Kjell Tore Guttormsen
b54e450c3e
Every long-form gate agent sees ONE edition, so «the reader has heard this
before» was structurally invisible — and the specifics-bank dedupe actively
encouraged re-surfacing the same material. At series cadence that is the
fastest-growing defect class. N11 makes it visible BEFORE the skeleton is
approved, at two grains:
- Series grain — new scripts/editions package: each locked edition's spent
anecdotes/arguments/hooks are written to <serie>/linkedin/series-distillate.json
at Step 8 lock (distil-append), and the next skeleton is checked against it at
Step 2.5 (distil-check) with the finding folded into the annotation gate the
operator already reads. Advisory, never blocking: a deliberate callback is a
legitimate move, an unnoticed retread is not.
- Material grain — specifics-bank usedIn log: record-usage stamps «used in
edition NN» on the specifics an edition actually consumed (read from the bound
slot-map, so abstrakt/ekstern stamp nothing). At lock, so it means published;
idempotent under a pivot re-lock. Additive-optional, schema stays v1.
Placement deviates from the plan text (${DATA}) after premise-verification:
per-series state belongs in the series root beside edition-state.json, where
Step 0 already resolves the path and no slug→path map is needed. Operator
approved. The distillate module also lands in scripts/editions now rather than
at N12, since the AC required a testable roundtrip and N12 planned that package
anyway — N12 extends it instead of creating it.
Similarity is character-trigram Jaccard, not word overlap: the plugin is
language-general and inflection (migrere/migreringen) breaks word tokens.
Calibrated on real paraphrase pairs — retellings 0.44-0.55, unrelated 0.04-0.06,
same-topic-different-story 0.24 — so the default threshold sits in the gap at
0.40. Word-Jaccard scored a shortened hook paraphrase at 0.11.
TDD (Iron Law): tests written first and verified red before implementation, for
both the new package and the bank logging.
Suites: editions 27/0 (new suite line + guard, floor 27) · specifics-bank
28 -> 45 · test-runner 163 -> 173 (Section 16r: 9 unconditional greps +
non-vacuity self-test; anti-erosion floor 146 -> 155) · trends 300/0 ·
brain 134/0 · hooks 140/0 · tests 35/0 · render 60/0. tsc --noEmit clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QxvWAjte7vPcF79QeSRvRJ
|
||
|---|---|---|
| .. | ||
| src | ||
| tests | ||
| package-lock.json | ||
| package.json | ||
| README.md | ||
| tsconfig.json | ||
editions — series-level memory for long-form (N11, serie-nivå-vern)
Every long-form gate agent sees one edition. At two editions a week the fastest- growing defect class is not a bad edition — it is a repeated one: the anecdote, the argument or the hook the reader already met three editions ago. That failure is structurally invisible today, and the specifics-bank dedupe actively encourages re-surfacing the same material.
So each locked edition leaves behind a distillate of the narrative units it spent, and the Step 2.5 skeleton gate checks the next skeleton against it — before prose exists, the only point where changing course is still cheap.
Where the distillate lives
<serie>/linkedin/series-distillate.json — beside edition-state.json in the series
root, not in the per-user data dir. It is series-scoped state, the series folder is
already where the edition's state lives, and putting it there means it travels with the
series and needs no slug→path map.
Series-scoped is the point: this answers "has this reader heard it?". Cross-series
re-use of the operator's raw material is a different grain, tracked by the
specifics-bank's usedIn log.
Setup
cd scripts/editions && npm install
Use
# Step 8 (lock): fold the edition's spent units into the series distillate
node --import tsx src/cli.ts distil-append \
--distillate "<serie>/linkedin/series-distillate.json" \
--series seres --extract "<abs>/extract.json"
# Step 2.5 (before prose): check the proposed skeleton against everything published
node --import tsx src/cli.ts distil-check \
--distillate "<serie>/linkedin/series-distillate.json" \
--skeleton "<abs>/skeleton.json" [--threshold 0.4] [--json]
extract.json = {editionId, title, lockedAt, anecdotes[], arguments[], hooks[]} — the
AI extract from the locked edition. skeleton.json = {anecdotes?, arguments?, hooks?}.
Exit codes: 0 on success — including a REUSE verdict. The check is advisory by
design (it reports into the annotation gate the operator already reads at Step 2.5),
unlike the binding gate's BLOCK. 2 on usage error.
Model
- Extraction is AI, comparison is code. The command layer distils the units from the locked prose; the store, the similarity and the verdict are plain deterministic code — no model in the loop, no run-to-run variance.
- Similarity = character-trigram Jaccard, not word overlap. The plugin is
language-general and the languages it is used in are inflection-heavy: migrere /
migreringen are the same unit to a reader but different word tokens. Calibrated
against real paraphrase pairs (see
tests/distillate.test.ts): retellings score 0.44–0.55, unrelated material 0.04–0.06, same-topic-different-story 0.24 — so the default threshold sits in the gap at 0.40. Word-Jaccard scored a shortened hook paraphrase at 0.11 and would have let it through. - Units are compared within their kind. An argument re-used as a hook is a new move, not a retread.
- A re-locked edition replaces its entry, never duplicates it —
/linkedin:pivotre-opens and re-locks the same edition, and the reader still only met it once. - A missing distillate is not an error. A series' first edition compares against nothing and is CLEAR.
Test
npm test # node:test, deterministic (callers supply lockedAt)
npm run build # tsc — must stay clean