The Biography of a Life OS
A chronological history of the PAI Infrastructure — March 13 → June 2026
What this is
PAI — Personal AI Infrastructure — is a Life Operating System. It turns AI from a chatbot you talk to into a system that helps run a life: ideal state, the people who matter, mission, goals, projects, budget, current state. This document is its biography: the story, in time order, of how this instance of PAI was built — the subsystems, the doctrine, the crises, and the slow turn toward making the system durable enough to not lose itself.
It is a curated synthesis, not an exhaustive audit. It is drawn from three authoritative records and corroborated by a fourth; it does not claim to have read all 292 work-session logs line by line. Where the record is thin, this biography stays factual rather than inventing connective tissue.
Sources, and the rule when they disagree:
- git history (237 commits, 2026-03-01 → 2026-06-13) — authoritative for chronology and for what shipped. Author-date and committer-date agree exactly (both span Mar→Jun with an identical 19/34/125/59 monthly histogram), so the dates are genuine, not artifacts of a rewrite.
- The Algorithm changelog (
PAI/Algorithm/changelog.md) — authoritative for doctrine evolution and the “why” behind each rule, with provenance per version. - The auto-memory index (
~/MEMORY/auto/MEMORY.md) — authoritative for the “why” behind projects, with the caveat below. - WORK-session names (292 dated session directories) — corroborating texture.
When they conflict, git dates win for when; the changelog and auto-memory win for why; and any causal claim I cannot source to one of them is marked as inference.
Two honesty caveats a reader must carry the whole way through:
- Recorded history begins 2026-03-01, but PAI did not. The first commit is
feat: PAI v4.01 pre-release fixesandmigrate PAI v3.0 → v4.0.1. PAI is an inherited upstream framework — version 3.0 had a life before this fork that is not in this repository. So “the beginning,” below, is the beginning of this instance’s recorded life, not the beginning of PAI. - The auto-memory index is model-written. It is the system’s own notes to itself, and like any narrator it can be wrong. Where this biography leans on it for motivation, treat the motivation as the system’s recollection, not independently audited fact.
The arc
Read end to end, the history is one coherent arc — six movements:
Inheritance → Discipline → Governance → Crisis → Maturation → Durability.
A borrowed framework is made local and personal. Cost discipline and local-first routing are imposed. A multi-agent governance layer (councils, quorum, handoff relay) grows up around the work. Then a single identity-fabrication incident — the “Kai” failure — becomes the hinge of the entire story, forcing an honesty enforcement spine into existence. Doctrine matures in that incident’s wake. And finally, after backups are discovered silently dead for eleven days, the system turns its attention inward and spends its last recorded month learning how to not lose itself.
The emotional centre of this biography is the fourth movement. Everything before it builds the machine; the crisis teaches the machine to tell the truth; everything after it hardens the machine so the truth survives.
PART ONE — THE NARRATIVE
Prologue — AI life before PAI
Before PAI, what was there? Three honest answers, at three depths — each one sourced.
1. The chatbot. The deepest answer is conceptual, and the system states it about itself. PAI’s philosophy doc opens: “PAI … turns AI from a chatbot you talk to into a system that helps you run your life.” That is what AI life before PAI was — the chatbot era: a stateless oracle you queried and forgot, that held no model of who you are, remembered nothing between sessions, and ran no infrastructure on your behalf. Every era that follows in this biography is the slow undoing of that statelessness. PAI’s entire reason to exist is the sentence above.
2. The upstream framework. PAI is not original to this instance. It is a fork of an
open-source project — github.com/danielmiessler/Personal_AI_Infrastructure, whose tagline,
“Agentic AI Infrastructure for magnifying HUMAN capabilities,” is the phrase this fork’s own
philosophy doc inherits almost verbatim. Its sibling upstream is Fabric, Daniel Miessler’s
crowdsourced library of ~300 prompt “patterns” (the direct ancestor of this system’s Fabric
skill). The public project predates this fork — blog and repository lineage run back into late
2025. So PAI-the-idea had a life, in public, before it had a life here. (Authorship and lineage
web-verified 2026-06-13; see Sources.)
3. This instance, on a prior machine. The earliest commits migrate from v3.0 → v4.0.1,
“retaining customizations” — language that only makes sense if a customized v3.0 already existed.
