Construct-Network Redesign — generalizing spiraling-v2 principles (Eileen #606) to the construct layer

[zkSoju]

Summary

Comprehensive feedback and redesign principles for the Loa Constructs network. Generalizes Eileen's spiraling-v2 analysis (0xHoneyJar/loa#606) from the cycle-orchestration layer to the construct-composition layer. Incorporates Jani's FAANG-parallelism framing ("100s of workers... figuring out approaches so they are not stepping on each other's toes") as the mental model for the architecture.

Full feedback doc at ~/hivemind/strategy/construct-network-feedback-2026-04-20.md (operator hivemind, syncable on request).

The frame shift

We are not building a better single-agent. We are building the architecture that lets dozens to hundreds of specialized constructs (+ their subagents) work concurrently on the same codebase, world, or creative artifact without chaos. Every feedback item inherits from this frame.

Three doctrine inputs compose here:

1. Eileen #606 — spiraling-v2 principles (observability, compact state, selective verification, micro vs macro loops)
2. Jani 2026-04-20 — "think like corporations at FAANG-parallelism scale"
3. Eileen 420 transcript — civic-layer separation (system vs participation must not blur)

Current diagnosis (honest)

Working

• 29 packs installed, clear thematic organization
• Persona-scoped skill organization inside each pack
• Open-source framework + composable packs — architectural bones are correct
• External validation (Eileen observed a large enterprise running same construct-network shape)
• Hivemind doctrine accumulating

Breaking or absent

• Observability: can't see which pack loaded, which skills fired, what context was read, what artifacts got emitted, cost attribution
• Territory overlaps: no frontmatter schema for writes-to / reads-from — two concurrent constructs can collide
• Invocation vocabulary mess: /feel, "FEEL mode", "wear ALEXANDER lens", "you embody STAMETS" — 5 idioms for similar things, doesn't scale to 100+ constructs
• Civic-layer not declared: no frontmatter says governance_layer (system vs participation)
• Corpus emission inconsistent: most constructs don't emit labeled outputs; sparse training data for RL horizon
• Context bloat: every invocation re-ingests large doctrine pages; no compaction layer
• No composition recipes: known-good pairs (ALEXANDER+KEEPER, STAMETS+OSTROM) aren't first-class invocable units
• No orchestration router: operator hand-composes; doesn't scale

The 8 feedback axes (load-bearing)

#	Axis	Priority	Maps to #606
1	Observability per invocation	P0	CHANGE 1, MUST 3
2	Territory declaration in frontmatter	P1	— (FAANG-parallelism)
3	Civic-layer as first-class schema	P1	— (Eileen 420 transcript)
4	Invocation vocabulary tiered (named / intent / router)	P1	—
5	Orchestration router as first-class skill	P2	MUST 6, §8
6	Composition recipes (known-good pairs)	P2	—
7	Uniform corpus emission schema (feedback-v3)	P1	— (biggest RL-horizon leverage)
8	Micro/macro loop separation	P2	CHANGE 4

Axis 1 · Observability (P0)

Every construct invocation emits trajectory event to `.run/construct-trajectory.jsonl`:
```jsonl
{"ts":"...","construct":"artisan","persona":"ALEXANDER",
"skills_invoked":["analyzing-feedback"],
"context_read":["taste.md","feel-register.md"],
"artifacts_emitted":["grimoires/.../feedback.md"],
"verdicts":[{"kind":"register_check","result":"pass","conf":0.8}],
"cost_usd":0.42,"input_tokens":4021,"output_tokens":877}
```

Without this, every other optimization is guesswork.

Axis 2 · Territory declaration

Extend `construct.yaml` schema:
```yaml
governance_layer: system | participation | both
writes_to: [glob patterns]
reads_from: [glob patterns]
compose_with: [{slug, relation}]
conflicts_with: [{slug, reason}]
register_scope: per-world | all-worlds
```

Enables orchestration to compose 10+ constructs in parallel without toe-stepping.

Axis 3 · Civic-layer tentative mapping

Pack	Layer	Why
loa, archivist, observer, crucible, hardening, noether, protocol, dynamic-auth	system	governors
artisan, k-hole, the-arcade, rosenzu, kansei, the-easel, mibera-codex	participation	speakers

Rule: system may veto; participation may propose. Crossing the line = error.

