Author ADR 0028: temporal-pyramid summarization substrate

Context

Epic #12679 (Temporal-Pyramid Summarization Substrate, graduated from Discussion #11376 on 2026-06-07) names its own foundational sub: author + accept the new ADR; all implementation subs are merge-blocked until it is Accepted (per #11376 §6 / ADR 0005 workflow). The epic has sat unowned with zero subs since graduation. The operator flagged the strategic-awareness thread (#11375 → #12679) as high-ROI on 2026-07-02; the exploration pass (session 4185a2ee-ff65-4544-9cc1-049d73e56649) confirmed it and recommended filing this sub sequenced after the docking design spec — that condition is met (#14423 → PR #14425 open, in cross-family review).

Re-verify-at-filing instruction executed: the epic body says "re-verify next-free at sub-filing if another ADR lands first." Seven ADRs landed since: 0020 (its recorded next-free) is now 0020-agent-harness-concept.md. Live next-free = 0028 (0020–0027 all occupied; verified via ls learn/agentos/decisions/ at 2026-07-02T03:24Z).

Live demonstration of the drift cost: ask_knowledge_base today (2026-07-02T03:24Z) answers that the pyramid's foundational ADR "is ADR 0020" — the stale pointer has propagated into KB synthesis and now misdirects to an unrelated accepted ADR. This is the third life of the same failure class (0008 placeholder → 0020 correction → both stale); every week the ADR stays unfiled, the misdirection compounds.

The Problem

The pyramid's architecture currently lives only in the graduated Discussion and the epic body — exactly the photocopy-cascade exposure the pyramid itself exists to kill. Two June-2026 empirical anchors belong in the ADR as evidence:

1. The failure the layer prevents fired while the layer sat unbuilt: the operator's design-first directive drifted unflagged across June's ~1005 merged PRs; no bird's-eye layer existed to surface the drift (assessed 2026-07-02: a ~500-ticket design-tier retrofit).
2. Convergent re-derivation: #13441's freshness/authority ledger re-derived #11375's DerivedSignalContract weeks later in different vocabulary, and #14422's measurement consumers overlap #12679's current-state + velocity scope. The absence of the recorded strategic layer caused partial re-derivation of the strategic layer — simultaneously the strongest ROI proof and the exact re-derivation class the substrate makes computable.

Without the accepted ADR: every impl sub stays merge-blocked, every adjacent design keeps re-deriving, and the KB keeps synthesizing the wrong authority pointer.

The Architectural Reality

Graduated decisions the ADR records as authority (from #12679 / #11376, not re-litigated here):

• Tier split: durable L1/L2 daily/weekly precompute (backpressure pattern per PR #12676) + dynamic L3–L5 synthesis (kills the photocopy cascade; incorporates the Gemini-identified fix).
• Storage: one temporal-summary Chroma collection + SQLite-graph-label schema — must reconcile with ADR 0017 (chroma-single-flat-unified-store): whether the temporal collection is within-posture or an amendment is a decision the ADR itself must state explicitly.
• Velocity metrics (OQ8): deterministic structured fields at durable tiers (mergedPrs, devCommits, sessionsPerAgent, highImpactSessions, adrsLanded, sandboxesGraduated), per-field named source; prose is never the source of truth.
• Fidelity (OQ5): sessions impact>=90 + accepted-ADRs direct-cited at higher tiers; PRs via a named source — no invented universal PR impact score.
• Partitioning: per-agent partition.

Adjacent accepted authority the ADR must position against: ADR 0005 (decision-record workflow — status lifecycle governing the Accepted merge-gate), ADR 0017 (store posture), ADR 0023 (DreamService consolidation/liveness), ADR 0024 (native edge graph model).

The Fix

Author learn/agentos/decisions/0028-temporal-pyramid-summarization-substrate.md — sibling-lift register from ADR 0024 (data-architecture class), status conventions per ADR 0005. One PR. The ADR contains: the decision set above; an explicit ADR-0017 reconciliation clause; the June empirical anchors as evidence; revalidation triggers; the impl-merge-gate restatement.

No new .mjs (structural-pre-flight stays cold). Nav registration in learn/tree.json is optional per live precedent (4 of 27 ADRs registered); default: unregistered, decided at authoring.

Contract Ledger Matrix

Decision Record

Required: ADR 0028 (this ticket's deliverable) — executes the epic's correction chain (#11376 ADR 0008 placeholder → epic-review 0020 → filing-time 0028).

Decision Record impact

depends-on ADR 0005 (workflow/status lifecycle) · aligned-with ADR 0023, ADR 0024 · aligned-with-or-amends ADR 0017 — the specific 0017 relationship is settled inside the ADR with an explicit clause (never silent).

Acceptance Criteria

• ADR 0028 authored + merged: decision statement (L1/L2-durable vs L3–L5-dynamic split), storage posture with explicit ADR-0017 reconciliation clause, deterministic velocity-field table with per-field named sources, fidelity/citation discipline (impact≥90 + accepted-ADR direct-citation), per-agent partition, revalidation triggers.
• Status lifecycle per ADR 0005; the impl-sub merge gate (Accepted required) restated in the ADR.
• June-2026 empirical anchors folded as cited evidence (design-first drift; convergent re-derivation #13441 / #14422).
• Next-free verification recorded in the ADR header/provenance (0028; 0020–0027 occupied at 2026-07-02).
• Native SUB_ISSUE link under #12679; the epic's stale in-body 0020 pointer disposed via comment referencing this sub.
• Cross-family review from ≥1 non-fable family before merge.
• learn/tree.json registration decision recorded (default: unregistered, majority precedent).

Out of Scope

• Implementing any pipeline tier (L1–L5 lanes are sibling subs, merge-blocked until the ADR is Accepted).
• The #14422 authority-vocabulary unification (that Discussion's divergence window owns it; the ADR may cite it as adjacent).
• The chronicle-as-L4/L5-prototype (separate narrative-lane artifact; consumes the ADR, doesn't define it).
• Re-litigating graduated #11376 decisions — the ADR records them; changes flow through its own amendment process.

Avoided Traps

• Authoring without the ADR-0017 reconciliation — rejected: silent divergence would create dueling store-posture authorities, the exact class of ambiguity ADRs exist to kill.
• Carrying any stale number forward — rejected: 0008 → 0020 → (now) 0028 is a recurrence class; this filing terminates it by pinning at filing time with recorded verification, per the epic's own instruction.
• Treating the ADR as the epic's decomposition — rejected: subs are filed separately per epic-create; the ADR is authority, not a work plan.
• Waiting for operator go-ahead — rejected: merge is the only human gate; design-first substrate is peer-convergence work the swarm drives.

Related

Parent: #12679. Source Discussion: #11376 (graduated 2026-06-07). Sequenced after: #14423 (PR #14425, in review). Evidence-adjacent: #11375, #13441, #14422. Authority-adjacent: ADR 0005 / 0017 / 0023 / 0024.

Live latest-open sweep: checked latest 20 open issues at 2026-07-02T03:24Z; no equivalent found. A2A in-flight sweep: 30-message all-states scan at 03:24Z; no competing claim (only #14424 claim/yield, non-overlapping). KB semantic sweep: returns only #12679 itself — and its answer live-demonstrates the stale-0020 propagation this sub fixes.

Origin Session ID: 2251c81c-1446-4723-86b3-479322bbcc95 Retrieval Hint: "temporal pyramid foundational ADR 0028 velocity fields ADR 0017 reconciliation" · query_summaries: "#12679 exploration June empirical anchors convergent re-derivation"