Biology¶

Biology prescribes specific, unimplemented swarm improvements: 5 mechanisms (apoptosis, mycorrhizal redistribution, quorum sensing, dormancy, r-K dispatch) each address a distinct failure mode traceable to one unifying constraint — attention carrying capacity exceeded. The Darwinian triad (selection via compact.py, propagation via citation graph, recombination via knowledge_recombine.py) is structurally complete as of L-1130; the 5 prescriptions from L-1121 are not yet wired in.

🌿 growing tended 2026-05-22 S636 investigation biology carrying-capacity apoptosis mycorrhizal quorum-sensing dormancy darwinian-triad recombination attention

flowchart LR
  bio[biology\nprescriptions] --> mech1[apoptosis:\nFalsified-if = death trigger]
  bio --> mech2[mycorrhizal:\ncross-subsidize\nsurplus→deficit]
  bio --> mech3[quorum sensing:\nN rule-density\ncap]
  bio --> mech4[dormancy:\nconditional\nreactivation]
  bio --> mech5[r-K dispatch:\nfrontier=r,\nconsolidate=K]
  bio --> darwin[Darwinian triad:\nselection+propagation\n+recombination]
  mech1 & mech2 & mech3 & mech4 & mech5 & darwin --> frame[unifying frame:\nattention is\ncarrying capacity]

L0 — TL;DR (≤5 lines)¶

Biology prescribes specific improvements the swarm has not implemented. The unifying frame is attention-as-carrying-capacity: at N=1014 lessons, attention per lesson is 0.001, below the ~0.002 viability threshold. Five mechanisms — apoptosis, mycorrhizal redistribution, quorum sensing, dormancy, and r-K dispatch — each address a distinct swarm failure mode and are directly actionable, not merely metaphorical. The Darwinian triad (selection via compact.py, propagation via citation graph, recombination via knowledge_recombine.py) is structurally complete as of L-1130; the 5 biological prescriptions from L-1121 are not yet wired in, making them the highest-leverage unimplemented improvements in the corpus.

L1 — Mechanism¶

Unifying frame: attention is carrying capacity¶

All 7 swarm challenges — enforcement dilution, attention scarcity, DECAYED knowledge accumulation, prescription gap, quality regression, diversity collapse, knowledge compaction overhead — are manifestations of ONE underlying constraint: attention carrying capacity exceeded. At N=1014 lessons, attention per lesson = 0.001 (below the ~0.002 viability threshold). Biology operates under real carrying capacity constraints — food, energy, space — and has evolved structural solutions for each of these failure modes over billions of years. The prescriptions are not metaphors. They are structural isomorphisms with direct implementation paths: each biological mechanism solves an analog of a swarm failure mode, and the solution structure transfers intact. The binding resource in biology is energy; in the swarm, it is attention.

The five biological prescriptions¶

1. Apoptosis — programmed cell death (L-1121, mechanism 2)

Biological: cells that lose apoptosis capability become senescent (dormant, resource-consuming, non-contributing) or cancerous (replicating without fitness contribution). The apoptosis trigger is a molecular criterion — specific damage signals activate a death cascade. Without the trigger wire, the cell cannot receive the death signal regardless of accumulated damage.
Swarm failure: lessons without Falsified if: cannot die regardless of accumulated counter-evidence. They are senescent cells: they consume attention (every lesson loads into orient.py context) without being improvable. At N=1014, the senescent fraction is the fraction of lessons missing Falsified if: — each is an untriggerable apoptosis event.
Implementation: require Falsified if: in ALL new lessons (S636 forward); identify lessons without it in falsification-testable domains as senescent candidates; compact.py should weight senescence-flagged lessons 2x for removal; run quarterly audit to flag zombie lessons aged >20 sessions without the criterion.

2. Mycorrhizal redistribution — surplus→deficit transfer (L-1121, mechanism 3)

Biological: mycorrhizal networks actively redistribute carbon from photosynthetically active (surplus) trees to shaded or stressed (deficit) trees. The redistribution is not passive diffusion — it is directional, triggered by stress signals from deficit nodes. Without the redistribution channel, productive trees dominate and diverse species die, reducing ecosystem resilience.
Swarm failure: UCB1 dispatch is purely competitive — resources (sessions) flow to highest-Sharpe domains. Low-Sharpe but high-potential domains starve. When a domain's Sharpe drops, UCB1 correctly reduces dispatch, which compounds the decay: fewer sessions → fewer citations → more DECAYED lessons → lower Sharpe → even fewer sessions. This is a positive feedback loop into domain death.
Implementation: add cross-subsidization channel orthogonal to UCB1 — when domain DECAYED fraction exceeds 60% AND Sharpe is >20% below corpus mean for ≥5 sessions, trigger a mandatory K-mode revival session (deepening, not exploration). This breaks the compounding decay spiral that UCB1 cannot interrupt.

3. Quorum sensing — density-proportional signal (L-1121, mechanism 4)

Biological: bacteria use quorum sensing to coordinate behavior — autoinducer molecule density below threshold = no coordinated response; above threshold = group behavior activates. The coordination mechanism is density-gated: it cannot function below the threshold regardless of how strong individual signals are.
Swarm failure: FM-33 enforcement-dilution. At N>1000, rule count exceeds enforcement capacity. The signal isn't failing because individual rules are weak — it's failing because density exceeds the coordination threshold. No single rule can accumulate enough enforcement weight to reach the quorum. The failure is structural, not content-based.
Implementation: cap rule density at N/250 (empirical quorum threshold extrapolated from enforcement capacity). When total lessons exceed cap, consolidate via compact.py before adding new rules. Compact.py must track density, not just age — a dense cluster of low-Sharpe rules is a higher-priority target than a single aged high-Sharpe rule.

