# Run Phase 0 shadow mode
> Phase 0 dry-run: exchange and track commitments, execute only the Grid 1 baseline, then graduate to Phase 1 local integration and Phase 2 federation.
- Repository: g2tf-org/g2tf-standards
- GitHub: https://github.com/g2tf-org/g2tf-standards
- Human docs: https://grok-wiki.com/public/docs/g2tf-org-g2tf-standards-436ba3f3e0ff
- Complete Markdown: https://grok-wiki.com/public/docs/g2tf-org-g2tf-standards-436ba3f3e0ff/llms-full.txt
## Source Files
- `architecture/design-principles.md`
- `architecture/federation.md`
- `architecture/failsafe-model.md`
- `CONTRIBUTING.md`
- `rfcs/grid-2-rfc-1.pdf`
- `README.md`
---
---
title: "Run Phase 0 shadow mode"
description: "Phase 0 dry-run: exchange and track commitments, execute only the Grid 1 baseline, then graduate to Phase 1 local integration and Phase 2 federation."
---
RFC #1 §Phased Adoption defines **Phase 0 — Shadow Mode** as: run the full protocol stack as a dry run; make and track commitments; execute only what would have happened anyway. In this repository that stack is **G2P** (`specs/g2p/spec.md`, v26.0-draft), **GCAP** (`specs/gcap/spec.md`), and **DESP** (`specs/desp/spec.md`). Physical dispatch stays on each element's **Grid 1 baseline**. There is no shadow-mode flag, CLI, or wire format in the living specs — Phase 0 is an operational constraint on an otherwise complete interval loop.
All three protocols and the member files are `Status: Draft — seeking input`. G2P §7 leaves wire format, encoding, element state machines, version negotiation, and conformance test vectors unspecified. A Phase 0 run is an implementation report against Draft text, not a conformance claim against a Stable `v26.Y` line.
## What Phase 0 does and does not do
| Surface | Phase 0 requirement | Not in Phase 0 |
|---|---|---|
| G2P message families | Exchange and persist Service Descriptors, Commitments, and Response / Telemetry on the one-minute clock | A published encoding or transport |
| GCAP | Allocator computes per-class `commit-%`, Buffer committed dispatch, and Source committed take from provisioned Grid 1 constraints | Binding those outputs to physical setpoints |
| DESP | Members declare A / B / C positions; Allocator records C → B → A curtailment order | Charging, class SLAs, or domain class-mix enforcement (commercial terms, outside the protocol) |
| Physical POI | Execute only the Grid 1 baseline that would have run without Grid 2 | Self-dispatch of the Commitment |
| Federation announcements (G2P §3.3) | Optional local bookkeeping | Phase 2 inter-domain interoperation |
RFC #1 recommends starting in **select grid pockets**, typically a hyperscaler, a utility transmission provider, a clean energy provider, and a load serving entity. The smallest Allocation Domain remains a **host utility plus one flexible load** (`architecture/allocation-domains.md`). Broader recognition (for example FERC-level) is deferred until the local path is proven (`architecture/design-principles.md`).
## Adoption lifecycle
```mermaid
stateDiagram-v2
[*] --> Phase0: RFC 1 recommended start
Phase0: Phase 0 Shadow Mode
Phase0: protocol runs, Grid 1 executes
Phase1: Phase 1 Local Vertical Integration
Phase1: solution proof in existing rules
Phase2: Phase 2 Regional Federation
Phase2: peer Allocators announce and listen
Phase0 --> Phase1: local pocket ready to bind commits
Phase1 --> Phase2: same-spec peers couple domains
Phase0 --> Phase0: interval loop, no physical Grid 2
```
| Phase | RFC #1 wording | Binding physical behavior |
|---|---|---|
| **0 Shadow Mode** | Run the full stack as a dry run; make and track commitments; execute only what would have happened anyway | Grid 1 baseline / standalone schedule / interconnection behavior (`architecture/failsafe-model.md`) |
| **1 Local Vertical Integration** | Groups of large loads, transmission owners, and load serving entities build a solution proof in existing rules structures | Members self-dispatch Commitments inside host-utility parameters |
| **2 Regional Federation** | Peer entities running the same spec interoperate; inter-domain advertisements and policy | Announce-and-listen across coupled zones; local constraints still win (`architecture/federation.md`) |
`architecture/federation.md` lists **Phase 2 milestone criteria** as an open item. Do not invent a graduation scorecard.
