David/dotfiles
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

🔧 Auto-commit from sysupdate on 2026-05-29 09:43

Browse Source
This commit is contained in:
David F Glidden
2026-05-29 09:43:53 +02:00
parent ec6e414867
commit fffcf173c1
32 changed files with 2387 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
PENDING.md
+64 -1
View File
@@ -702,4 +702,67 @@ The Jurist's six-phase authorization map governs sequencing. Cross-cutting succe
**What works:** the literal-question discipline (structurally enforced; survives compression).
**What doesn't yet:** hook-aware deposit detection; pause-statement symmetry; thread validity gate; Symmetria lineage anchor in pulse; §3 self-contamination flags; suspend outcome.
**Artifacts:** three methodology documents in `~/_Dev/CapableMind-AI/docs/thinking/David/methodology/`; two feedback memories (`feedback-load-bearing-not-by-immediate-weight.md` + `feedback-skill-success-is-reexplanation-reduction.md`); one new KG drift-pattern (`under-valuing-small-discipline-marks-by-immediate-visible-weight`); session ledger entries.
**Artifacts:** three methodology documents in `~/_Dev/CapableMind-AI/docs/thinking/David/methodology/`; two feedback memories (`feedback-load-bearing-not-by-immediate-weight.md` + `feedback-skill-success-is-reexplanation-reduction.md`); one new KG drift-pattern (`under-valuing-small-discipline-marks-by-immediate-visible-weight`); session ledger entries.
---
## PENDING-22 — Hermes Agent scout deliverable (for jurist review)
**Date:** 2026-05-27
**Tag:** RESEARCH / SCOUT — awaiting jurist review (contains NO proposals per brief; each candidate adaptation would become a separate [PROPOSAL] only after jurist review)
**Summary:** Completed the steward-authorized, jurist-drafted Hermes Agent scout mission — structured comparative analysis of `NousResearch/hermes-agent` (read from a clone, HEAD `c819bc5`; ~134k★) against the four L1 pain points.
**Deliverable:** `~/_Dev/CapableMind-AI/docs/thinking/David/l1-reliability/hermes-agent-scout-2026-05-27.md` (uncommitted working-tree file on `capableMind_docs` main — awaiting steward decision to commit).
**Two stale-fact corrections to the brief (steward-requested pass), verified against source:**
- **Pain #1 ("confidence discarded") is substantially STALE** — Amendment 61 shipped it end-to-end (I-CF floor `0.35` `base.ts:64`; I-CC ceiling; `source_classification_confidence` persisted; recall composite weights it `query-router.ts:657-680`; `recall.ts:116-118` exposes it). Reframe to "built; open question is calibration, not existence." Materially changes the crosswalk.
- **Pain #2 substrate claim imprecise** — current BMF is a hybrid (SurrealKV + per-module better-sqlite3 + LanceDB + file-logchain), SurrealDB mid-retirement. Not a completed "shift to SQLite+LanceDB."
**Headline findings for the jurist:**
- **Pain #1:** Hermes's *default* memory has NO confidence/quality/provenance (provenance computed-then-discarded; only the opt-in `holographic` plugin has a `trust_score`, and it is usage-feedback not classification-time). **CapableMind is ahead** here.
- **Pain #4 (skill/procedural memory)** is the high-value lesson: `SKILL.md` artifact (minimal enforced schema: name+description+body), dual creation triggers, progressive-disclosure retrieval, never-delete curator lifecycle (maps onto CapableMind's existing `lifecycle_state` enum), agentskills.io portability. CapableMind has no procedural-memory architecture (though `knowledge_type='procedural'` already exists on entities).
- **Pains #2/#3:** concrete inputs (hard-coded curation exclusion taxonomy; trigram-FTS for multilingual; progressive disclosure; explicit model-driven retrieval) but no confidence-gated admission and no auto-retrieval-fidelity solution.
**Governance flag (most important for jurist):** Hermes's headline feature — an autonomous background-review fork that writes skills/memory **without human authorization** — is exactly the autonomous self-modification CapableMind's constitution gates (loop-is-load-bearing; DN-GOV-05). Any borrowed pattern must re-introduce the authorization boundary Hermes omits. §7 candidate adaptations are all marked SPECULATIVE for this reason.
**Awaiting:** Jurist review of the deliverable before any adaptation work. No code, no spec, no proposal produced this session.