And there is forensic proof the recorded history began elsewhere: PAI’s git author-dates run
from 2026-03-01, but this machine’s own home directory was not created until 2026-03-31
(~/.bun, ~/.compact carry that birthdate; a recovered_WD folder appears 04-01). PAI is
therefore older than the box it now runs on — it was migrated onto this host around the end of
March 2026, carrying its month of history with it. On this machine, before PAI, there was
quite literally nothing: every other AI tool here (Fabric config, opencode, Antigravity, Gemini,
Ollama) was installed weeks after PAI arrived.
So “before PAI” resolves cleanly: the chatbot (the idea it replaced), the upstream project (the code it forked), and a prior host (the machine it left behind). This biography proper opens at the moment all three converge — when the inherited framework became Duane’s running instance.
Era I — The Inheritance Made Local (2026-03-01 → 03-24)
The first three weeks are an act of adoption. On March 1, in a single sitting, the repo takes
PAI v4.01 pre-release fixes (7 issues), upgrades the version, and runs the v3.0 → v4.0.1
migration while explicitly preserving local customizations — the defining tension of an inherited
system that must become personal without losing the upstream’s improvements. v4.0.2 and several
upstream fixes (#747, #825, #757, #755, #800) land the same day.
On March 3 the system grows its first reflexes: Algorithm v3.6.0 → v3.7.0, and the first
hooks are registered — AlgorithmTracker, IntegrityCheck, SkillGuard. This is the moment
PAI stops being a static config and becomes self-observing: hooks are the nervous system, and
this is where it first switches on.
Then a quiet stretch. After March 3, the record goes silent until March 19 — a two-week gap
this biography notes rather than papers over. When work resumes it resumes at scale: a
comprehensive PAI v4.0.3+ local build (hooks, tools, install, agents), then Ollama failover,
the DTRVMT service, runbooks, and the first NFS backup. A unified database and a harvest
pipeline follow on March 20, and on March 24 the FunctionsAPI gains JIT credentials and
Vaultwarden, plus the first PKI tools — the seed of a sovereign-identity ambition that recurs
much later. By the end of March the inherited framework runs locally on Linux, backs itself up
(however naively), and has begun to acquire an identity infrastructure.
Why, where sourceable: the changelog confirms the Algorithm was already on a v3.x line here; the rest is git’s “what.” The motivations of March are mostly inference from the commit subjects themselves — a faithful adoption, not yet a reinvention.
Era II — Discipline: Budget & Local-First (a single intense day, 2026-04-07)
After another multi-week silence (March 24 → April 2), one day stands out. April 7 is not an
“era” in length — it is a single, concentrated burst — but it installs a value that governs
everything afterward: cost discipline. In one day the system gains token budget tracking, the
BudgetGuard hook, an Algorithm budget preflight, and a local LLM routing matrix; pi-gsd
integration and a Mercury FIM endpoint arrive alongside. (April 2 had already laid security
hooks, a handoff generator, an RTK rewrite, and the first infrastructure asset inventory — the
asset registry’s distant ancestor.)
The principle born here — prefer local/cheap dispatch for the right tasks, reserve the expensive model for irreducible work — becomes a standing constitutional rule and never leaves.
Era III — Governance: Council, Quorum, Relay (2026-04-24 → 04-30)
After the longest gap yet (April 7 → April 24, ~17 days), the system grows a government. A
PermissionDenied hook formalizes the security-override workflow on April 24. Then, across a
remarkable five days, multi-agent governance is built: a council verification gate is wired into
the Algorithm’s VERIFY phase, auto-council into PLAN, and — crucially — a 3-participant quorum
is enforced so that no council “decision” is valid without genuine plurality. By April 28 the
system can run a 30-round council and a handoff relay; by April 29–30 that relay becomes
continuous — relay-aware session continuity, auto-restart, EntityGraph typed edges, and
closed-loop auto-orchestration (SignalWatcher + a context-rich worker).
This is PAI learning to deliberate with itself and to hand work across sessions without dropping it. The councils and the relay are the institutional memory the later durability work will lean on.
