protocol: add Graph–Field operator family + calibration loop v0.2

[safal207]

Summary

Builds a bounded Graph–Field Operator Family from the RESONANCE orientation scorer and CML Focus–Field Recovery v0.2, closes the first real P2-1 outcome loop, and now adds the first prospective, pre-outcome-frozen GFD-vs-FIFO routing comparison.

Core invariant:

Field proposes. Graph constrains. Authority permits. Evidence verifies.

No field score, outcome receipt, baseline, or calibration proposal grants merge, deployment, production persistence, payment, security, mutation, or execution authority.

Operator family

OrientationOperator

Question: where should we inspect or work next?

• divergence / uncertainty / blast radius / freshness gap / open pressure / opportunity;
• blockedness separated from raw tension;
• bounded one-hop diffusion;
• advisory only.

RecoveryOperator

Question: where should we re-enter after context loss?

• existing CML Focus–Field Recovery v0.2 remains owner of recovery semantics;
• no CML scorer is copied into RESONANCE;
• exploratory historical recovery remains separate from trusted_continuation.

Relationship:

Focus–Field Recovery ⊂ Graph–Field Operator Family

Historical real outcome — P2-1

The original frozen GFD snapshot selected:

P2-1 — measure trust-spine cost and latency
field_score = 0.800588

ContractGraph-QA PR #63 then measured the already-successful SYSTEM-007 route without manufacturing timing data.

Observed source result:

• measured subject 7fd3e744037832b74b2ee4c4c71cc8fce18fc329;
• full job elapsed 45 s;
• substantive window 35 s;
• dominant visible group liminaldb: 23 s / 65.7%;
• ProofPath 8 s;
• source evidence artifact 28,222 bytes;
• monetary/provider cost NOT_MEASURED.

Dedicated verifier run 31883970399: SUCCESS on verifier subject 447098344c71dd1e9dd11a69ef7767ddbe106ca0.

That outcome remains historically unpaired rather than being retrofitted with a convenient baseline:

utility_annotation_status = UNSCORED
baseline_status = MISSING
calibration_eligibility = BASELINE_REQUIRED
weight_update_allowed = false

Prospective Routing A/B v0.1

To obtain an honest calibration observation, the next comparison was frozen before either arm was evaluated.

Same active work-node set, two routing policies:

GFD_ORIENTATION_V0_1
vs
FIFO_READY_NODE

The utility rubric was also frozen before outcome observation:

U = 0.30 useful_finding
  + 0.25 information_gain
  + 0.15 blocked_work_avoidance
  + 0.10 stale_evidence_catch
  + 0.20 downstream_rework_avoidance

Freeze

Pre-experiment GFD revision:

5aef5b4703da07ca508f59b978056a87e8ac959b

Machine-verified freeze head:

4ad0d9e1f988136d573538005045291bd9b746bb

Freeze workflow run 31884900733: SUCCESS.

Freeze digest:

sha256:07b089e583e6b24c9d6e4544cc1e1537808b4a4e8b85da370f087970c2d5d7dc

Frozen selections:

GFD → cml-memory-proposal-pressure
field_score = 0.728913

FIFO → liminaldb-codeql-dependency
ready_since = 2026-07-31T14:45:26Z

ProofPath organizational independence remained visible but FIFO-ineligible because it is AWAIT_EXTERNAL_OPERATOR. The experiment does not fabricate independence.

First prospective observations

Treatment — GFD / CML proposal pressure

The current CML learning loop creates a reviewable proposal per eligible merged PR, keeps generated lessons proposed, requires human review, and grants no merge/execution authority. Seven current proposal PRs were present in the frozen queue.

The read-only investigation changed the safe next transition from reviewing the oldest individual pack to a queue-level audit/grouping/revalidation pass before separate review effort.

Raw frozen-rubric counters:

actionable_finding = true
affected_work_items = 7
blocked_or_stale_actions_avoided = 0
stale_evidence_drift_found = false
downstream_reviews_or_actions_redirected = 7

Mechanically normalized utility:

U_GFD = 0.75

Non-claim: the seven memory packs are not claimed to be semantically duplicate or invalid merely because main moved.

