Theta Consolidation

ThetaAgent executes a deterministic consolidation pipeline for cross-temporal knowledge linking. Named after theta rhythms in hippocampal replay (Zhang et al., 2015), it compresses temporal experience into durable knowledge connections.

The agent operates in two phases: SITREP (situational awareness) and Consolidation (cross-temporal linking).

Phase 1: SITREP

The /sitrep command generates a structured SituationReport through three components:

Horizons — Time-based attention windows inspired by hippocampal theta waves:

AttentionSchema — Models what information is currently attended to. Targets have priority (0.0–1.0) and decay rates. Active targets above a threshold are included in the report.

SituationReport — Aggregates objectives, recent thoughts, agenda items, current events, and capability notes into a single markdown-formatted output.

Phase 2: Consolidation

Stage 1: Temporal Slicing

Documents are organized into weekly slices (YYYY-WNN). The Monday 06:00 cycle consolidates the previous ISO week for that run’s logical date (the week that closed relative to that Monday), not the empty week that starts that morning, and not “everything since the last success.”

DAG gaius_theta_cycle stays catchup=False: unpausing must not dump every missed Monday onto the LIGHT token (that window grows until Theta starves other Airflow workloads). Missed weeks are a persistent failure.

Remediate from the Signals tree with daily LIGHT increments (just theta-backfill --from-date … --to-date …). Each UTC day encodes that day’s thoughts and merges centroid, CLT, and groundings into one theta_consolidation_runs ledger row for the containing ISO week. Completing that row is LIGHT coverage, not the cortical product. max_active_runs=1 serializes the days. --weekly is the coarse one-Monday-run path.

A backfill of a missed week starts on that week’s earliest UTC day and walks toward the present. That is the default. --latest-first is the old order and does not approximate how the week would have consolidated. Do not open the week as one out-of-band LIGHT occupancy. Leave the week in progress to Monday 06:00. One finished week is a point; the next natural consolidation is the start of a trajectory.

The core objective is to converge on the best OWL-grounded SKOS-in-Qdrant aperture for the week-or-longer timeframe under consolidation, and to stay effective when that aperture is backtested on the accumulated corpus. The ontology that aperture must satisfy is the entailed stack OWL ⊨ SKOS ⊨ SHACL (Aegir): SKOS is an entailed annotation face of certified OWL, and SHACL is the closed-world view of that same OWL. Not a third vocabulary, and not a set of hypothesized class–class links.

Class–class “pairs” — CLT co-activation scored by BERTSubs — are not that design. They were carried forward from the subsumption-alignment habit Atelier has already retired (the subsumption/LLM-mediated alignment dance is gone; see Atelier docs/src/architecture/classification.md). Do not refine the pair count, the 0.8 gate, or the TBox-holdout MRR as if they measured the objective. A completed ledger row is not the objective either. Replacement measures have to speak to efficacy of the entailed stack.

The authentic Signals data product (gaius.theta.consolidation) has not been published. Intended product split and evidence planes (booked Gaius/Theta session + rustfs materials, week (X_i), CLT/MaxSim week binding, replaceable KB week-block, ShadowStrategy aperture delta, sealed transcript, History) are considered in docs/scratch/2026-09-20/160356_theta_consolidation_data_product_aspects.md. AgentRTC wiki current/design/data-product-aspects.md is non-normative vocabulary (Aspect = contract, not the blob; session series is a distinct product). The ledger remains for Nautilus / sitrep.

Stage 2: NVAR Dynamics

Nonlinear Vector AutoRegression via reservoir computing (Gauthier et al., 2021) computes a consolidation urgency signal from embedding centroid trajectories.

Given slice centroids c_1,…,c_t in R^128 (ColBERT-Zero agg centroids):

• NVAR predicts c_hat_{t+1} using a polynomial basis over delayed embeddings
• Drift = ||c_hat_{t+1} - c_t||_2
• Urgency = sigmoid(alpha * drift)

High urgency indicates rapid semantic drift — the knowledge base is changing faster than consolidation is linking it.

Stage 3: Subsumption Inference

Candidate A ⊑ B hypotheses are scored by BERTSubs Intra (Chen et al., 2023, via DeepOnto) against one TBox. The layers are fixed and none is minted at run time:

The generated kb_current.owl and gaius_domain.owl of the 2025 pipeline are retired: minting owl:Class from markdown or thought tokens was a category error. Requires the JVM (JPype) that DeepOnto starts.

Stage 4: Knowledge Gradient Selection

Candidate relationships are filtered using the Knowledge Gradient policy (Powell & Ryzhov, 2012). KG balances exploration (uncertain candidates) against exploitation (high-confidence relationships):

KG(x | S) = E[max_{x’} mu_{n+1}(x’) | S_n=S, x_n=x] - max_{x’} mu_n(x’)

Only candidates whose expected improvement exceeds a cost threshold are selected. The BeliefState tracks prior value and observation variance for each candidate.

Stage 5: Document Augmentation

Selected relationships are injected into source documents as:

• Wikilinks: [[Target Document]] for navigation
• Action links: [action:search "query"] for deferred execution

Effectiveness Tracking

The EffectivenessTracker measures augmentation impact by recording whether users follow injected links. When SHAP is available, feature attribution analysis identifies which augmentation types and document characteristics predict user engagement.

ShadowStrategy

Gaius consumes external/sdg-strategy (SdgAperture). ShadowStrategy there means a branch of the strategy repo (lens / voices / knobs / targets), not a Theta time grain. See external/sdg-strategy/README.md (Shadows are branches).

Daily increments (zndx.window_date) are not shadows and not day-bounded releases. The ledger row completes when Monday–Sunday of that slice are incorporated. DAG catchup=False; just theta-backfill (Signals) serializes days with max_active_runs=1.

A shadow Theta comparison is: same ISO week, two strategy_ids, two week-level Aspect sets (B–E), not two SQL rows.

CLI Commands

# Situational report
uv run gaius-cli --cmd "/sitrep" --format json

# Enqueue ThetaCycleFlow (same vessel as Airflow Monday 06:00)
uv run gaius-cli --cmd "/theta consolidate" --format json

# View consolidation stats
uv run gaius-cli --cmd "/theta stats" --format json