## Interval loop (dry-run)
Phase 0 uses the same Member ↔ Allocator pattern as live operation. The split is at the point of interconnection: the Commitment is recorded; the physical asset does not follow it.
```mermaid
sequenceDiagram
participant Use as Use / Buffer / Source
participant Alloc as Allocator
participant TO as Grid 1 custodians
participant POI as Physical POI
TO->>Alloc: Provision constraints (capacity, transmission, reserves)
Use->>Alloc: G2P Service Descriptor (signed positions)
Alloc->>Alloc: GCAP clear (headroom, buffers, C then B then A)
Alloc->>Use: G2P Commitment (interval, domain, inputs version)
Note over Use: Track commit vs baseline; do not bind setpoints
Use->>POI: Execute Grid 1 baseline only
Use->>Alloc: G2P Response / Telemetry (actual vs Commitment)
```
Where a Grid 1 dispatch instruction conflicts with a Grid 2 allocation, **Grid 1 prevails** (`architecture/temporal-position.md`). In Phase 0 that rule is absolute: the allocation is never the executed instruction.
### Messages still required
G2P defines four families. Phase 0 still needs the first, second, and fourth. The third belongs to Phase 2.
Signed per-interval (or reissued) statement of requests/offers over a forward time vector. Use: one consume position per DESP class. Buffer: signed charge (+) / discharge (−). Source: offer to supply. Units: energy in the interval (v26.0).
Per-class `commit-%` to Use; committed dispatch to Buffer (cleared before any load curtailment); committed take to Source. MUST identify interval, domain, and the clearing-inputs version.
Verified self-dispatch report of actual per-class dispatch against the Commitment so the overlay footprint is measurable. In Phase 0, **actual** is the Grid 1 baseline; the Commitment is the counterfactual.
Announce-and-listen across coupled zones. Out of Phase 0 scope. A missing peer announcement MUST be treated as no extra boundary headroom if you exercise this path early.
Timing budget (descriptor deadline, clearing deadline, commitment publication, clock sync) is an open Engineering discussion. Allocations apply to **whole one-minute intervals** only.
## Element behavior in shadow mode
Treat Phase 0 physical output as the same table the failsafe model uses when context is lost. Protocol software still runs; setpoints do not move off baseline.
| Element | Protocol still does | Physical execute (Phase 0) |
|---|---|---|
| **Use** | Broadcast per-class consume positions; listen for `commit-%` | Grid 1 firm load limit or the site's existing manual curtailment; honor agreed ramp rates |
| **Buffer** | Broadcast charge/discharge offers; listen for committed dispatch | Standalone Grid 1 schedule; **no Grid 2 support commitment** |
| **Source** | Broadcast supply offers; listen for committed take | Grid 1 interconnection behavior; agreed ramp rates |
| **Allocator** | Deterministic GCAP from TO-provisioned constraints; publish Commitments and (SHOULD) an inputs digest | Does not command breakers, setpoints, or market awards. On stale inputs or partition: **MUST NOT clear** |
The Allocator has **no discretion** (`members/allocator.md`, GCAP §1). Identical inputs MUST produce identical allocations so a shadow log can be reproduced by any party holding the same descriptors and constraint set.
Pool sizing still applies as a planning bound: the participating pool MUST be sized so simultaneous fallback stays inside manageable Grid 1 ramp limits. Phase 0 does not relax that constraint; it is the same bound Phase 1 will inherit.
## Run a pocket
Pick the smallest viable pocket: host utility transmission owner plus at least one Use Member. Authority for constraints is the transmission owner; a regional operator MAY add capacity limits. Commercial terms (flex enforcement, class mix, charging) stay **outside** G2P/GCAP/DESP.
Feed the Allocator each interval: zone production capacity, transmission constraints, target reserves; optionally distribution factors and DLR. Peer announcements MUST NOT relax local limits. GCAP MUST withhold clearing on stale constraint inputs.
Implement Use / Buffer / Source / Allocator against `members/*.md` and the three specs. Choose any local encoding and transport; this repository does not specify one. Authenticate every Service Descriptor and Commitment (G2P §6). Transmission-level deployments are expected on private networks.
On the one-minute clock: Members send descriptors → Allocator runs GCAP → Allocator publishes Commitments (interval, domain, inputs version) → Members **do not** apply them to the POI → Members send telemetry of actual (baseline) versus Commitment.
