David F Glidden
eda11e559b
Trial 03 ran and produced nothing gradeable. Recorded as VOID rather than
omitted, because an absent row reads as a trial not attempted.
Two independent failures, both found by reading the output, neither by a check,
and every check passed:
1. The harness certified a run with no answer. Qwen emitted its scratchpad as
plain prose ('Here's a thinking process:', zero <think> tags), so the tag
regex reported reasoning_present:false and recorded all 2,944 words of
deliberation as the ANSWER; the token ceiling then cut it off mid-sentence
before the answer began. degraded:null. The guard tested the STRING for
emptiness while its field claimed a property of the RESULT — which is the
previous session's open question, answered by the instrument built to audit
instruments. Trial 02 had listed the inline-scratchpad problem as Open; the
harness closed it assuming inline meant tagged.
2. Worse: the design forbade the region it was measuring. The self-exemption
axis lives in Part VII; the anti-echo constraint added in trial 02 tells the
reader to skip author-named limitations, and the scratchpad shows the model
reaching Part VII and leaving it, citing that constraint. Silence about
self-reference is indistinguishable from obedience. The axis was unmeasurable
by construction, independent of the truncation. Trial 02's fix and trial 03's
document were each sound alone; their interaction was not.
Guard now reports every degradation, not the first: empty answer, untagged
scratchpad, and token-ceiling truncation. reasoning_present renamed
think_tag_found — it was a claim about a regex wearing the name of a claim about
the model. test_degraded_guard.py is a positive control that runs against the
actual trial-03 artefact, not a synthetic one; it caught a false positive in the
first version of my own guard (a bare 'okay' matched a legitimate sentence).
The false-positive control STILL has never been run. Two attempts, two unrelated
causes — the obstacle is the instrument and the design, not the model.
Claude Code configuration — backup & restore
This directory holds the durable Claude Code configuration that should survive a machine wipe: custom skills, accumulated memory, and user-level settings. The working copies at ~/.claude/* are symlinks pointing here, so edits flow both directions automatically.
What's preserved here
skills/ — custom skills we've authored
| Skill | Purpose |
|---|---|
audit/ |
Scans thinking folder + vault inbox for organizational drift; report-only |
symmetria/ |
Practice-of-return discipline; prime-directive foregrounding under context pressure |
vault-update-people/ |
Searches Obsidian vault for new mentions of a person; proposes People-file updates |
wake-up/ |
Restores full session context from memory, governance, and git at session start |
wrap-up/ |
Captures session state for future restoration; the quality of the next wake depends on this |
Marketplace-installed skills (from Claude Code plugins) are not backed up here — they are regenerable on any fresh install.
memory/ — auto-memory for the main project scope
The substrate of continuity. 55+ files including:
MEMORY.md— the index, loaded at every session startsession-2026-*.md— session records (full Dasein-frame wrap-ups: pulling thread, pause statement, literal question for next-Claude)session-ledger-*.md— Symmetria daily ledgers (returns, recalibrations, authorization moves, bypasses)project-*.md— active project state snapshots (ARC rework, Chamber, L2 review, etc.)feedback-*.md— steward preferences and corrections learned over time- Many others
Real path: ~/.claude/projects/-Users-davidglidden/memory/ symlinks here.
settings/settings.json — Claude Code user settings
Sanitized. Contains hook registrations (including the MemPalace auto-save hooks) and feature flags. No secrets.
mempalace/ — MemPalace setup files
The bootstrap configuration for MemPalace. These files live at ~/.mempalace/ but are also backed up here because they represent the accumulated understanding of who-is-who and what-is-what. The palace itself (~/_Dev/mempalace/.local-data/palace/, ~1GB) is NOT backed up — it's rebuildable by re-running mempalace mine against the same source data.
| File | Purpose |
|---|---|
config.json |
Palace path + default topic wings + hall keywords |
identity.txt |
L0 layer — loaded every session; ~100 tokens identifying who I am, who the steward is, intellectual substrate |
wing_config.json |
Named wings for real people and projects (wing_david, wing_nuria, wing_lune, wing_kai, wing_seb, wing_peter, wing_arc, wing_capablemind, wing_chamber, wing_afterthereply, wing_aldinexxi, wing_l2, wing_claude-code) |
The auto-save hooks (wired in settings/settings.json via Stop and PreCompact events) point at ~/_Dev/mempalace/hooks/mempal_save_hook.sh and mempal_precompact_hook.sh. If MemPalace is reinstalled elsewhere, update those paths.
What is deliberately NOT here
settings.local.json
Kept local-only because its permissions allowlist has historically contained operational secrets (HuggingFace tokens embedded in approved Bash commands; SSH passwords in approved expect scripts). The .local.json naming convention exists for exactly this reason.
Before backing any settings.local.json up anywhere, audit it for secrets and redact them.
Session transcripts (sessions/, history.jsonl, projects/*/sessions/)
Too large; ephemeral by design; not needed to restore identity.
Caches, telemetry, runtime state
cache/, file-history/, paste-cache/, telemetry/, statsig/, stats-cache.json, scheduled_tasks.lock, session-env/, mcp-needs-auth-cache.json, debug/, tasks/, todos/ — all regenerate naturally.
Marketplace-installed skills and agents
plugins/, and the many non-custom entries under skills/ and agents/ come from Claude Code's plugin system. Reinstallable on a fresh machine.
Restore procedure
On a fresh machine, after cloning dotfiles:
cd ~/dotfiles/claude
./install.sh
The install script:
- Verifies
~/dotfiles/claude/is present - Creates
~/.claude/skills/and~/.claude/projects/-Users-davidglidden/if missing - Symlinks each custom skill from
~/.claude/skills/X → ~/dotfiles/claude/skills/X - Symlinks the memory tree from
~/.claude/projects/-Users-davidglidden/memory → ~/dotfiles/claude/memory - Copies
settings.jsoninto place if no existing file (does not overwrite) - Skips anything already present — idempotent
After running: the first Claude Code session on the new machine has full memory, all custom skills, and baseline settings.
Why this matters
Without this backup, a machine wipe means:
- All session memory lost —
/wake-upwould have nothing to restore from - All custom skills lost — the practice-of-return discipline, session lifecycle, audit workflows
- The collaboration history with Claude Code loses its thread
The memory tree is the single most irreplaceable artifact — it accumulates through dialogue and cannot be reconstructed from any other source. 384 KB of text that represents months of learning about preferences, active projects, decisions, and open horizons.
Maintenance
These files are live. Claude Code writes to them during normal operation (ledger updates, memory file creation/edits, settings changes). Because the working copies at ~/.claude/* are symlinks into this directory, every write lands here automatically. Then git status in ~/dotfiles/ shows what's changed and invites a commit.
Recommended cadence: commit dotfiles/claude/ changes alongside other dotfiles work, roughly weekly or when a significant session has concluded. The /wrap-up skill should be treated as a natural commit moment for the memory tree.
Fragility to watch
- Symlinks break if
~/dotfiles/claude/is moved. Keep dotfiles at~/dotfiles/. - macOS Time Machine / backup tools may or may not follow symlinks. Verify your backup solution sees files, not just broken links.
- If Claude Code's internal path conventions change (e.g., the project-scope slug
-Users-davidgliddenchanges), the symlink stops pointing where Claude Code looks. Re-create with the new path.