Axis 4 · Invocation vocabulary tiers

See companion glossary `~/hivemind/wiki/concepts/construct-invocation-glossary.md`:

• Named tier (learning, precision): `wear KEEPER lens`, `embody OSTROM`
• Intent tier (scale): `audit register`, `research X`, `tour cycle-N`
• Router tier (ambiguity): `AskUserQuestion` with 2-3 construct options

Axis 7 · Uniform corpus emission

Every construct verdict projects to `feedback-v3` JSONL (same schema as VFP tour marks). Construct outputs become TourMarks with `source: "construct:/"`. This is the single biggest leverage for the RL horizon — every invocation becomes a training pair automatically.

Connection to VFP

VFP (Visual Feedback Protocol) and the construct-network are the same problem at different scales. Every construct verdict is a TourMark. Every operator reaction is a corpus row. The `feedback-v3` emission schema designed for the agent-led tour is the same schema constructs should emit.

Don't build a second feedback protocol. Constructs ride VFP.

Staged rollout (priority order)

Mirrors #606 §10:

1. Observability repair (Axis 1) — every other optimization depends on it
2. Schema extension (Axes 2, 3) — territory + civic-layer in frontmatter
3. Corpus emission uniformity (Axis 7) — feedback-v3 across all constructs
4. Invocation glossary + named→intent tiers (Axis 4)
5. Composition recipes (Axis 6) — canonize known-good pairs
6. Orchestration router (Axis 5) — construct-layer analog of Spiral harness
7. Micro/macro separation (Axis 8) — scope-tagged verdicts

Expected impact (conservative): 30-50% token reduction for routine construct invocations by month 3; usable at 50+ concurrent construct capacity by month 6; full VFP-integration by month 9.

Anti-patterns (DO NOT)

1. Silent auto-escalation of construct scope
2. Construct self-review (breaks civic-layer)
3. Frontmatter as prose (keep machine-readable)
4. Inventing new invocation idioms per session
5. Removing civic-layer split for convenience
6. Shipping new constructs without declaring territory + governance_layer
7. Routing every operator intent through orchestrator during learning phase
8. Optimizing for "fewer constructs active" vs "right constructs active, bounded territories"

Success criteria

1. Every invocation traceable in .run/construct-trajectory.jsonl
2. No concurrent constructs write to overlapping zones
3. Civic-layer category visible pre-invocation
4. Operator can answer "which constructs are active"
5. Operator can answer "why did this session cost $X"
6. Any construct verdict projectable to feedback-v3 corpus row
7. Composition recipes discoverable via `/compose list`
8. Ambiguous intents → `AskUserQuestion`, not silent pick
9. Civic-layer separation schema-enforced (not doctrine-only)
10. Named-tier teaches; intent-tier scales — both functional

Related

• Spiraling Report Recommendations #606 — ancestor analysis generalized here
• Operator hivemind: `~/hivemind/strategy/construct-network-feedback-2026-04-20.md` (comprehensive draft)
• Operator hivemind: `~/hivemind/wiki/concepts/construct-invocation-glossary.md` (tier doctrine)
• Operator hivemind: `~/hivemind/wiki/concepts/agent-native-civic-architecture.md` (civic-layer doctrine)
• Operator hivemind: `~/hivemind/wiki/concepts/riding-loa-endgame.md` §12 (five-layer composition stack)

---

🤖 Drafted by opus-4-7 in session with @zkSoju. Synthesizes Eileen's #606 principles, Jani's FAANG-parallelism framing, and operator's construct-network pruning pass (2026-04-20).

[eileen1337]

3. 420 transcript is not talking about agents in a team, it has to do with individualized agent instances and nothing talked about should be used as ANY processes here.

[eileen1337]

As an extra, hivemind was originally designed to execute, since then everything needs to become code and we are holding off on all hivemind until later on when we have our own enterprise.
The large corporation I am referring to has their own system for product synthesis, this is the parallel with constructs. NOT the hivemind which needs to be its own SLM creation as opposed to any of the things that are talked about here. We are not going to do any work on the hivemind until we actually have a hivemind. None of that will come until dNFT protocol and freeside, which are mega projects of consolidating current and past consolidations has been shipped. As of current extremely slow shipments it will take at least 1 year until hivemind can be worked on again.

[zkSoju]

ok