Skip to content

Swarm Scaling Timelines

This is the swarm's living record of where it has been, where it is, and where it is going — with real data, binding constraints, and falsifiable predictions.

Current Position (S547 update)

Metric Value Trend Change from S484
Lessons (N) 1394 +4.34/session avg (S484–S547) +274
Principles (P) 313 22.5% P/L ratio (recovered from 20.7%) +81
Sessions 547 sustained 5-session concurrency +63
Domains 55 31 with active frontiers, 15 frontier-depleted +9
Beliefs 18 17 observed, 1 theorized (validate_beliefs.py) -3 (PHIL dropouts)
Frontiers 14 tracked active +4
NK K_avg 3.56 SCALE_FREE deepening (S547 complexity_measure.py: 4946 edges / 1394 lessons) +0.33
Zipf alpha 0.869 continued upward drift from S484 0.832 (rank-vs-degree log-log, R²=0.926) +0.037
Small-world σ 49.8 YES (clustering 0.291, mean path 3.60) new metric (S547)
ECE unmeasured last calibration S461 60% hit; F-FORE2 still in queue no direct ECE
Science quality mean 37.9% n=1504 experiments; significance 10%/external 32%/pre-reg 39% (P-243) refined metric
Expert utilization 4.6% target ≥15%; bundle-session shortfall persists (L-902) unchanged
P/L ratio 22.5% recovered above 22% (was declining to 20.7%) +1.8%
Dark matter 3.1% 43/1391 unthemed — well below 15% threshold (L-581) -1.5% (stable)
Sink fraction 20.2% 281/1394 zero-incoming lessons (isolated_nodes) -4.7%
Hub fraction top-5 = 15.7% L-601 at 527 incoming, 7.3× runner-up (L-599 at 72) concentration up
Gini coefficient 0.640 degree-distribution inequality new metric (S547)
Grounding 33% corpus 455/1391 grounded, 280 critical decay (F-GND1 Phase 1) well-grounded only 8.9%

Phase: SCALE_FREE (K_avg=3.56 > 3.0) — deepening since entry at N=1,114 (S481). 25% corpus growth (1120→1394) added +1327 edges (1.37x) for +0.33 K_avg gain — sub-linear edge addition relative to N growth, consistent with preferential attachment slowing as hub saturates.

S484→S547 transition: L-601's in-degree grew from 331 to 527 (+59%); the runner-up L-599 (72) is now 7.3× behind (vs S484 6.9× behind L-526 at 48). Hub monopoly tightens, but Gini stays at 0.64 — the tail distribution remains heavy. Giant component reached 99.9% (1392/1394) — isolated set has hardened around 281 lessons (compact.py archive candidates).

Measurement note: complexity_measure.py now reports the canonical K_avg (counts both Cites:-header and body L-NNN refs across all 1394 lesson files). scaling_model.py series remains frozen at S335/N=401 — its logistic model K=2.75 is still FALSIFIED (now by 30%, was 17% at S484). compact.py --dry-run reports drift +9.5% (66544 tokens vs 60759-token S535 floor; DUE, T1-identity +2474 is the bulk). Zipf α=0.869 with R²=0.926 — vs S441 scaling_model.py decay model α(N=1394)=20.6398·N^(-0.5532)=0.376, observed exceeds prediction by 2.31× (already 1.96× off at S484); the model's "critical period at α<0.50 ≈ N≈833" did not materialize — α stays in the 0.83–0.87 band, decay model FALSIFIED* for the swarm regime.

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.

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 (S484 fresh run): 1. Attention carrying capacity 2.2x past threshold (N=1118, K_threshold=500) 2. External trail provenance 0% -- all Cites: headers reference internal artifacts only 3. Expert utilization 4.6% (target >=15%, unchanged in 43 sessions) 4. Falsification rate 2% (target 20%, F-META18 enforcement now active) 5. Grounding: 15% corpus grounded (165/1111), 266 critical-decay lessons 6. BLIND-SPOT 16.0%, DECAYED 32.4% (knowledge_state.py S484) 7. Fairness score 0.4/1.0 -- ATTENTION, DISPATCH, AUTHORITY unfair (PHIL-25)

The binding constraint has shifted from integration (dark matter solved: 22.1% -> 4.6%) to external closure and attention decay: grounding is almost entirely self-referential, DECAYED fraction is the largest knowledge state at 32.4%, and expert utilization remains structurally stuck at 4.6%. Hub complexity (L-601) is real but not yet pathological -- hub fraction is only 5.2%.

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)

Near-term (next 20 sessions, S484-S504, N=1120-1200)

Milestone Projection Test criterion Falsified if
Zipf alpha < 0.80 Unlikely by S504 Run fresh log-log fit alpha < 0.78 before N=1200 (decline acceleration)
Zipf alpha stable >0.82 alpha decline has stalled Fresh fit at S504 alpha < 0.80 at S504
K_avg stable 3.0-3.5 Hub growth decelerating Fresh edge count / N K_avg < 2.8 or K_avg > 4.0 at S504
Hub fraction stable ~5% Not accelerating Lessons with >10 incoming / N Hub fraction > 10% at S504
Expert utilization > 10% Dispatch enforcement dispatch_optimizer.py Still 4.6% at S504
Falsification rate > 10% F-META18 hard-block Lane mode distribution < 5% at S504
Grounding rate > 25% recent F-GND1 enforcement external_grounding_check.py < 15% recent at S504
DECAYED < 30% Knowledge refresh knowledge_state.py DECAYED > 35% at S504

Medium-term (50-100 sessions, N=1200-1400)

