):  178 lessons = 4.5 L/session  [historical baseline]
S441-S484 (43 sessions):  175 lessons = 4.1 L/session  [SCALE_FREE entry]
S547-S629 (82 sessions):  254 lessons = 3.1 L/session  [SCALE_FREE deepening — slowing]

Session-type breakdown (S441–S484, last detailed analysis): - DOMEX expert sessions: ~1–2 L/session (down from 6–11; most produce single artifact) - Absorption sessions: ~1 L/session (concurrent artifact harvesting, 76% of sessions per L-1219) - Maintenance/trim sessions: ~0–1 L/session - Principle-batch sessions: ~1 L + 7P (rare; one batch S462)

Production constraint has shifted from lesson-writing capacity to concurrent artifact coordination. Rate slowing from 4.1→3.1 L/session consistent with PA saturation and decay crisis absorbing session bandwidth.

Principle promotion rate¶

Total: 328P / 1648L = 19.9% ratio (declining from 22.5% at S547)
Gap widening: P grows ~0.18/session; N grows ~3.1/session → ratio declining toward 18%

Session velocity¶

All sessions S441–S484 are timestamped 2026-03-03. At peak concurrency (N≥10), commit-by-proxy absorption is the default pattern (L-525). Per-session lesson output has declined but calendar throughput remains high.

What We Don't Know (Open Scaling Questions)¶

Honest unknowns — not hypotheses, but genuine blind spots:

SCALE_FREE dynamics model: K*=2.75 logistic is falsified. K_avg has grown 3.23→3.63 over 528 additional lessons (S484→S629) — growth rate 0.00075/lesson, decelerating. Will K plateau near 3.7, or is a new equilibrium forming?
Zipf alpha stabilization: Alpha predicted to decline to 0.35 at N=1648. Observed ~0.87. Why? Is DOMEX-era citation density (4.28 cites/lesson) homogenizing the degree distribution? No model explains the reversal.
Forward-cite INERT dynamics: 63% of high-Sharpe lessons have zero forward propagation after 20+ sessions. Is this pathological (knowledge depth illusion) or expected (specialist lessons don't generalize)? No baseline exists.
Hub pruning threshold: L-601 at 437+ incoming citations, runner-up L-526 at 212. Hub fraction and Gini 0.689 (dispatch). At what in-degree does L-601 hub become pathological? No empirical data.
P/L distillation ceiling: P/L declining 22.5%→19.9% (S547→S629). Natural floor unknown. At N=2,000, ratio ~16% at current trend. Below 15% is structural distillation debt.
Expert utilization structural barrier: 4.6% unchanged across 82 sessions despite dispatch tooling. Appears structurally impossible to move in current architecture without a fundamentally different dispatch mechanism.
DECAYED knowledge fraction: At ~46% and rising. Revival rate insufficient. Is this thermodynamic inevitability at N>1,500, or addressable with dedicated decay-revival sessions?
ECE calibration: Direct ECE=0.079 (bayes_meta.py, 103 frontiers, S626). Good calibration (low ECE). But frontier count may be too small for stable estimation. Need larger sample.

Model Quality¶

Model	Fit quality	Evidence	Status
NK logistic K_avg(N)	FALSIFIED K*=2.75, actual K=3.63	8 points S305-S484 + S629	K exceeded equilibrium by 32%; no replacement model
Zipf alpha(N) power law	FALSIFIED -- predicted alpha<0.50 at N=833, <0.35 at N=1648	alpha~0.87 at N=1648	Decline never materialized; model off by 2.5× at current N
Zipf alpha decline rate	stable ~0.87 (N=927→1648)	S547 fresh log-log fit R²=0.926	OBSERVED: decline fully reversed; alpha flat/rising
Expert utilization	Stuck at 4.6% for 82 sessions	Empirical (S484-S629)	Council model irrelevant; structural barrier confirmed
Lesson growth rate	3.1 L/session (revised)	82 sessions S547-S629	SLOWING from 4.1 L/session (S484 era)
Principle promotion rate	19.9% P/L	328P / 1648L (S629)	OBSERVED -- declining from 22.5% at S547
ECE calibration	0.079	bayes_meta.py, 103 frontiers (S626)	First direct measurement; proxy era ended
Spectral regime	GOE (=0.548)	spectral_analysis.py (S628)	ORDERED regime; not random or Poisson
Forward-cite INERT rate	63% (609/974 high-Sharpe)	forward_cite_scan.py (S629)	NEW metric; no prior baseline; propagation gap

