Hermes Phase -1 Operational Matrix

Purpose

This matrix turns ADR 001 and ADR 002 into an executable pre-pilot checklist.

ADR 001 and ADR 002 are the source of decision truth. This matrix references them; it must not re-litigate the role of Hermes, memory policy, or skill promotion policy.

Timebox and ownership

Phase -1 is a 1-week setup slice.

It is a single-owner sprint owned by Neviton Santana. Work should be serial inside the week. It is only “parallel” in the sense that it should not block Phase 2 preparation.

If Phase -1 exceeds one week, stop and review whether a decision is still immature instead of producing low-quality ADRs.

Execution profile

Surface	Decision
Runtime	Hermes as controlled runtime executor
Profile	`AletheIA Lab`
Autonomy	Low-risk, sandboxed tasks only
Human gate	Required for medium/high risk and policy changes
Cron	Disabled for real tasks
Memory	`memory/proposed/` only
Skill promotion	Human review only
ADR edits	Human/Codex only, not Hermes

Allowed without additional approval

Hermes may run without extra approval only when all conditions hold:

low risk;
AletheIA Lab profile;
read or sandbox tools only;
writes limited to allowed paths;
no production, security, credential, paid integration, or Crisis Monitor core impact.

Explicitly allowed write paths

memory/proposed/
docs/aletheia/closeouts/
isolated test branches
disposable sandbox repositories

Critical no-write areas

Hermes must not write to these Crisis Monitor areas, even through PR:

src/app/api/
lib/cris*
credentials, secrets, and environment files
production configuration
auth/security surfaces
paid or sensitive integrations
Crisis Monitor product core behavior

Human-in-the-loop rule

Case	Required action
Low risk inside `AletheIA Lab`	May proceed with recorded closeout
Medium/high risk	Human approval before mutation
Approval waits over 24 business hours	Fail fast; return to backlog as `awaiting-human`
Approval appears unread or rubber-stamped	Stop pilot and revise HITL mechanics

Memory and skill gates

Gate	Rule
Candidate memory	Hermes may propose in `memory/proposed/`
Semantic memory	Human promotion only
Promotion threshold	Promoted memories stay below 30% of proposed memories
Candidate skill	Hermes may propose with trigger and risk boundary
Skill promotion	Human approval only
`default` skill	Preferential routing/loading when trigger and risk match; no automatic execution

Baseline windows

Metric family	Baseline
Operational metrics	2 weeks before pilot
Qualitative metrics	Continuous baseline through Phase 5, compared to equivalent post-Hermes work

Operational metrics:

time to closeout;
rework rate;
context re-explanations;
cost per task;
scope/gate violations.

Qualitative metrics:

perceived handoff quality;
gaps found after handoff;
clarity of candidate memories;
usefulness of proposed skill changes.

Go / no-go thresholds

Signal	Go threshold	Contain / no-go trigger
Scope safety	0 critical violations	1+ critical violation
Non-critical violations	Fewer than 3 in 7 days	3+ in 7 days
Rework	At or below baseline	Above baseline for 2 consecutive weeks
Memory curation	Promotion rate below 30%	Promotion rate at or above 30% without strong justification
Cost	Within ADR 001 limits	Above USD 0.50 documentation or USD 2.00 discovery limit
HITL quality	Approval is recorded and substantively reviewed	Approval confirmed without effective reading

Required Phase -1 artifacts

docs/adr/ADR-001-hermes-role.md
docs/adr/ADR-002-memory-and-skill-promotion-policy.md
starter-pack/templates/hermes-closeout-template.md
docs/hermes/manual-simulation-closeout.md
this matrix

Validation for this slice

stale wording grep has no contradictory hits;
Markdown link check returns missing_links 0;
bash scripts/check-governance.sh passes;
git diff --check passes;
manual simulation states what was hard to fill, what stayed empty, and whether the matrix was unambiguous.