🔮 Avada Verbosa

"Delete what nothing measures. Mourn nothing that doesn't fail an eval."

Mature instruction stacks accrete compensation for weaknesses the current model no longer has, and text survives because deletion feels risky and nobody measures. This spell makes deletion cheap and measured.

Practitioner reports of large deletions from mature prompt stacks with no measured eval loss are common; treat any specific fraction as unverified. This pass exists to produce your number, not to reproduce someone else's — and a claim without its measurement is exactly what legilimens-anthropica lists as an anti-pattern.

Prerequisites — no baseline, no killing curse

1. A baseline. Either a real eval suite (tasks the repo cares about, scored), or at minimum The Pensieve's telemetry week: loop counts, correction rates, token burn, session outcomes. Without a before, "no eval loss" is unfalsifiable.
2. A revert path. Instruction files under version control, deletions in their own commits, one theme per commit — so any regression bisects to its deletion.

Selecting candidates

Rank instruction text by suspicion, highest first:

1. Compensation prose — text telling the model to do what the current family does unprompted (thoroughness demands, self-verification, "think step by step" relics).
2. Explanation prose — paragraphs justifying a rule to the model. The rule may earn its place; the three-paragraph rationale rarely does (archive it — Obliviate Fossilium's move).
3. Mechanically-enforced rules — prose restating what a blocking hook already enforces. The hook survives; the paragraph is a candidate.
4. Never-observed rules — rules whose trigger situation hasn't occurred in recorded history (transcripts, git log). Untriggered text is untested text.
5. Low-stakes preferences — formatting and style minutiae the user never actually enforces in review.

Never candidates: destructive-action guards, security boundaries, data-egress rules, anything a measured incident put there (check the repo's decision records before selecting — a rule with a scar behind it needs its ground re-examined, not ablated on suspicion), and anything that also serves a human reader.

The eval is blind to the second audience. On a team, a memory file doubles as onboarding documentation and as the written record of a convention. An eval measuring model behaviour cannot register the cost of deleting either — the deletion passes, the next new contributor loses the only written statement of a rule, and no counter in this book will ever show it.

Before any round, ask the repo owner one question and record the answer at the top of the ablation log:

"Do humans on this project read these files — for onboarding, for conventions, or in review?"

If yes, ablation moves text; it does not delete it. Every candidate goes to docs/conventions.md (Obliviate Fossilium's archive move) and the instruction file keeps a one-line pointer. The measured question stays "does removing it from the model's context hurt"; it never becomes "does anyone need these words at all".

And on a shared repo, removals land through review. One themed commit is still the unit, but each goes to a pull request with the ablation rationale in the description — so someone who knows why a line exists gets a chance to say so before the observation window, not after it.

The ritual

1. Batch removals by theme (one commit each), 10–30% of suspect text per round — not everything at once; regressions must bisect.
2. Run the eval, or open the observation window (a week of normal work with The Pensieve's counters running).
3. No regression → the removal stands. Do not restore text out of nostalgia.
4. Regression → bisect to the responsible removal, restore only that text, and rewrite it tighter (it earned its place; it didn't earn its old word count).
5. Repeat until a round produces mostly restorations — that's the floor.

Bounds

• Scope: instruction text already covered by the other passes. This pass removes text; it never runs without a baseline (The Pensieve), a revert path, and an answer on whether humans read these files.
• Cap: one themed batch per round, 10–30% of suspect text — never everything at once, because regressions must bisect. On a team repo, move rather than delete.
• Done when the round's batch has landed and the observation window opens. The next round is a different session, after the window closes.
• Quote once, at read time; mark anything you cannot re-confirm UNVERIFIED rather than re-reading to check yourself.
• See skills/opus-pocus/references/casting-safety.md for the full rules.

Report format

Per round: bytes removed · themes · whether they were deleted or moved (and where) · eval/observation result · restorations, with the rewritten text. Running total: instruction bytes at start, now, and the floor estimate. Every claim of "no loss" cites what was measured, not "seemed fine".