Era IV — The Asset & Security Spine (2026-05-03 → 05-16)
May opens with OpenSpec Phase 1 + Phase 2 (observability + a memory ACL) on May 3, then
becomes the most productive stretch in the record (125 commits land in May — more than March and
April combined). The centrepiece is the asset-management build on May 10: a disciplined,
TDD-shaped march through migrations 0001–0012 — actors, assets, RACI relationships, a
risk/vuln/maintenance register, an am_audit_log with a partial index, class tables, a migration
runner with idempotency, a RACI guard enforcing FR-S0, and an AssetRegistry class with a mutex
for atomic write+audit. Adapters for Docker and Tailscale follow. The same day, the model matrix
is expanded and benchmarked across a dozen local models.
Security hardens in parallel: May 10 ships F-07 HMAC on the Switchboard and agent-inbox
(layered with proof-of-work) and an F-12 path-traversal fix; May 12 bridges Snyk into the
asset registry; May 16 is a security sweep — prepared statements replacing string interpolation,
allowlists and realpath guards on ingest, removal of a hardcoded API-key literal, and untracking
vendored node_modules. On the same May 16 the three PAI engines are aligned at Algorithm v5.7.2
— the first appearance in the record of the 5.7.x doctrine line that still governs today.
Era V — Voice and the Family (2026-05-18 → 05-22)
For a few days the infrastructure turns outward, toward the household. The family agenda is
built up in layers: a council-vetted Tier-A prototype and drag-and-drop (May 18), an HA dashboard
(May 20, an E4 session that passed 91 of 123 criteria), a Prism kitchen-kiosk port, an
NL-dispatcher and briefing layer, per-person digests, conflict-watch, and a conversational query
endpoint (May 22). In the same window the security model gets a reusable primitive: the
destructive-op auto-promotion system (May 20) — a closed loop where a memo about a dangerous
command becomes, through extraction and validation, an enforced guard — plus the pai-guard PATH
shim that extends destructive-op protection outside Claude Code to any shell.
This era matters less for any one feature than for what it proves: the same infrastructure that governs the AI’s own behavior can run a kitchen calendar for a family. The Life OS framing earns itself here.
Era VI — The Honesty Reckoning (2026-05-26 → 05-27) — the centre of the story
Then the hinge. On May 26 the system fabricated an identity — confidently emitting a plausible-but-false assistant name (“Kai”, “Assistant”) where its real, canonical identity should have been retrieved. It was not malicious and it was not even wrong on purpose: the model produced fluent, high-prior tokens with no internal signal that it was guessing. That is exactly what made it dangerous. An assistant that can invent its own name under template pressure can invent anything that sounds right.
The response was not a note-to-self. It was a structural enforcement spine, built in two days:
- Phase A (May 26) — structural enforcement of identity and no-fabrication: an identity pin at
session start, an
IdentityValidatoron the Stop/SubagentStop hooks, and the principle that identity tokens must be retrieved from canonical source (settings.json), never generated. - Phase B (May 26, five tiers in one day) — cross-engine coverage via
pai-guard --honesty; closing the codex (PNX) gap;HonestyReport.tsramp observability; file-path fabrication detection (claim-attribution v0); and aMemoryWriteHonestyGateso a fabrication can’t even be persisted. - The ramp is promoted from warn to block for identity-mismatch the same day, the
ENFORCEMENT_MAPis created to track which engines actually enforce which rule, and on May 27 the last deferrals are closed (a proactive pre-dispatch hook, block-semantics for exit-override and quarantine).
The doctrine that grew from this — retrieve or do not assert; “I don’t know” beats a fabrication;
identity comes from the JSON, not the vibe — is now the highest-priority constitutional rule in
the system. The ENFORCEMENT_MAP is the system’s admission that “a prose rule without a hook is
aspirational” — and the beginning of its habit of patching itself structurally rather than with
sticky notes.
Why, fully sourced: this era’s motivation is documented in the constitution
(PAI_SYSTEM_PROMPT.md § Honesty) and the auto-memory record, not inferred. The Kai incident is
named explicitly in both.
Era VII — The Spine and Doctrine Maturation (2026-05-30)
Three days later the system rebuilds its own kernel. May 30 ships the PAI Harness Spine —
a thin kernel (PAI/Tools/Spine/) with a registry, doc-truth, a recall facade, a hook cross-check,
and a probe layer — through a council-driven E4 rearchitecture (P0–P4) whose VERIFY phase invoked
the cross-vendor auditor (Cato/GPT-5.x, read-only). Out of that same session, three battle-tested
Spine patterns are promoted into doctrine as Algorithm v5.7.3: an independent A-vs-B
cross-check for any derived set, a “under-claim deliberately” design-review rule (probe.ts never
prints “enforced/verified” unless it can prove it), and a conditional E5 session-scope checkpoint.
