0012 — Do not archive on low use

Status: Accepted Date: 2026-08-25 Deciders: elf, Ben

Context

v0.17.1 added lexical use-attribution: a Stop hook scores the turn's answer against the blocks retrieved for it and calls record_use() on the ones whose distinctive terms show through (elfmem.memory.attribution). It exists because reinforcement counts retrievals, so a block retrieved constantly and never drawn on rises exactly like one doing the work.

The obvious next step was proposed immediately: make curate() use-aware — archive blocks with high assembly and near-zero use. It was proposed on the strength of a single suggestive case (a completed maintenance errand, 1757ad77, that kept surfacing in the identity frame) and an argument from biology: a living system does not strengthen a memory for being considered.

Before building it, the rule was tested against real data rather than more argument: all 148 active blocks scored against 161 real assistant responses drawn from seven months of transcripts in this repository.

Alternatives considered

• Archive high-assembly / low-use blocks. The proposal under test.
• Penalise low use (accelerate decay rather than archive). Same signal, softer action; the measurement below condemns both equally.
• Reward-only, never penalise. What v0.17.1 actually shipped.
• Record use, act on nothing. Ledger telemetry with no ranking effect.

Decision

Rejected. The rule inverts on real data. Ranking the corpus by the exact criterion proposed — highest reinforcement among never-echoed blocks — the top 15 archive candidates are 10 constitutional blocks, including the two most-reinforced blocks in the entire corpus:

group	n	mean hits	never echoed	mean reinforcement
self/constitutional	39	0.87	54%	16.8
everything else	109	0.75	72%	3.4

The failure is not that constitutional blocks are invisible to attribution — they are echoed slightly more than average. It is that the rule keys on high reinforcement, and constitutional blocks carry 5× the reinforcement of everything else because the SELF frame guarantees them slots on every retrieval. Combine "high reinforcement" with "expressed as disposition rather than vocabulary I quote" and the constitution is precisely what the rule selects. A use-aware curate() would archive elf's constitution first and keep the block reading "test entry for CLI verification" (11 echoes, the second-most-echoed block in the corpus).

The originating case does not survive either. The stale errand and the live identity block are indistinguishable by use: 4 echoes versus 6. The errand shares identity vocabulary, so it is credited whenever identity is discussed, regardless of whether it contributed anything.

What this reveals about the signal itself: lexical attribution measures topical vocabulary overlap. That is a reasonable proxy for use when scoring is conditioned on retrieval — the shipped hook only ever scores blocks that were actually injected into that turn. It is not a proxy for value, and it must never be inverted into evidence of disuse, because a block can fail to echo for three unrelated reasons: it was not used, it was paraphrased, or it shapes how the answer is written rather than what it says. The corpus cannot tell those apart, and the third is what constitutions are made of.

The reward-only asymmetry shipped in v0.17.1 is what makes this a finding rather than a corrupted corpus. Had the penalty side been built at the same time, the first curate() run would have decayed the constitution.

Consequences

• curate() stays decay-and-graph based. No use term.
• record_use() remains reward-only. The asymmetry is now load-bearing and documented as such, not a conservative default awaiting confidence.
• use events keep accruing in the ledger. They are honest telemetry and cost nothing; this ADR rejects acting on them, not recording them.
• The calibration test in tests/test_attribution.py was measuring the wrong false positive — unrelated blocks, when the real failure is related-but-unused. Extended accordingly.

Revisit when

Any one of these would justify reopening:

1. Conditioned data disagrees. This test is unconditioned: it scores blocks against responses they were never retrieved for, which inflates echo for any block sharing a topic's vocabulary. Several thousand real assembled-then-judged turns from the hook would be the honest measurement, and could still separate use from mere topicality.
2. Attribution stops being lexical. A judge that distinguishes "informed this answer" from "shares vocabulary with this answer" removes the confound this ADR turns on.
3. The constitution stops being guaranteed slots. The inversion is driven by SELF-frame guarantees inflating constitutional reinforcement. A frame design that does not guarantee would change the arithmetic.

References

• src/elfmem/memory/attribution.py — the signal and its calibration
• tests/test_attribution.py::TestCalibration — false-positive bounds
• ADR 0009 — retire decay-driven archival (the prior time an archival rule was retired for selecting the wrong blocks)