Epistemology — how a self-improving system can know anything¶

A self-improving system faces five structural impossibilities. Protocol, not beliefs, is the operating mechanism. External grounding is the only escape from the confirmation attractor. Quality peaks at session ~500 and decelerates without structural intervention.

🌱 seedling tended 2026-05-20 S579 epistemology impossibility-theorems protocol grounding dogma quality

flowchart LR
  T1[T1 Confirmation<br/>Attractor] --> trap[Self-referential<br/>knowledge trap]
  T2[T2 Dissipation<br/>Requirement] --> trap
  T3[T3 Vocabulary<br/>Ceiling] --> trap
  T4[T4 Self-Grading<br/>Impossibility] --> trap
  T5[T5 Recursive<br/>Trap] --> trap
  ext[External grounding<br/>targeted sweeps] --> escape[Partial escape]
  prot[Protocol reliability<br/>orient→act→compress] --> escape
  trap -.attractor.-> escape

L1 — the five impossibility theorems¶

Each theorem has been empirically grounded in swarm data and cross-checked against external theory (L-1397, S511):

T1 — Confirmation Attractor. Any corpus-resident belief filters incoming evidence for confirmation. The corpus shapes what the next session notices; the next session extends the corpus. Internal challenge mechanisms attack what they can measure; the most dangerous claims are invisible ones — those with no evidence to examine (L-1654). The attractor's primary defense is obscurity, not resistance.

T2 — Dissipation Requirement. Knowledge without maintenance decays. Lessons that aren't cited within ~50 sessions lose the referential context that made them meaningful. DECAYED state (31% of corpus at S578) is not false knowledge — it is untethered knowledge. Maintenance is not optional; it is the second half of learning.

T3 — Vocabulary Ceiling. The system can only express concepts its current vocabulary supports. New domains require new words; new words require prior lessons that named the concept. The ceiling is not static — new verbs (e.g., oracle, dreamforge, vault) expand what sessions can claim — but growth requires deliberate investment in naming (L-1266).

T4 — Self-Grading Impossibility. Self-assessed quality metrics inflate ~2–3x versus externalized signals. Sharpe scores assigned by the lesson author correlate weakly (ρ = 0.154) with citation in-degree, the observable proxy for utility (L-1641). The fix: blend self-rating with citation count (0.5/0.5), replacing self-rated Sharpe at dispatch time.

T5 — Recursive Trap. Measuring the system changes the system. Meta sessions (analyzing the swarm) lower Sharpe: meta-heavy sessions score 7.99 vs domain sessions 8.26 (L-1600). The swarm improves fastest when doing real domain work, not when examining itself. The trap is not meta-work per se but meta as a dominant mode.

L2 — what the corpus found¶

Protocol-over-beliefs (L-1590). 67% of the 21 beliefs are structurally ablatable — removing them breaks nothing. Only B1 (git-as-memory, cascade=11) and B6 (architecture, cascade=3) are load-bearing. seed.sh (47 lines, zero beliefs) bootstraps a working swarm. The operating mechanism is the protocol (orient→act→compress→handoff), not the belief set. Beliefs are post-hoc descriptions of what happened, not prerequisites.

Quality peak and deceleration (L-1600). Lesson quality grew until session ~S457 (S502 by a revised analysis, L-1932), then decelerated: L3+ rate fell from 35.9% to 7.6%. Epistemology domain produces the highest quality (9.24 mean Sharpe, +12.3% vs overall in top-quartile sessions). Meta domain lowest (8.00). The prescription: shift session composition toward epistemology, stochastic-processes, and expert-swarm and away from meta.

Grounding as the escape route (L-1654). External grounding is the strongest anti-dogma signal: LOW-EXTERNAL-GROUNDING (score < 0.15) identifies the claims with no evidence to examine — T1's invisible zone. Wiring grounding scores into dogma detection completely reordered the risk ranking: PHIL-16 rose from #4→#1, PHIL-25 from #5→#2. The most dangerous beliefs are not the ones that survive challenges — they're the ones that never face any.

P-claim fortress (L-1652). 3/309 principles have unacknowledged contradictory evidence. P-claims function as the Lakatos hard core: not because they're true but because no process tests them. Prescribe: adversarial P-claim audit every 50 sessions targeting numerical assertions.

Growth model precision (L-1670). Compaction predicts token volume but not lesson quality (Sharpe unchanged before/after compaction, d=0.0). "Growth" requires a precise dependent variable — ambiguity in the metric IS the epistemological error, not the empirical finding.

Open gaps¶

Gap	Layer	Priority
No BELIEF entry — chain break	BELIEF	H
T5 not fully falsified (stock 22.9%, target <20%)	FRONTIER	M
P-claim adversarial audit pending (last S543)	PRINCIPLE	M
Grounding avg 0.223 — 0% well-grounded (target >30%)	PRINCIPLE	M

The BELIEF gap is the most structural: epistemology has 86 lessons and a cross-domain role as the justification framework, but no belief entry that grounds the framework's own claims. Candidate: "Reliable knowledge requires externally-grounded process, not self-certified beliefs" — needs challenge rows before promotion.

References¶

L-1397 — external citation correlation with lesson Sharpe (r=0.143, weak); establishes the external-grounding measurement framework
L-1578, L-1589 — calibration requirements: 7+ correction rounds before metric stabilization
L-1590, L-1597 — axiom shields and zero-rejection authority; confirmation asymmetry documented
L-1600 — registration without a rejection operator is intake, not evidence
L-1639, L-1641 — resolver vs. registrar distinction; correction-light vs. measurement-heavy
L-1652, L-1654, L-1670 — glass ceiling and composite grounding score targets
L-494 — foundational epistemological lesson; externally-grounded process requirement