Keep a per-interval log that a third party can replay: inputs digest, descriptors, Commitment, actual MW/MWh, and the delta (counterfactual Grid 2 vs executed Grid 1). That delta is the Phase 0 evidence, not a dispatch order.
File implementation reports in GitHub Discussions (`CONTRIBUTING.md` comment-cycle step **Run**). Use Issues only for spec defects or RFC #1 errata. Running Phase 0 experience is what informs the next living-spec revision.
### Suggested per-interval record
There is no mandated schema. A replayable log needs at least:
| Field | Role |
|---|---|
| `interval` | One-minute window the messages bind to |
| `domain` | Allocation Domain id (format open) |
| `inputs_version` / digest | Constraint set the Commitment was computed against (GCAP §5 SHOULD publish) |
| Service Descriptor set | Signed Use / Buffer / Source positions |
| Commitment | `commit-%`, committed dispatch, committed take |
| `actual` | Metered or scheduled Grid 1 baseline |
| `delta` | Commitment minus actual (tracking only) |
| `cleared` | Whether the Allocator published or withheld (stale/partition) |
## Graduation
Specs do not define numeric exit criteria. RFC #1 sequences the work as **local proof first**, then federation.
**Phase 0 → Phase 1** is appropriate when the pocket can bind Commitments to self-dispatch **inside existing Grid 1 rules** (connect-and-manage / host-utility parameters), still meeting NERC, FERC, utility, and TSO/ISO obligations. Phase 1 is the first interval in which executed MW may follow `commit-%` rather than the prior baseline.
**Phase 1 → Phase 2** is appropriate when a second domain runs the same spec and Allocators exchange G2P federation announcements. Partition MUST NOT stop local clearing and MUST NOT trip Members in unaffected domains. Missing announcements clear as **zero extra boundary headroom**.
Phase 1 does not require a market-rule change. Grid 2 sits below ~5-minute market dispatch and above sub-second ancillaries, so a voluntary pocket can go live without replacing Grid 1 settlements (`architecture/temporal-position.md`).
## Verification signals
A Phase 0 run is healthy when:
- Every interval has either a Commitment that names `interval`, `domain`, and inputs version, or an explicit withhold (stale/partition).
- Replaying stored inputs through a second GCAP implementation reproduces the same allocations.
- Physical telemetry matches the site's **pre-Grid 2** schedule or firm limit, not the Commitment.
- Aggregate ramps, if you simulated applying Commitments, stay inside the host-agreed ramp rates and the domain pool-sizing bound.
- Grid 1 measures (frequency, ACE, flows) are unchanged by the overlay software, because the overlay is not yet moving the POI.
Grid 1 custodians verify the overlay by **aggregate footprint**, not by inspecting internal workloads (`architecture/overview.md`). The Allocator never sees facility-internal mapping of jobs onto A / B / C.
## Failure modes and open items that block a closed-loop test
G2P §7. Interoperability between two vendors in one pocket is a local agreement until encoding and test vectors land. Report the encoding you used in the Discussion.
Failsafe says allocations are interval-only and a missed `commit-%` is fallback. Quantitative missed-interval thresholds and re-entry hysteresis are open. In Phase 0, fallback equals the behavior you are already executing — still log withhold vs miss.
Commit-before-open is required in spirit (`architecture/temporal-position.md`) but the millisecond/second budget is an Engineering discussion. Do not claim a spec deadline.
G2P §3.4 verification granularity, metering source of truth, and attestation are unspecified. Domain certification criteria for “spec compliance” are an open item on allocation domains. Phase 0 reports should state how actuals were measured.
Do not treat federation announcements, loop prevention, or import policy as Phase 0 exit gates.
Do not size or operate a shadow pool that would be unsafe if software later bound Commitments by mistake. Ramp discipline and whole-pool fallback limits are normative now, including in dry-run.
## Related pages
The Grid 1 baselines Phase 0 executes on purpose, and that Phase 1+ must revert to on fault.
Host utility plus one flexible load: the on-ramp pocket for a first shadow run.
The four message families, one-minute clock, and authentication MUSTS the dry-run still exercises.
No-discretion clearing, withhold-on-stale, and constraint-authority inputs.
What Phase 2 adds: announce-and-listen, conservative missing-announcement clearing.
Where Phase 0 implementation reports go (Discussions) and how they feed the next spec revision.