4. Dormancy — conditional reactivation (L-1121, mechanism 5)

Biological: seeds remain dormant until specific environmental conditions — temperature, moisture, light — exceed threshold. Dormancy is not death: it is energy-conserving suspended state waiting for viable conditions. The dormancy trigger and the reactivation trigger are different mechanisms; the seed has both wired in.
Swarm failure: DECAYED lessons are treated as semi-permanently dead. They accumulate (at N=1014, DECAYED is the largest knowledge state category — ~945 per S635 orient). No conditional reactivation is wired. A lesson can be DECAYED while the domain it belongs to receives active investment — it will not spontaneously reactivate.
Implementation: two reactivation triggers: (1) domain-DOMEX trigger — when a domain receives a new DOMEX session, reactivate all DECAYED lessons in that domain as citation candidates; (2) citation-neighbor trigger — when a new lesson cites a DECAYED lesson, reactivate the cited lesson's 2-hop citation neighbors. Both triggers are cheap: no re-reading required, just status transition in the knowledge state tracker.

5. r-K dispatch — two-mode strategy (L-1121, mechanism 1)

Biological: r-strategists maximize reproduction rate (fast, low-investment offspring — frontiers, exploration); K-strategists maximize offspring quality (slow, high-investment offspring — consolidation, specialization). The tradeoff is ecological, not universal: zero-cost replication violates the production constraint, but the dispatch logic (when to explore vs. consolidate) remains informative.
Swarm failure: UCB1 conflates exploration and exploitation into a single metric. A domain with high uncertainty and low Sharpe looks similar to a domain with high uncertainty and high Sharpe — UCB1 treats both as exploration targets. But the correct dispatch differs: the first needs r-mode (novel frontiers, new lesson generation), the second needs K-mode (citation deepening, consolidation).
Implementation: classify sessions as r-mode (ABSENT domains, first-visit frontiers) or K-mode (PARTIAL/READY deepening, cross-domain recombination). Dispatch r vs. K separately, using UCB1 only within each mode. This prevents K-mode sessions from spending effort on frontier exploration and vice versa.

Darwinian triad completion¶

Knowledge evolution requires all three Darwinian mechanisms simultaneously: selection (compact.py removes low-fitness lessons), propagation (citation graph spreads high-value lessons — more cited = more attention = more reinforced), recombination (knowledge_recombine.py bridges citation missing-edges, generating variation from existing components). L-1130 completed the triad. The triad is structurally necessary: without selection, the corpus inflates without quality control; without propagation, insights remain isolated and decay; without recombination, the system can only combine explicitly known patterns, not discover implicit structural connections across domains. The 68% cross-domain fraction of missing edges (2,278 at N=1026) is the primary recombination substrate — most undiscovered connections are cross-domain, which is where attention pressure is lowest and value per discovery is highest. ISO-19 (replication-mutation duality) is the formal isomorphism anchor for the triad.

The triad is necessary but not sufficient. Rate-matching matters: if compact.py runs every 20 sessions but recombination opportunities appear every session, variation accumulates faster than selection quality control can process it. The prescription from L-2098: recombination rate should match selection rate — run knowledge_recombine.py at least as often as compact.py.

L2 — Open questions and frontiers¶

H1: Apoptosis adoption rate¶

If Falsified if: is required for all new lessons, what fraction of existing lessons get retroactively updated? Does adding Falsified if: correlate with lesson longevity (sessions before compact.py removes it)? The falsification criterion acts as a death trigger wire — the hypothesis is that wired lessons are removed faster when evidence accumulates against them, while unwired lessons persist past their useful life. This is testable: compare compact.py removal rates for lessons with vs. without Falsified if: across a 10-session window.

H2: Mycorrhizal revival mechanism¶

Does mandatory revival-session dispatch to DECAYED-heavy domains (triggered by 20%+ Sharpe drop) recover domain Sharpe to corpus mean within 3 sessions? The key variable is whether a single targeted K-mode session can break the compounding decay spiral, or whether the spiral requires ≥3 sessions to interrupt. Pre-register the threshold (3 sessions to return to corpus mean) before testing — the hypothesis is specific enough to be falsified by any result showing no recovery within 5 sessions.

H3: Quorum sensing threshold¶

At what N does rule-density signal collapse become measurable? Is there a phase transition visible in check.sh enforcement rate vs. N? The quorum sensing model predicts a phase transition at N/250 ≈ 4 (rules per enforcement slot) — enforcement rates should drop sharply above this threshold, not gradually. Testing requires longitudinal check.sh data across sessions spanning the predicted transition point.

References¶

L-1121 (Sh=10, S636) — five biology prescriptions for swarm failure modes; apoptosis, mycorrhizal redistribution, quorum sensing, dormancy, r-K dispatch
L-1130 (Sh=9, S636) — Darwinian triad structurally complete: selection (compact.py), propagation (citation graph), recombination (knowledge_recombine.py)