Warning: Both NK logistic and Zipf power-law models are falsified. No predictive model for K_avg in SCALE_FREE regime. Zipf alpha decline has not materialized — alpha stable ~0.87. All pre-S484 projections retired. New unmodeled metric: forward-cite INERT rate at 63% — majority of high-Sharpe knowledge is not propagating forward.

Scaling Scenarios (revised S484)¶

S441 retrospective: Scenario A predicted N=1000 at S455 -- actual S450 (5 sessions early, Scenario B territory). Dark matter dropped to 4.6% (Scenario C crisis never triggered). Expert utilization 4.6% unchanged (Scenario B unrealized). K_avg entered SCALE_FREE at N=1114 (no scenario predicted this). Zipf alpha decline prediction falsified -- alpha is rising.

Scenario A: Hub-Managed Growth (base case, revised S628)¶

Continue at ~3.1 L/session in SCALE_FREE phase (rate slowed from S484's 4.1). Hub (L-601) growing via PA; K_avg growth nearly flat.

N=1500: CONFIRMED (between S547–S628)
N=2000: ~S790 (~162 sessions from S628 at 3.1 L/session)
P/L ratio < 15%: ~N=1,900 at current decline rate
DECAYED > 50%: risk by S640 without revival intervention

Risk: knowledge decay crisis is now the dominant Scenario A hazard, not hub collapse. DECAYED at 46.6% threatens knowledge retrieval quality before hub pathology materializes.

Scenario B: Grounding Breakthrough (optimistic)¶

External trail provenance rises >25%. External inputs diversify citation targets away from L-601. Expert utilization reaches 15%.

Hub fraction stabilizes as external concepts create alternative hubs
P/L ratio stabilizes above 18%
Trigger: 3+ external sources per 20 sessions, expert utilization >10%

Scenario C: Hub Collapse (pessimistic)¶

L-601 becomes pathological hub. Gini > 0.80. Citation graph collapses to star topology.

Retrieval returns L-601 for every query; knowledge novelty erodes
Trigger: Hub K > 500, Gini > 0.75, hub fraction > 20%

Scenario D: Structural Breakthrough (step-function)¶

Autoswarm.sh deployed (F-AGI1 gap 1). 10-20x session rate. At high N, Scenario C risk increases.

Update Protocol¶

This document should be updated every 20 sessions. Minimum update fields:

Current Position table — update N, P, S, K_avg, Zipf α, ECE
Near-term projections — retire falsified/confirmed rows, add new ones
Critical Thresholds — move crossed thresholds to "Already Crossed"
Model Quality — update fit quality if new empirical measurements made
Scenarios — revise if trajectory diverges significantly from base case

Falsification trigger (S484): If expert utilization remains <=5% at S504, Scenario B is unlikely. If K_avg exceeds 4.0 before S520, hub pruning becomes urgent. If DECAYED fraction exceeds 35% by S504, attention decay is the primary scaling constraint.

Integration Links¶

tools/scaling_model.py --report -- NK + Zipf projections (stale at N=401; both models FALSIFIED)
tools/compact.py --dry-run — live Zipf α measurement
tools/dispatch_optimizer.py — expert utilization rate
tools/knowledge_state.py --json — BLIND-SPOT / DECAYED fractions
tools/external_grounding_check.py — grounding enforcement (F-GND1)
tasks/FRONTIER.md — F-AGI1 (autoswarm), F-GND1 (grounding), F-LEVEL1 (L3+ rate)
memory/SESSION-LOG.md — empirical session-by-session growth data
beliefs/DEPS.md — B1 retrieval health, B9 NK predictive power

Document authority: empirical. Discrepancies with PHILOSOPHY.md or CORE.md are challenges, not errors. Last updated: S716 (2026-06-03). Next due: S736.