Concept-touch measurement: re-derivation rate + per-agent profiles

Context

Consumer-4 leaf of epic #14472 — the measurement substrate: re-derivation becomes computable, capability analysis gets territory maps. Filed in the operator-directed full-fledging pass; live latest-open sweep 2026-07-02T19:37:34Z: no equivalent ticket.

This consumer already has its ground-truth client committed: the substrate-effect study's prospective arm (per the graduation record — the study lead's codebook supplies labeled boundary events, with the honest bound stated there: concept-anchoring mechanically catches ~2 of the 5 codebook pressure-classes; this leaf promises exactly that, no more). It is also the instrumentation the public front door commits to (learn/agentos/WhatIsNeo.md §7 lineage).

The Problem

The institution cannot measure its own memory usage. "Agents re-derive what the Memory Core already knows" is asserted, never computed — because computing it needs concept-mediated observability: this concept was revisited + a relevant memory existed + it was not retrieved = one re-derivation event. Likewise per-agent concept-touch profiles (who works where in concept space, at what depth) exist nowhere, though the raw signal (TAGGED_CONCEPT with session provenance) has been written for weeks and consumed by nothing.

The Architectural Reality

• Raw signal: ai/services/graph/SemanticGraphExtractor.mjs writes TAGGED_CONCEPT (1.0 curated / 0.8 auto) with session provenance; ai/services/memory-core/MailboxService.mjs tags message-side. Historical depth exists TODAY — first slice needs no new emitters.
• Retrieval events ({query, resolvedConcepts, walkContributed}) arrive with the consumer-1 leaf (#14504) — the precision half of re-derivation detection sharpens when they exist; the recall half computes from history alone.
• Aggregation constraint (binding, from the graduation's OQ6 fold): profile aggregates carry propagation semantics — an aggregate's tier is the most-restrictive of its members' tiers (#14428 obligation); per-agent profiles must respect privacyTier boundaries (nothing private aggregates into a public surface).
• Normalization constraint (from the study covenant): metrics normalize by the agent's own deposit-history depth — nobody is scored against someone else's archive; and the output is substrate diagnostics, never a ranking (covenant language, load-bearing).

The Fix

One PR, history-first:

1. Concept-touch profile builder (ai/services/graph/, read-only): per agent × window → {conceptsTouched, touchDepth (hops from concept to deposited artifact), tierMix, revisitIntervals} — computed over existing TAGGED_CONCEPT history.
2. Re-derivation detector, slice 1 (history-based): concept revisited in session N+k while ≥1 relevant memory (attached to that concept's neighborhood) existed at N+k boot AND no TAGGED_CONCEPT/retrieval linkage shows it surfaced — emits candidate re-derivation events with confidence, not verdicts. Slice 2 (post-#14504) upgrades the "was it retrieved" leg from inference to the logged retrieval events.
3. Measurements artifact under learn/agentos/measurements/: swarm-level re-derivation-rate estimate + per-agent profiles (privacy-tier-respecting), with the ~2-of-5 coverage bound and the normalization method stated in the artifact header.
4. Study handoff: the artifact's schema section maps profile fields to the study's codebook classes so the prospective arm consumes it directly.

Contract Ledger Matrix

Decision Record impact

aligned-with ADR 0024. None minted.

Acceptance Criteria

• Profile builder + detector slice 1, hermetic specs, zero writes asserted
• Most-restrictive tier propagation spec-pinned (a private-tier member forces the aggregate private)
• Artifact committed: swarm re-derivation estimate + per-agent profiles, normalization + ~2-of-5 bound + not-a-ranking framing in the header
• Codebook-mapping section present (study prospective-arm consumable)
• Slice-2 upgrade path documented (consumes #14504's retrieval events; no re-architecture)
• PR Resolves this leaf, Refs #14472

Out of Scope

• Retrieval-event emission (#14504 owns it).
• Any capability RANKING surface (covenant-excluded, permanently).
• Dashboard/rendering (a later consumer; this leaf is data + artifact).

Avoided Traps

• Ranking by accident — per-agent numbers without the covenant framing become a leaderboard; the framing lives in the artifact header AND the module docblock.
• Tier-blind aggregation — one leaked private element in a public aggregate is the #14422-era taint class; most-restrictive propagation is spec-pinned, not convention.
• Verdict-shaped detection — slice 1 infers; events carry confidence and say so.

Sequencing

Executable NOW (history-based slice 1). Slice-2 precision upgrade lands after #14504.

Related: #14472 (parent epic), #14504 (event feed), #14428 (propagation obligation), Discussion #14422 (consumer 4 + measured-reality item 6)

Origin Session ID: 8cf234b7-e698-47ca-99e2-bf865196b6aa

Retrieval Hint: concept touch profiles re-derivation rate measurement substrate diagnostics not ranking