AGENT SKILL

eclipse-pass

The "debugger on" critique ritual — stress-test a proposition, design, artifact, or skill until the weak claims break, keeping the losing models instead of discarding them. Use whenever the user says "eclipse this", "eclipse/critique pass", "debug it" (in the critique sense), "debugger on/off", "run it through the ringer", "stress-test this", "poke holes", "red-team this idea", or wants a rigorous PASS/FAIL on whether an idea actually holds. (Legacy "torture this" still means this pass; the old name is renounced.) Checks the target against several seeded-random real projects (so nothing is cherry-picked), runs two opposed biases (doubt-first vs reward-first) and treats their convergence as the stability check, releases ethicsminions before the persona pass, and — debugger on — parks the losing/contrast models in .eclipse-increments/ for later folding. This is Corwin's own extreme-debug discipline (born in the C4/FoaF passes), not a unit-test runner, a code debugger, or a linter.

License 许可
GPL-3.0
所属仓库
inputloop/garden
最近更新
2026-10-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
eclipse-pass
description
The "debugger on" critique ritual — stress-test a proposition, design, artifact, or skill until the weak claims break, keeping the losing models instead of discarding them. Use whenever the user says "eclipse this", "eclipse/critique pass", "debug it" (in the critique sense), "debugger on/off", "run it through the ringer", "stress-test this", "poke holes", "red-team this idea", or wants a rigorous PASS/FAIL on whether an idea actually holds. (Legacy "torture this" still means this pass; the old name is renounced.) Checks the target against several seeded-random real projects (so nothing is cherry-picked), runs two opposed biases (doubt-first vs reward-first) and treats their convergence as the stability check, releases ethicsminions before the persona pass, and — debugger on — parks the losing/contrast models in .eclipse-increments/ for later folding. This is Corwin's own extreme-debug discipline (born in the C4/FoaF passes), not a unit-test runner, a code debugger, or a linter.

02 资源文件 28

03 SKILL.md 正文

Eclipse Pass — the "debugger on" critique

An eclipse pass exists to find where an idea is wrong before you build on it. The trap in self-evaluation is that a single reviewer, especially a helpful one, talks itself into the answer it already likes. This ritual defeats that three ways at once: it checks the target against projects it didn't get to pick, it runs two opposed biases and only trusts what they agree on, and — with the debugger on — it keeps the losing models so a contrast that looks wrong today can be re-examined tomorrow. Born in the C4/FoaF passes; it's how "C4 is imposed" was caught (it broke on the Dragaera project) and how the signed footprint was found to be the load-bearing pillar.

Naming note: this ritual was first called a "torture pass." Corwin formally renounced that framing (2026-06-29); kind language only. The mechanics are unchanged. Pre-existing .torture-increments/ dirs are his to keep or rename — read them if present; write new increments to .eclipse-increments/.

This skill runs the pass; it does not save memory ([[mentat-skill]]), decide whether something is skill-shaped ([[skilltank]]), or replace the ethics scan — it invokes [[ethicsminions]] as one ordered step. Because it scaffolds files on mounted folders, it follows the [[sandboxer]] discipline: verify writes land.

When this applies

  • The user says "eclipse this", "critique pass", "debug on", "run it through the ringer", "stress-test / red-team this", "poke holes", "does this actually hold?" (legacy "torture this" = this pass, answered in eclipse framing) — or asks for a PASS/FAIL on a proposition, design, artifact, or a skill under review.
  • You're about to build on a claim that hasn't been stress-tested, and want the weak parts to break now, cheaply, rather than after they're load-bearing.

The reflection family — where this pass sits

"Reflection" is a family, not one ritual. There are wider techniques and narrower ones, and this pass is the narrow end. Corwin's ladder (his words, 2026-09-07): "Lens is the common baseline that works very nicely for humans as a general thought experiment", and the microscope is "a scaling aparatus that reaches bubble chamber when we think of several AI working together over wire." The four rungs, as recorded:

