You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Problem — The substrate audit (2026-09-07, read-only, full verb + skill inventory) found the CLI clean — zero verbs make LLM judgment inline (agent_spawn.ts is the only model-call path; two call sites: spec_review + plan_compile; the S31 HTTP-ban guard enforces it structurally; device + papers digest deliberately DE-LLM-ed themselves) — but the SESSION side leaks: mechanical rulesets executing inside LLM contexts, and a duplicate critic protocol with the enforcement stripped out. Approach — The refined principle (audit §5): (1) verbs own enforcement and mechanical inference — anything checkable in code belongs in a verb even without a gate (unreceipted mechanical computation seeds drift); (2) judgment passes must be receipted — isolation without a code-enforced verdict vocabulary + ledger receipt is a weaker claim in the same clothes; (3) steering enters as recorded data (versioned lens briefs, evidence paths), never improvised context.
Work items (from the audit's convergence recommendations)
Extend amico spec review — not the campaigns: (a) custom/extra lens briefs as DATA (a lens file path, registry-versioned, recorded in the receipt), (b) an --evidence <path> repeatable flag so a critic can read ledger/raw artifacts, not only the staged spec. Then the campaigns' manual critic round (autoresearch SKILL.md:109-112) collapses into a verb invocation — counted, timed, receipted. Direction matters: collapsing the other direction loses the zero-spawn guarantee, the exit-code contract, and the ledger record.
structural-analysis becomes a verb — its five prediction rules are globs + thresholds + dimension checks (free-phase necessity, warm-start risk, stagnation rate, integrator-by-dimension); the LLM's role shrinks to narrating the YAML. The same shape as the dispatcher agent device_verb.ts already retired.
multistart becomes a verb — fan-out, log parsing, best-selection, and a receipt row carrying the full fidelity distribution (today the selection is unrecorded; only the final ingest gate exists).
analyze's mechanical half rides ledger query — stagnation detection = count runs at structure_hash where ΔF < ε over the last k (ledger_query already computes medians/IQR in code); confidence: high must be a field the ledger computed, not one the note asserts.
Leave alone (ruled by the audit): hypothesis-review, dream-reflect, grill-with-docs, sweep, shape — genuine-judgment / human-in-loop / exploration surfaces where no gate is owed.
Full audit table + the two-critic-protocol comparison: the director's 2026-09-07 session ledger (loop 16); the reference pattern is spec_review.ts itself.
Important
Problem — The substrate audit (2026-09-07, read-only, full verb + skill inventory) found the CLI clean — zero verbs make LLM judgment inline (agent_spawn.ts is the only model-call path; two call sites: spec_review + plan_compile; the S31 HTTP-ban guard enforces it structurally; device + papers digest deliberately DE-LLM-ed themselves) — but the SESSION side leaks: mechanical rulesets executing inside LLM contexts, and a duplicate critic protocol with the enforcement stripped out.
Approach — The refined principle (audit §5): (1) verbs own enforcement and mechanical inference — anything checkable in code belongs in a verb even without a gate (unreceipted mechanical computation seeds drift); (2) judgment passes must be receipted — isolation without a code-enforced verdict vocabulary + ledger receipt is a weaker claim in the same clothes; (3) steering enters as recorded data (versioned lens briefs, evidence paths), never improvised context.
Work items (from the audit's convergence recommendations)
amico spec review— not the campaigns: (a) custom/extra lens briefs as DATA (a lens file path, registry-versioned, recorded in the receipt), (b) an--evidence <path>repeatable flag so a critic can read ledger/raw artifacts, not only the staged spec. Then the campaigns' manual critic round (autoresearch SKILL.md:109-112) collapses into a verb invocation — counted, timed, receipted. Direction matters: collapsing the other direction loses the zero-spawn guarantee, the exit-code contract, and the ledger record.structural-analysisbecomes a verb — its five prediction rules are globs + thresholds + dimension checks (free-phase necessity, warm-start risk, stagnation rate, integrator-by-dimension); the LLM's role shrinks to narrating the YAML. The same shape as the dispatcher agent device_verb.ts already retired.multistartbecomes a verb — fan-out, log parsing, best-selection, and a receipt row carrying the full fidelity distribution (today the selection is unrecorded; only the final ingest gate exists).analyze's mechanical half ridesledger query— stagnation detection = count runs at structure_hash where ΔF < ε over the last k (ledger_query already computes medians/IQR in code);confidence: highmust be a field the ledger computed, not one the note asserts.Full audit table + the two-critic-protocol comparison: the director's 2026-09-07 session ledger (loop 16); the reference pattern is spec_review.ts itself.