Tempus Mutatur

---
name: tempus-mutatur
description: Apply a model-era delta to this repo's instruction layer — establish from evidence which model family the files were written for, then list what current guidance changes, as edits to your files rather than as research. Use for "these instructions were written for an older model", "what changed for the 5-family", "are we still writing for Sonnet 3.5", "is any of this advice out of date", "we have a model ID hardcoded somewhere", or as part of /opus-pocus. Consumes legilimens-anthropica's digest when one exists. NOT for fetching the guidance itself (→ legilimens-anthropica, which gathers inputs; this one produces edits). NOT for removing text that is merely unused (→ avada-verbosa).
---

# 🔮 Tempus Mutatur

> *"You wrote this for a model that no longer exists."*

Instruction layers are written against a model and then outlive it. The advice that was correct for
the family you started on becomes, at best, dead weight — and at worst an active instruction to
compensate for a weakness that is gone, which the current family will obey.

`legilimens-anthropica` answers *what does current guidance say*. This pass answers the different
and more useful question: *given what it says, what in these specific files is now wrong?* One
gathers inputs; this one produces a change list.

## What to scan

- The resolved memory files and rule files, plus their **git history** — when were they written,
  and against what
- Skill and agent frontmatter, for pinned model identifiers
- Scripts, CI config and settings files, for model IDs and API parameters
- Any dated decision records the repo keeps
- `legilimens-anthropica`'s digest from this session, when it ran

**Establish the era from evidence, not assumption.** A file's first commit date, a model ID named
in it, a decision record's date, or a piece of advice that was current in a known window. Say which
evidence you used. If the era cannot be established, this pass is `? — could not verify` — the
delta between an unknown era and today is not a finding, it is a guess.

## Anti-patterns to find

1. **Compensation for a fixed weakness.** Text instructing the model to do what the current family
   does unprompted — thoroughness demands, self-verification prose, step-by-step relics, elaborate
   role-play framing. It was load-bearing once, which is why it reads as harmless now.
2. **A dated model identifier that has aged out.** A specific model ID in agent frontmatter, a
   script, a settings file, or CI. IDs are retired on a schedule; an alias is not. Distinguish a
   pin that is deliberate (reproducibility, a known-good evaluation) from one that is merely old.
3. **Advice current guidance reverses.** Not "no longer needed" but "now counterproductive" — the
   most valuable class, and the one nobody finds by re-reading their own file, because it still
   reads exactly as sensible as the day it was written.
4. **Capability the layer does not use because it did not exist.** Path-scoped rules, newer hook
   events, frontmatter fields, plugin mechanisms. The absence is invisible: nothing errors, the
   layer just does by hand what the harness would now do for it.
5. **Era mismatch between files.** One file written for the current family, another two
   generations back, both live and both obeyed. The older one usually wins in novel situations,
   because it was written when instructions had to be broader.
6. **A version fact stated as a constant.** A hardcoded cliff, cap, limit or version number that
   was measured once and typed. It ages silently and is quoted forever — the same defect
   `obliviate-fossilium` calls a typed fact.

## Fix prescriptions

| Finding class | Fix |
|---|---|
| Compensation prose | Remove, or narrow to the part that is genuinely non-inferable. A rule naming a concrete command survives; a rule asking for diligence does not. Cite the guidance that makes it redundant, with its date |
| Aged model ID | Replace with the alias unless the pin is deliberate. Where it is deliberate, say so in the file, so the next reader does not "fix" it |
| Reversed advice | Change it, and record why — an inverted rule with no note gets reverted by whoever remembers the old guidance |
| Unused capability | Name the capability, what it would replace, and hand it to the spell that owns the migration (`accio-rulebook` for path-scoped rules, `sorting-hat` for tier fields, `protego` for hook-enforceable checks) |
| Era mismatch | Bring the older file forward or retire it. Two eras coexisting is a `contradicto-revelio` finding waiting to happen |
| Typed version fact | Replace with the derivation, or with the pass that measures it. A number nobody re-measures is wrong on a schedule |

## Bounds

- **Prefer `legilimens-anthropica`'s digest from this session over anything remembered.** Guidance moves faster than any snapshot in this file, and training data is not a source.
- Scope: this repo's instruction layer and the config that pins models for it. Not application source, not dependency versions, not another plugin's files.
- Cap: at most 15 findings, aged model IDs and reversed advice first — those two classes act today; the rest are hygiene.
- Done when the layer has been read once against the established era. Not when the era feels fully mapped: there is always another guidance page.
- Offline, or where the delta cannot be sourced, report `? — could not verify` per finding rather than reasoning from training data. Guessing a version delta is worse than declining one.
- 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

Lead with the era finding: which family these files were written for, and the evidence that
established it (commit dates, model IDs, decision records). Then per finding: `file:line` ·
class (1–6) · the quoted text · what changed · the prescription · the guidance source **with its
date**. A prescription citing no source is a `?`, not a finding.

`SCANNED NOTHING` if no instruction layer was found; `? — could not verify` if one was found but
its era could not be established, or if no current guidance was reachable.