Empirical contract-card ablation studies, successor-operator amortization protocols, and cognitive decision research converge on sharp operational gates. Teleon encodes nine formal decision rules directly into queues aq23..aq35.
Every Persistent Store item exposes a <=50-token SKOS L0 scent card and Expected Value of Information (bits per 100 tokens) before any agent fetches full bytes.
Retrieval never dumps 50 raw hits into an agent prompt; Stage 1 screens 5,000+ items via Roaring bitmaps into a <=5-item Consideration Set.
Non-compensatory B_NEG bitmaps immediately veto incompatible licenses, PII risks, wrong runtime OS, or known regression signatures regardless of similarity score.
A reusable Contract Card or skill is promoted only when min(per_model_deltas) >= 0 across tested model families and mutation-kill score >= 0.80.
When baseline pass probability >= 0.98 and task ambiguity < 0.15, Teleon bypasses context injection completely, avoiding +62% latency overhead.
Every decision record carries a decay rate lambda and explicit review_trigger predicates so stale assumptions automatically enqueue re-verification.
Capability claims are audited under identical tool/prompt scaffolds on uncontaminated tasks, flagging custom-wrapper inflation.
Compares retained operators vs rebuild, conventional cache, and direct expert construction; keeps historical evidence, current qualification premises, and production authority strictly separate.
Gates intrusive inspection or high-impact escalation behind explicit evidence thresholds and logs an immutable docket of both why the system looked and why it held back.
aq01..aq61), version-fenced stores (S01..S278), 3-Receipt substrate separation, retained failure tombstones, and 0-call / multi-harness test gates.aq28)When baseline pass probability is already saturated (≥ 0.98) and ambiguity entropy is low (< 0.15), injecting extra instructions adds up to +62% latency for zero accuracy gain. aq28 returns BYPASS_INJECTION_BASELINE_SATURATED.
≤ 2,048B)Cards exceeding 2,048 bytes or injected mid-turn disrupt reasoning trajectories. aq28 enforces strict byte caps and pre-turn placement.
aq32)Retained operators beat rebuilding under affine drift (-12.69% calls), tie a conventional cache under static premises (KEEP_PRESENT_DESIGN_CACHE_SUFFICIENT), and halt when unamortized on one-offs or out of bracket.
aq27)An instruction card that lifts Model A by +0.08 while regressing Model B by -0.149 is blocked. Promotion requires min(per_model_deltas) ≥ 0 and mutation-kill score ≥ 0.80.
aq31)Capability audits hold the tool scaffold, retry budget, and token ceiling strictly identical across candidates on uncontaminated holdouts, isolating true lift from wrapper gaming.
aq35)Logs immutable dockets in Store S248 recording both why the system escalated for human co-sign and why it deliberately held back when proportionality was unmet.