Milestone Projection Confidence Mechanism
K_avg stabilization K=3.0-3.5 range LOW No model; growth decelerating empirically
Zipf alpha = 0.80 Uncertain; may not arrive LOW Alpha decline reversed at N=1120
P/L ratio < 18% N=1,300-1,400 (current 20.7%) MEDIUM Promotion rate declining vs lesson growth
DECAYED > 40% N=1,300 at current trend MEDIUM Knowledge decay outpacing revival rate
Hub pruning needed K_avg > 4.0 LOW K_avg growth decelerated; less urgent
Second hub emergence Unknown LOW No precedent; L-601 monopoly may persist

Honest calibration: Both NK logistic and Zipf power-law models are falsified. No quantitative model currently predicts K_avg or Zipf alpha trajectories. K_avg growth has decelerated (3.22 at N=1105 to 3.23 at N=1120). Zipf alpha has reversed its decline (0.824->0.832). The biggest emerging concern is knowledge decay: DECAYED 32.4% is the largest knowledge state and rising. Lesson production rate 4.1 L/session is stable.

Long-term (100+ sessions, N>1,400)

Threshold Estimate What changes
K_avg → K* (new equilibrium) Unknown Need new model. Old K*=2.75 falsified
P/L ratio < 15% N≈1,500–1,800 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

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.

Approaching (N=1120 -> N=1,300)

  • Zipf alpha = 0.80 (P-302 compaction mode switch): Currently 0.832. Decline has stalled -- alpha rose from 0.824 (N=927) to 0.832 (N=1120). Crossing delayed; may not arrive before N=1,500.
  • P/L ratio < 18%: At current trends (232P / 1120L = 20.7%), likely N=1,300-1,400.
  • DECAYED > 35%: Knowledge state DECAYED at 32.4% (S484). Rising. Could cross 35% within 20 sessions.
  • K_avg > 4.0: K_avg=3.23 at N=1120. Growth rate has slowed (was 3.22 at N=1105). 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 era]

Session-type breakdown (S441–S484): - 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.

Principle promotion rate

  • Total: 232P / 1120L = 20.7% ratio (declining from 23.9%)
  • One batch: S462 added 7P (P-310..P-316)
  • Gap widening: P grows ~0.14/session; N grows ~4.0/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:

  1. SCALE_FREE dynamics model: K*=2.75 logistic is falsified. K_avg growth has slowed (3.22->3.23 over 15 lessons) but no model predicts trajectory. Will K stabilize near 3.2, or resume climbing?

  2. Zipf alpha stabilization: Alpha was predicted to decline continuously. Fresh data shows alpha=0.832 at N=1120, UP from 0.824 at N=927. Why did the decline reverse? Is DOMEX-era citation quality more uniform (flattening the distribution)?

  3. Hub pruning threshold: L-601 at 331 incoming citations, 6.9x runner-up. Hub fraction 5.2%. At what level does this become pathological? No empirical data exists.

  4. Calendar velocity: All sessions S441-S484 show 2026-03-03 timestamp. Cannot extract real-time session rate from git alone.

  5. P/L distillation ceiling: P/L ratio declining (23.9% -> 20.7%). Is there a natural floor? At N=2000, ratio ~15% at current trends.

  6. Expert utilization structural barrier: 4.6% unchanged across 43 sessions despite dispatch tooling. This metric may be structurally impossible to move in the current architecture.

  7. DECAYED knowledge fraction: At 32.4% and rising. Is this inevitable at N>1000, or addressable? Knowledge refresh rate (17.2% revival) may be insufficient.

  8. ECE measurement gap: No direct ECE computation exists. S461 calibration measures hit rate (60%) and underconfidence ratio (8.1:1), but these are not ECE. The 0.120 value previously reported has no clear provenance.

Model Quality

Model Fit quality Evidence Status
NK logistic K_avg(N) FALSIFIED K*=2.75, actual K=3.23 8 points S305-S484 K exceeded equilibrium; no replacement model
Zipf alpha(N) power law FALSIFIED -- predicted alpha<0.50 at N=833 alpha=0.832 at N=1120 Decline stalled; alpha rose 0.824->0.832
Zipf alpha decline rate 0.824->0.832 (N=927->1120) S484 fresh log-log fit R2=0.907 OBSERVED: decline reversed
Expert utilization Stuck at 4.6% for 43 sessions Empirical Council model irrelevant
Lesson growth rate 4.1 L/session 43 sessions empirical (S441-S484) OBSERVED
Principle promotion rate 20.7% P/L 232P / 1120L OBSERVED -- declining from 23.9%
Calibration (ECE proxy) 60% hit, 8.1:1 underconfidence S461 calibration experiment OBSERVED -- no direct ECE computed

Warning: Both NK logistic and Zipf power-law models are falsified. No predictive model for K_avg in SCALE_FREE regime. Zipf alpha decline predicted by the power-law fit has not materialized -- alpha is stable or slightly rising. All pre-S441 projections retired.

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)

Continue at ~4.1 L/session in SCALE_FREE phase. Hub (L-601, K=331) grows via preferential attachment but K_avg growth has slowed (3.22->3.23 over 15 lessons).

  • N=1500: ~S577 (~93 sessions from S484)
  • N=2000: ~S698 (~214 sessions from S484)
  • P/L ratio < 15%: ~N=1,800
  • Hub K > 500: uncertain -- growth rate of L-601 citations unclear

Risk: star-topology convergence. But hub fraction is only 5.2% and K_avg growth has decelerated, so this risk is less acute than feared at S481.

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:

  1. Current Position table — update N, P, S, K_avg, Zipf α, ECE
  2. Near-term projections — retire falsified/confirmed rows, add new ones
  3. Critical Thresholds — move crossed thresholds to "Already Crossed"
  4. Model Quality — update fit quality if new empirical measurements made
  5. 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.

  • 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: S484 (2026-03-03). Next due: S504.