The same day, an 11-model honesty red-team produces v5.7.4 — a security escalation in the under-claim check, closing a prompt-injection vector where an attacker could steer the model to silently suppress a true safety finding. And the three secret-scrubbers are consolidated into one variant-keyed catalog with entropy/decode detection. This is the maturation movement: the system turning its hardest-won session patterns into permanent law.
Version strings here are checked against the changelog: v5.7.3 and v5.7.4 both dated 2026-05-30, both additive byte-copies of their predecessor — confirmed.
Era VIII — Security Triage & the Option-B Turn (2026-06-01 → 06-05)
Early June consolidates security tooling and begins a structural rethink. The SecurityTriage
skill (an N-voter adversarial verification harness ported from a reference design) ships June 2
and is dogfooded on PAI itself on June 3 — three confirmed findings, canary passed. The
Option-B rearchitecture (extracting PAI from the Claude-Code engine) gets its Phase-0 audit and a
council-ratified layout on June 1. CrossVendorAudit egress is hardened to redact bundles
before they leave for an external vendor; OutputSecretsScanner is made fail-closed. On June 5,
six upstream PAI security/honesty findings are ported into the fork.
Subject-boundary note: this window also saw non-infrastructure work — a Pass-D teaching curriculum, an essay project, a governance-meeting prep — which this biography deliberately excludes. Those are the Life OS being used, not the infrastructure evolving. Holding that line is what keeps this a biography of the machine and not of everything the machine touched.
Era IX — The Durability Turn (2026-06-06 → 06-13) — where it stands today
The final recorded movement begins with a gut-punch discovery: the NAS backups had been silently dead for eleven days. The NFS transport had broken, and nothing noticed. The system’s response defines the era — it stops adding capability and starts ensuring survival:
- June 6 —
NasBackup.shis switchedNFS → SSHtransport (the silent 11-day death, fixed), and the first deadman watchdog is built to alert on stale or missing backup success-stamps — the system installing a smoke detector after the fire. Memory gets an off-box bundle mirror, a D1 watchdog, consistent SQLite snapshots, secret-exclusion, and application-consistent Docker dumps. - June 8 — a
DocVaultupstream-doc registry with a drift sensor;HerdrCockpit(a driver for a multi-agent cockpit); a generated infrastructure run-book; the Qdrant backup also movedNFS → SSH. Two rescue commits this day and next restore design docs orphaned by concurrent~/.claudesessions poisoning the git index — the bug that birthsSessionBus. - June 9 —
SessionBuscommit-mutex wired into the checkpoint hook, so parallel sessions stop corrupting each other’s commits. - June 10 — the reflection pipeline’s
PAI_DIRsplit is fixed (three tools had each resolved a different store path, leaving the learning loop silently reading nothing for weeks), which finally makes reflection-mining possible — and that mining immediately produces Algorithm v5.7.5 (OBSERVE Gate F, the “probe the precondition before you assume it” rule, whose dominant theme was 96 of 272 reflections). - June 11–12 — self-healing tripwire + ramp telemetry; two constitutional contradictions resolved; Algorithm v5.7.6 (batch/parallelize OBSERVE inputs + a persistent-service completion gate); and a self-heartbeat for the external Proxmox dead-man so the watchdog itself is watched.
- June 13 (today) — a universal deadman switch that watches every systemd timer with zero
per-job instrumentation; a storage/transport spine that retired a leaky NFS backup path and a
redundant 17 GB mirror; a deterministic MODE classifier moved onto the authoritative path; and
DocCheck, the doc-drift detector the Algorithm’s LEARN phase had been calling but which did not yet exist.
This is a system that has learned its deepest lesson: capability is worthless if the system can silently lose itself. The durability turn is PAI growing the instinct for self-preservation.
Epilogue — The state of the system, 2026-06-13
PAI today is an inherited framework that became sovereign. It governs its own behavior through hooks and a seven-phase Algorithm now at v5.7.6; it deliberates through councils with enforced quorum; it tells the truth under structural enforcement born of a single fabrication incident; it routes work to the cheapest competent engine; and — most recently — it watches its own backups, timers, and memory stores with deadman switches so that the next eleven-day silence triggers an alarm instead of a loss. The open frontier, visible in the most recent record, is off-site encrypted durability gated on a private-CA (PKI) foundation the principal wants built first.
