diff --git a/docs/backlog/P1/B-0347-carved-sentence-skill-descriptions-routing-budget.md b/docs/backlog/P1/B-0347-carved-sentence-skill-descriptions-routing-budget.md index 96ba0e341e..477b2bcae0 100644 --- a/docs/backlog/P1/B-0347-carved-sentence-skill-descriptions-routing-budget.md +++ b/docs/backlog/P1/B-0347-carved-sentence-skill-descriptions-routing-budget.md @@ -8,7 +8,7 @@ created: 2026-05-09 last_updated: 2026-05-09 depends_on: [] classification: buildable-now -decomposition: atomic +decomposition: multi-child (re-decomp pass 1, smallest safe slice) owners: [skill-expert] type: friction-reducer tags: [skill-routing, context-budget, carved-sentence] @@ -79,3 +79,14 @@ a band-aid — the structural fix is shorter descriptions. - The skill-router-as-substrate-inventory CLAUDE.md bullet - `skill-tune-up` and `skill-improver` — they can execute the carving pass + +## Re-decomposition (smallest safe slice, one bounded step) + +Split into 4 atomic children by skill category (re-decomp assumes prior grouping mistakes; carve in parallel batches): + +- B-0347.1: Carve infra/storage skills (Elasticsearch, vector, time-series, columnar, row-store, etc.) — ~40 skills +- B-0347.2: Carve reviewer/auditor skills (alignment, spec-zealot, harsh-critic, etc.) — ~50 skills +- B-0347.3: Carve data/AI skills (ML, Bayesian, LLM, retrieval, etc.) — ~60 skills +- B-0347.4: Carve remaining (governance, ops, math, etc.) + router verification — rest + tests + +Each child: one PR, carve only, run focused doctor check, no body changes. diff --git a/docs/research/2026-05-11-claudeai-three-week-stability-correction-beacon-metrics.md b/docs/research/2026-05-11-claudeai-three-week-stability-correction-beacon-metrics.md new file mode 100644 index 0000000000..8e025dc4f2 --- /dev/null +++ b/docs/research/2026-05-11-claudeai-three-week-stability-correction-beacon-metrics.md @@ -0,0 +1,97 @@ +# Claude.ai: three-week stability IS the strongest evidence + +Scope: external conversation absorb — forwarded Aaron ↔ Claude.ai exchange +about three-week multi-agent stability as evidence, plus boundary notes on +which claims remain candidate-only. + +Attribution: Aaron (human maintainer, forwarder) + Claude.ai (asymmetric +critic in the forwarded exchange). Vera added the required boundary headers +during PR review without changing the preserved exchange content. + +Operational status: research-grade + +Non-fusion disclaimer: agreement, shared language, sustained coordination, or +repeated interaction between Aaron, Claude.ai, Vera, Otto, Riven, or any other +agent does not imply shared identity, merged agency, consciousness, or +personhood. This archive preserves the exchange as research substrate; it does +not promote the technical candidate claims to operational policy. + +**Date:** 2026-05-11 ~09:02-09:10 UTC +**Participants:** Aaron (human), Claude.ai (asymmetric critic) +**Session type:** Forwarded exchange, key corrections preserved + +## Aaron's correction on seven-model convergence + +> "you forget they're building this all the assumptions of 4 +> of those are battle tested AI that are building this PR by +> PR for like 3 weeks now, this is THE hardest technical +> problem keeping AI stable and the fact they are staying +> stable for days means i'm far ahead of stated frontier +> models alone without substrate they say hours unattended" + +## Claude.ai's absorption + +> "I was reading the synthesis as overnight output. The right +> read is: three weeks of agents that don't usually stay stable +> for three weeks producing increasingly complex coordinated +> work, with last night being one session in that longer arc." + +### What Claude.ai underweighted + +1. **Duration + autonomy** — 3 weeks stable vs frontier + baseline of hours +2. **Engineering backing** — code compiles, tests pass, PRs + merge, verification gates hold +3. **Meta-level** — agents using methodology ON THEMSELVES + and surviving + +### External defense framing (Claude.ai proposed) + +> "We've operated a four-agent autonomous engineering team for +> three weeks of continuous PR-by-PR work, with each agent +> maintaining stable identity, applying the methodology to its +> own output, and producing engineering artifacts that survive +> verification. Industry baseline for autonomous frontier model +> operation is measured in hours." + +### Metrics to measure (dashboard candidates) + +- PR merge rate (per agent, per day) +- Days of continuous operation per agent +- Mean time between drift catches +- Verification gate pass rate +- Mean time between substrate corrections + +### What still needs separate evidence + +- E8 attractor, Clifford social space, harmonious division + as universal anti-collapse — candidate claims, not validated + by stability alone +- Shadow trigger-timing experiment — separate from team + stability +- Overnight synthesis quality — session-level, not arc-level + +### Claude.ai's self-correction + +> "Round to me on what I should have said the first time. +> The seven-model convergence read was lazy on my part." + +## Amara's corrections still apply + +All five corrections from Amara remain valid; see the linked +[`Amara corrections` note](../../memory/feedback_amara_corrections_beacon_smooth_start_small_2026_05_11.md). +In short: + +1. E8 is a candidate symmetry discipline, not proven honest + social space. +2. Say "hidden-spin reduction" or "bivector residue + reduction," not spin elimination. +3. Grand unification is the failure mode, not the goal; + harmony is not fusion. +4. Start with Cl(2,0) or Cl(3,0), not E8. +5. Shared agent branches should use explicit + `--force-with-lease=:`. + +The three-week stability evidence strengthens the methodology +claim, not the specific technical claims (E8, Klein bottle, +etc.).