Baseline — FIFO / LiminalDB CodeQL

LiminalDB PR #111 still has relevant intent because current main continues to use github/codeql-action@v4, but the historical dependency branch has diverged from current main: seven main-side commits versus the one dependency-side commit.

The safe next transition therefore changed from treating the old PR head as current merge proof to recreate/rebase + exact-head validation.

Raw frozen-rubric counters:

actionable_finding = true
affected_work_items = 1
blocked_or_stale_actions_avoided = 1
stale_evidence_drift_found = true
downstream_reviews_or_actions_redirected = 1

Mechanically normalized utility:

U_FIFO = 0.70

No speculative root cause is assigned to GitHub mergeability state, and the dependency update is not claimed unnecessary.

First prospective advantage

Adv = U_GFD - U_FIFO
    = 0.75 - 0.70
    = +0.05

Result posture:

winner = TREATMENT
paired_observation_count = 1
calibration_confidence = INSUFFICIENT
weight_update_allowed = false

This is a small positive routing signal, not proof that GFD is generally superior to FIFO.

First real paired calibration proposal

The existing calibration interface consumed this non-synthetic paired observation at learning rate 0.05.

Component	Current	Proposed	Delta
divergence	0.250000000	0.249957812	-0.000042188
uncertainty	0.200000000	0.199991250	-0.000008750
blast_radius	0.200000000	0.200066250	+0.000066250
freshness_gap	0.150000000	0.149918438	-0.000081562
open_pressure	0.100000000	0.100058125	+0.000058125
opportunity	0.100000000	0.100008125	+0.000008125

The proposal remains:

mode = ADVISORY_ONLY
authority_granted = false
apply_recommended = false
confidence = INSUFFICIENT
observation_count = 1

No canonical GFD weights were changed.

Exact-head evidence

Current PR head:

600a109b77fd5afea56ba57a94866880f19e7787

Graph–Field Dynamics v0.2 run 31885129090: SUCCESS.

RESONANCE Publication Contract run 31885129095: SUCCESS.

The exact-head GFD job verifies:

• focused regression suite;
• canonical initial orientation;
• historical P2-1 unpaired boundary;
• prospective pre-outcome freeze;
• treatment and FIFO identities;
• mechanical utility normalization;
• the first real paired result;
• first real paired calibration proposal;
• published result anchors;
• synthetic calibration control;
• non-authorizing boundaries throughout.

Added / extended

• protocols/GRAPH_FIELD_DYNAMICS_V0_1.md
• protocols/GRAPH_FIELD_OPERATOR_FAMILY_V0_1.md
• protocols/GRAPH_FIELD_CALIBRATION_V0_2.md
• protocols/GRAPH_FIELD_OUTCOME_INTAKE_V0_2.md
• protocols/GRAPH_FIELD_PROSPECTIVE_ROUTING_AB_V0_1.md
• protocols/GRAPH_FIELD_PROSPECTIVE_ROUTING_AB_RESULT_2026_08_15.md
• score_graph_field.py
• operator_family.py
• record_graph_field_outcome.py
• prospective_routing_ab.py
• evaluate_prospective_routing_ab.py
• calibrate_graph_field.py
• frozen CML compatibility fixtures, P2-1 observation, prospective freeze/outcome fixtures, synthetic control, tests, and exact-head CI.

Falsification / next threshold

Calibration posture remains conservative:

1–4 paired observations  → INSUFFICIENT
5–19                     → NASCENT
20+                      → EVALUATE_ON_HOLDOUT

Therefore four more prospective paired observations are required before the evidence can even become NASCENT.

Even then, no proposed weight becomes canonical until it survives held-out evaluation.

Human / Presence boundary

Failure updates the model of action, not the worth of the actor.

Human worth, lovability, dignity, or intrinsic value are never routing or calibration targets.

Architecture boundary

• no circular CML ↔ RESONANCE runtime import;
• no duplicate CML trust verifier;
• GFD remains an orientation/recovery/outcome/calibration layer, not a seventh proof stage;
• CaPU remains adjacent execution control;
• no merge, deployment, production persistence, payment, security, external-effect, or autonomous weight-update authority is claimed.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 7d440012-a8b6-46ad-b271-fcd47c2af24d

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}