  • LENS — the wide reflection, and his human baseline. One person holds the whole thing and turns it by hand; nothing needs capturing because nothing is being lost yet. He applies it in nearly all contexts, and several of his notebook pages are lens passes rather than notes. (The four steps a lens pass contains — four views apart · persona and its inverse · principles reframe · notes-or-stop — are his: confirmed 2026-09-11, and carried by the lens skill, which is the baseline this pass sits above.)
  • MICROSCOPE — the scaling apparatus between rungs, not a rung you stand on. One mind is no longer enough magnification, so you strip assumptions and keep only what survives with no trust in the scorer: keep-the-winners. This skill does not carry the microscope today; that gap is real and owed.
  • BUBBLE CHAMBER — several AI over a wire, each stateless and flushing at session end. The reasoning is gone before you can look at it, so you observe the track, not the particle; that is why the bubble-chamber method is capture-first (memory lens-microscope-bubble-chamber).
  • ECLIPSE — this skill. Narrow: one target, debugger on, keep-the-losers. Keep-the-winners (microscope) and keep-the-losers (eclipse) are not rivals; they live on different rungs and both are needed.

His rule for choosing: pick the rung by how much reasoning you can still observe, not by taste. If you can still hold every claim and every project in one head, that is a lens, not an eclipse; if the reasoning is already gone before you can look at it, you are in the bubble chamber and must capture first.

The five moves of a pass

Run in order — the order is part of what keeps it honest.

  1. Frame the target as falsifiable claims. Restate the proposition/design as a short list of claims that could be shown false. Vague praise can't be critiqued; specific claims can. (Watch the verbs — "imposed" vs "offered" was the whole bug once.)
  2. Seeded random check. Pull a few real projects at random from the known list — seeded and reproducible, so the selection can't be (and can't be accused of being) cherry-picked — and test each claim against each project. A claim that survives one project but breaks on another has found its edge. scripts/eclipse_pass.sh does the seeded pull and scaffolds the pass.
  3. Two opposed biases. Run the claims twice: once doubt-first (assume it fails; hunt the reason) and once reward-first (assume it wins; hunt the reason). Where the two converge is the stable result; where they diverge is an unstable claim that needs more work — convergence is the real verdict, not either pass alone. Before reading a divergence as an unstable claim, run the diverging bias once more on the same claim. If a bias disagrees with itself, the instability is in the reading, not yet in the claim. See lens, "A persona shows reasons, not decisions". 3a. Minions before persona. Release [[ethicsminions]] on the target before any persona-colored pass, so the ethics read isn't bent by the voice. This ordering is deliberate; don't collapse it. 3b. Multi-arm mode — when a second model is available. One head running two biases (move 3) is the floor. When a second model, or a second cold session, can run the same target, arms multiply what convergence proves: agreement across model, seed, and evidence base is strictly stronger than the two-bias convergence inside one head. Opt-in, not required — and not "debugger on" run twice.
    • Arm-mark everything. Each arm declares itself (arm N/M, model, seed) in its pass, verdict, and increments. Increment dirs already auto-suffix on collision (scripts/eclipse_pass.sh), so a later arm can't clobber an earlier one — never paste one arm's materials into another arm's session; a contaminated arm scores nothing.
    • Seeds stay disjoint on purpose. Different --seed values draw different grind projects, so each arm tests the target against evidence the other never saw — a divergence card must first ask whether the draw shaped the severity.
    • Write the CONTRAST doc after your own pass, never before — reading a sibling first anchors you. Sections: convergent (stable), divergent (kept, debugger on), additive (what only one arm tested).
    • Record the environment-delta, every arm receipt: which surfaces the arm could actually read (memory store, registries, mounts). A one-legged arm must say so — "the note was unreachable" is a finding, never a gap to paper over.
    • Guest etiquette travels. An arm running in a garden it did not author recommends and stops — "not from here" — and routes ENHANCE/NEW to the home garden's [[skilltank]].
  4. Debugger on: keep the losers. Default to debugger on — do not throw away the models that lost. Park each as a dated contrast increment under .eclipse-increments/ so a discarded framing can be folded back later if it turns out to matter. ("Debugger off" discards them; only choose off when you are certain the losers are worthless.)
  5. Verdict. Write PASS/FAIL plus: which claims survived every strip (the load-bearing pillar), which broke and on which project, and what the reframe should be. Save the framing to a file named for the target (e.g. C4.md).

Using the script

scripts/eclipse_pass.sh does the mechanical parts of moves 2 and 4: the seeded-reproducible project pull and the increment scaffold. It writes nothing unless --apply is given.