---
## PENDING-23 — Skill-harvest practice added to the wake/wrap continuity discipline
**Date:** 2026-05-27
**Tag:** [HARDENING] — continuity-triad; steward-authorized direct implementation this session
**Summary:** Refactored "skills improve from what we learn" into our standing way of working — the *governed* analog of Hermes's autonomous self-improvement fork. `/wrap-up` gains **§1.6 "Skill harvest"** (propose create/patch/retire skills from the session + ledger; never autonomous), a **§8 output field**, and a propose-only constraint. `/wake-up` gains a **glance** for skill-harvest proposals left unauthorized (§2.a + §3). Improved skills now carry a one-line provenance note (added to wake/wrap themselves).
**Rationale:** Yesterday's wake/wrap improvements were this practice run by hand; this makes the reflex standing. The governed translation (propose → steward-authorize → apply → record) is the `[PROPOSAL]→[REVIEWED]` model turned on our own tooling — dogfooding the CapableMind thesis: self-improvement that is governed, auditable, never autonomous. It explicitly **inverts** Hermes's "nothing-to-save should not be the default" — "no harvest" is valid; manufacturing changes is the contamination shape. Yardstick: the reexplanation-reduction memory.
**Files affected:** `~/.claude/skills/wrap-up/SKILL.md` (§1.6, §8, constraints, provenance); `~/.claude/skills/wake-up/SKILL.md` (§2.a, §3, provenance).
**Governance note:** Touches the continuity triad the 2026-05-18 S-cluster audit treated with jurist shape-review (REVIEWED-24/25/26). Steward authorized direct implementation this session (additive, low-risk — same shape as REVIEWED-24's §8 additions). Surfaced for jurist awareness; jurist may refine §1.6 wording or elevate the practice.
**Storage (corrected):** No duplication. `~/.claude/skills/{wake-up,wrap-up,symmetria}` are already SYMLINKS into `~/dotfiles/claude/skills/` (the clean pattern, same as `audit` and `landscape-scan`). The edits therefore landed directly in the canonical, version-controlled files — nothing to consolidate. (Earlier this session I mis-asserted duplication from an `ls -la` that silently followed the symlink; corrected here via `-L`/`diff` check. Drift: asserting-fs-state-from-a-misread-listing — verify with `-L`, not `ls -la` of a symlinked dir.)
**Awaiting:** First live test at this session's `/wrap-up` (§1.6); jurist refinement if desired. The skill changes are in the canonical `~/dotfiles` tree (currently uncommitted).
---
## PENDING-24 — Hindsight deep-read & the L1 epistemic-vs-mechanical analysis (umbrella; contains proposals)
**Date:** 2026-05-27
**Tag:** RESEARCH / ANALYSIS — umbrella for sub-items tagged below ([PROPOSAL] A1/A2/B1, [HARDENING] C1/C2/D1). Steward-authorized deep read ("take all the time you need, do it once"); Symmetria active throughout.
**Summary:** Source-grounded deep read of Hindsight (arXiv 2512.12818 / `vectorize-io/hindsight`, Seb's flag) and of L1's spec + runtime (`BetterMemories.io@3bc8b75`), through the steward's thesis (*an epistemic system should think epistemically end to end, not mechanically*). Three sub-agent reads under the Symmetria §5 preamble + executor re-verification of every load-bearing claim against source.
**Deliverable:** `~/_Dev/CapableMind-AI/docs/thinking/David/l1-reliability/hindsight-deep-read-and-l1-epistemic-analysis-2026-05-27.md` (uncommitted working-tree file on `capableMind_docs` main — awaiting steward decision to commit).
**The verified reversal (changes the strategic picture):** Hindsight's *shipped code is not its paper.* The four-network epistemic typing + per-fact confidence + CARA belief-revision were **removed** (migration `g2h3i4j5k6l7_remove_opinion_fact_type.py`, 2026-04-02: deletes opinion rows, drops `confidence_score`, CHECK → `('world','experience','observation')`); no `reinforce`/`cara`/α math in the engine. They ship a pragmatic 3-type hybrid and still hit 91% on LongMemEval — **because the benchmark gives no credit for epistemic integrity.** CapableMind's governed/epistemic angle is therefore *unmeasured by the field* — its risk and its moat. The steward+Seb bet is vindicated, not threatened.