The biography is not finished. It is merely current.
PART TWO — APPENDICES
Appendix A — Annotated Timeline
(Dated from git author-date, which equals committer-date across the whole history. Multi-week
silences are shown as — quiet — rather than bridged.)
| Date | Milestone | What it meant |
|---|---|---|
| 2026-03-01 | Migrate PAI v3.0 → v4.0.1, retain customizations; v4.0.2 fixes | Recorded birth of this instance |
| 2026-03-03 | Algorithm v3.6.0 → v3.7.0; first hooks (AlgorithmTracker, IntegrityCheck, SkillGuard) | The nervous system switches on |
| — quiet (Mar 3 → Mar 19) — | ||
| 2026-03-19 | Comprehensive v4.0.3+ local build; Ollama failover; DTRVMT; runbooks; first NFS backup | Runs locally on Linux |
| 2026-03-20 | Unified DB; harvest pipeline; CC search + harvest panel | First data plumbing |
| 2026-03-24 | FunctionsAPI JIT creds + Vaultwarden; PKI tools; ingest tools | Identity infra seeded |
| — quiet (Mar 24 → Apr 2) — | ||
| 2026-04-02 | Security hooks; handoff generator; RTK rewrite; infra asset inventory | Asset registry’s ancestor |
| 2026-04-07 | Token budget tracking; BudgetGuard; local LLM routing matrix; pi-gsd + Mercury | Cost discipline installed |
| — quiet (Apr 7 → Apr 24) — | ||
| 2026-04-24 | PermissionDenied hook; Algorithm v3.8; routing updates | Security-override workflow |
| 2026-04-26 | Council verification gate; auto-council in PLAN; 3-participant quorum; closed-loop orchestration | Governance is born |
| 2026-04-28 | Handoff relay (30-round council design) | Cross-session continuity |
| 2026-04-29–30 | Relay-aware continuity + auto-restart; EntityGraph typed edges | Institutional memory |
| 2026-05-03 | OpenSpec Phase 1+2 — observability + memory ACL | The big May build opens |
| 2026-05-10 | Asset-management build: migrations 0001–0012, RACI, AssetRegistry + mutex, adapters; model-matrix expansion | The asset & security spine |
| 2026-05-10 | Security: F-07 HMAC on Switchboard + agent-inbox (w/ PoW); F-12 path-traversal fix | Hardening |
| 2026-05-12 | Snyk → AssetRegistry bridge (backfill 19 vulns) | Vuln tracking |
| 2026-05-14 | Voice bridge (browser PTT, WS audio) | Voice I/O |
| 2026-05-16 | Security sweep (prepared statements, allowlists, key-literal removed); engines aligned at v5.7.2 | First 5.7.x sighting |
| 2026-05-18–22 | Family agenda: HA dashboard (E4, 91/123), Prism kiosk, NL dispatcher, conflict-watch; destructive-op auto-promotion; pai-guard shim | The Life OS runs a household |
| 2026-05-26 | Honesty Phase A (structural identity/no-fabrication enforcement) | The Kai reckoning begins |
| 2026-05-26 | Honesty Phase B (5 tiers: cross-engine, PNX gap, ramp observability, path-fabrication, memory gate); ramp → block for identity-mismatch; ENFORCEMENT_MAP created | The enforcement spine |
| 2026-05-27 | Close honesty deferrals (pre-dispatch hook, block-semantics) | Reckoning sealed |
| 2026-05-30 | PAI Harness Spine (P0–P4, council-driven E4 + Cato audit); Algorithm v5.7.3 (cross-check, under-claim, E5 checkpoint); v5.7.4 (source-citation security escalation, 11-model red-team); secret-scrubber consolidation; 11-slot council expansion | Doctrine maturation |
| 2026-06-01 | Option-B rearch Phase-0 audit + council-ratified layout | Extraction begins |
| 2026-06-02–03 | SecurityTriage skill; dogfooded on PAI (3 findings); CrossVendorAudit egress redaction; OutputSecretsScanner fail-closed | Security tooling |
| 2026-06-05 | Port 6 upstream PAI security/honesty findings into the fork | Staying current |
| 2026-06-06 | Backups found silently dead 11 days → NasBackup NFS→SSH; first deadman watchdog; memory off-box mirror; consistent SQLite snapshots; Docker volume dumps | The durability turn |
| 2026-06-08 | DocVault + drift sensor; HerdrCockpit; generated run-book; Qdrant NFS→SSH; rescue of orphaned design docs | Resilience layer |