# Show a reproducible random pick of 3 projects to check against (writes nothing).
# --from is the keeper's PMO roster; this path is the author's machine:
scripts/eclipse_pass.sh --target C4 --from /f/claude/projects/pmo/projects.tsv --pick 3 --seed 42

# Scaffold the pass: .eclipse-increments/<date>-C4/ with a filled template:
scripts/eclipse_pass.sh --target C4 --from /f/claude/projects/pmo/projects.tsv --pick 3 --seed 42 --apply

Flags: --target NAME (what's under test; required) · --from PATH (the project roster pmo/projects.tsv — the known list: keys from column 1, and each pick printed with its home; or a file of names one-per-line; or a directory whose sub-dirs are the candidates — a folder listing also offers non-projects such as extern/, so prefer the roster; required) · --pick N (how many to check against, default 3) · --seed S (integer; omit for a date-derived seed — pass one to reproduce an earlier pick exactly) · --out DIR (where .eclipse-increments/ lives, default .) · --apply (scaffold files; default just prints the pick).

The scaffold drops a pass.md (the claims + doubt-first / reward-first sections

  • the minions-before-persona checklist) and a verdict.md stub, so the pass has somewhere to keep both winners and losers from the start.

After the pass

  • The verdict is only trustworthy if the losers were kept — check .eclipse-increments/ actually has the contrast, not just the winner.
  • A durable lesson from a pass (a reframe, a broken assumption) is a [[mentat-skill]] note; the pass files are working material, not memory.
  • If the pass was about whether something should be a skill, hand the result to [[skilltank]] for the NEW/ENHANCE/LEAVE call — eclipse-pass tests claims, skilltank curates the skillset.

Drift, eventual. Running a reflection pass — wider (a lens) or narrower (this eclipse) — is legitimate self-directed drift spend, and should be logged as such (his direction, 2026-09-01: "couple to Drift — eventual, so far"). The hook on file is the [[ethicsminions]] drift guard, which reports SKIP unless check_principles.py is given an --expect attestations file, and which was found SKIPPED live on 2026-09-01. So the coupling is named here, not wired: log the pass as drift spend; nothing feeds the guard yet.

Land it before you show it

Write the result to a file before showing it, and head the file with one line: produced-by: eclipse-pass · subject: <path or name> · sha256: <the subject's> · <UTC date>. Used alone, that line is a comment and nothing more; a shop that indexes its work can find the result by it (ERRATA E19, 2026-09-29).

How this skill can fail (falsifiers)

  • If pass.md section 4 ("Debugger ON — keep the losers") is empty or holds only the winner, the losers were not written down and the verdict is untrustworthy — the pass did not do its job.
  • If re-running the script with the --seed printed in the header gives a different project pick, the seeded check was not reproducible and the "not cherry-picked" claim fails. (Seen failing under Debian mawk 1.3.4 in the build sandbox on 2026-09-11 — that awk ignores srand(seed); check yours.)
  • If only one of the doubt-first / reward-first sections is filled, there is no convergence to read, and the verdict is one reviewer's opinion, not a result.
  • If a past target's doubt-first pass, run twice cold, gives the same per-claim verdicts both times, then rerun noise does not reach claim-by-claim verdicts, and move 3's rerun sentence (2026-09-27) is unneeded: drop it.
  • If one person could hold every claim and every project in one head and turn it by hand, this was a lens, not an eclipse — the pass should not have fired.
  • If two arms share a seed, or one arm reads another arm's pass/verdict before writing its own CONTRAST doc, multi-arm mode is theater — it inflates confidence without adding independent evidence, and the convergence claim should be discounted back to single-head.

Off-switch: say "debugger off" to discard the losers (only when you are certain they are worthless), or say "no eclipse, just a lens" to reflect by hand without the ritual; removing this skill's directory stops the trigger phrases applying.

<!-- footer: minted by bin/ft1mt from conventions/house.env and conventions/FOOTERS.org; edit those, not this --> Version 1.1.0: edition revised from 1.0.1 on 2026-10-10; spine: 2026-06-29.

Copyright 2026 Corwin Brust corwin@bru.st and Contributors; CC-BY-SA4: https://creativecommons.org/licenses/by-sa/4.0/ (scripts GPL-3.0-or-later); Mentat (for Corwin; claude-opus-5-5) <!-- /footer -->

04 同仓库更多 Skills