**Answer to the steward's question** (*refactor with our tools, or are they showing us the way?*): **mostly "our tools."** At the parts level L1 is even/ahead — RRF (k=60, module-weighted), cross-encoder rerank, BM25+vector hybrid (LanceDB), honest read-path degradation, **and a fully-wired _numeric_ confidence chain** (I-CF floor → I-CC ceiling → persisted `source_classification_confidence` → recall weight 0.15; all verified live). Our gap is not missing tools — it is: **(a)** the epistemic *kind* signals (`means_of_knowing`, `earned_confidence`) are computed at write and **read by nothing in recall** (verified — orphaned exactly as the numeric confidence was *before* Amendment 61); **(b)** the similarity probe / observation-recall coupling is **dead code** (`setSimilarityProbe` has zero callers — REVIEWED-18 inert and silent); **(c)** the causal subsystem is an **ungoverned inference-generator** (N6: ~42 edges/event, 97%+ coherence-unevaluated, `json_each` full-scan in the ingest hot loop — *the epistemic failure and the operational crash are the same failure*); **(d)** no external benchmark to tune recall against.
**Where they genuinely show us the way (borrowable with our tools):** (1) **bounded graph growth** — per-unit link caps (`_cap_links_per_unit`: temporal 20 / semantic 50) + anti-hallucination causal `target_index < i` (prior-only) — the exact governor N6 lacks; (2) **always-on local recall quality** — their cross-encoder rerank runs unconditionally on an 80 MB local model, where **L1's rerankers no-op unless inference slots are graduated** (cold/teacherless → heuristic-only recall); (3) the **LongMemEval/LoCoMo benchmark harness** (plug-in seam: dataset/generator ABCs + an L1 adapter exposing `retain_batch_async`+`recall_async`).
**Sub-items surfaced (none unilaterally committed):**
- **A1 [PROPOSAL] (jurist territory):** thread `means_of_knowing`/`earned_confidence` to recall as output provenance (+ optional ranking signal) — *"Amendment 61 for the qualitative epistemic axis."* L1-only (existing fields); the L2-coupled belief-schema version stays PARKED.
- **A2 [PROPOSAL]:** if we adopt the benchmark, record it as a **floor not a ceiling** (it cannot score epistemic integrity; Hindsight is the cautionary case of optimising it away).
- **B1 [PROPOSAL] (architectural):** an epistemic governor on causal-edge generation — Hindsight's per-unit cap + prior-only constraint (mechanical half) **+** mint causal edges as held/low-confidence `means_of_knowing=inference`, promotion gated on coherence (epistemic half). Defuses N6 *and* prevents the next one.
- **C1 [HARDENING]:** bundle a local always-available cross-encoder fallback so recall quality doesn't depend on slot graduation.
- **C2 [HARDENING]/issue:** fix or honestly remove the dead similarity probe (`orchestrator.ts:363`, zero callers).
- **D1 [HARDENING]:** wire L1 to the LongMemEval/LoCoMo harness via an adapter (bind to A2).
**Set aside on record:** BMF-on-Hindsight-substrate (relational — L1 is the co-authored mechanism since 2026-05-23 (steward + Seb), conceived from the steward's Chamber prototype; substrate change touches both co-authors' work; sovereignty — Postgres/Oracle vs L1's local-first sqlite+LanceDB+file-logchain; governance — Hindsight has no authorization loop / logchain immutability / external-review hook). We take technique + validation, not substrate. Paper-vs-code divergence is itself a caution: borrow from *their code*, not their paper.
**Caveat:** checkout `3bc8b75`; Seb's later commits (`bd70ceb`, `e8c5fb7` w/ D1–D10) are not on disk and may move some findings.
**Audit update (2026-05-28):** four-pass pre-build audit completed (`pre-build-audit-2026-05-28.md`). Findings (a)/(b)/(c)/(d) were re-tested against substrate; A1's persistence-finding (no schema for `means_of_knowing`; only numeric value of `EarnedConfidence` persisted) corrected; B1's "prior-only constraint" borrow ruled redundant (BMF enforces by construction); B1's `_cap_links_per_unit` borrow validated as 1–3 lines; the parent amendment's items 9 (numeric confidence in recall response) and 10 (epistemic state in health) found NOT shipped, reshaping A1's scope. Co-author branch deferred. Audit document is part of the Monday package.
**Awaiting:** steward review of the audit; jurist review of A1/A2/B1; co-author engineering review of B1/C1/C2/D1 (Seb on resumption from Peter block 2026-06-01+; steward + executor continued joint work during).