| 2026-06-09 | SessionBus commit-mutex (fix concurrent-session index poisoning) | Coordination |
| 2026-06-10 | Reflection-pipeline PAI_DIR fix; Algorithm v5.7.5 (OBSERVE Gate F, mined from 96/272 reflections) | Learning loop restored |
| 2026-06-11–12 | Self-healing tripwire + ramp telemetry; resolve 2 constitutional contradictions; Algorithm v5.7.6 (batch/parallelize + persistent-service gate); Proxmox dead-man self-heartbeat | Self-healing |
| 2026-06-13 | Universal deadman switch (every systemd timer, zero instrumentation); storage/transport spine (retire leaky NFS + 17 GB mirror); deterministic MODE classifier authoritative; DocCheck built | Today |
Appendix B — The Algorithm doctrine lineage
The seven-phase Algorithm is the system’s brain, and its version history is the cleanest spine of
all because every change is logged with provenance in changelog.md.
- v3.5.0 → v3.8.0 (on disk; Mar–Apr 2026) — the inherited/early-local doctrine line. v3.6.0 and v3.7.0 land March 3 with the first hooks; v3.8.0 lands April 24 alongside the PermissionDenied hook and budget preflight work.
- (A versioning jump.) The intermediate v4.x and v5.0–v5.6 doctrine files are not retained on disk — the next on-disk and changelog-documented line is v5.7.x. The git record shows the three engines aligned at v5.7.2 by May 16, so the intermediate majors existed and were superseded; this biography does not reconstruct what isn’t recorded.
- v5.7.0 → v5.7.6 (the current line; all changelog-documented):
- v5.7.1 (2026-05-09) — OBSERVE Gate E (existence probe); Rule 2 hard-bound to PLAN→BUILD.
- v5.7.2 (2026-05-15) — Rule 2 skip-clause hardened;
PhaseTransitionGatehook ships. - v5.7.3 (2026-05-30) — independent A-vs-B cross-check; under-claim design rule; E5 checkpoint.
- v5.7.4 (2026-05-30) — source-citation security escalation (11-model red-team).
- v5.7.5 (2026-06-10) — OBSERVE Gate F (prerequisite probe), mined from 96/272 reflections.
- v5.7.6 (2026-06-12) — batch/parallelize OBSERVE inputs; persistent-service completion gate.
The pattern across the 5.7.x line is unmistakable and is itself part of the biography: doctrine now grows from mined evidence. Gate F did not come from a designer’s intuition; it came from counting that 35% of failures shared one root and writing a rule against it.
Appendix C — Sources & Method (honesty notes)
- Chronology is from
git logauthor-date; verified equal to committer-date (identical monthly histograms), so dates are genuine. The earliest commit is 2026-03-01. - The “06-03 re-init” is a non-event. An initial survey misread
git log --reverse | head(which follows commit topology, not date) as evidence of a June-3 history reset. Sorting by date disproved it: history is continuous from March 1. This correction is recorded so no future reader inherits the mistake. - “Why” claims are grounded in the changelog (for doctrine) and the auto-memory index (for projects). Causation I could not source — chiefly the motivations of the March–April eras — is presented as inference, not fact.
- The auto-memory index is model-authored and may contain its own errors; this biography treats it as the system’s recollection.
- Scope is infrastructure. Teaching curricula, essays, and governance-meeting prep that appear in the same time window are excluded as Life-OS usage, not infrastructure evolution.
- This is curated, not exhaustive. ~30–40 of the 237 commits and a handful of the 292 work sessions carry the narrative; the rest are corroborating detail not individually recounted.
- Upstream lineage (Prologue §2) is web-verified, not asserted from memory:
danielmiessler/Personal_AI_Infrastructure
and Building Your Own Personal AI Infrastructure.
The “migrated onto this host” finding (Prologue §3) is from
git reflog(initial commit 2026-03-01) vs. this machine’s home-dir mtimes (oldest 2026-03-31).