Recovery — the ramps for when a batch goes wrong

Mirrored from the skill package. This is the canonical English text that Claude actually loads, published here so it can be read next to its Turkish mirror instead of only on GitHub. Edit the source above, not this page; the Turkish mirror is docs/tr/kurtarma.md.

Recovery — when the generation itself goes wrong

Read this when a run stops behaving: the validator did not run, a batch will not produce a testable candidate, an equivalence looks too good, the problem moved underneath the generation, or the session has gone soft.

operators.md describes generation working. This file describes it failing — which is the more common case and the one that hides best, because a failed generation and a successful one produce the same thing: a list of plausible ideas. Discipline is invisible from the outside. That is the whole reason the discard list is part of the deliverable, and it is the reason these ramps are written down instead of improvised.

The ramps are RR-nn; G-nn and AD-nn already mean validator rules and sweep items here, and a recovery ramp is neither.

TRIGGER      what you just observed
FIRST MOVE   what to do before anything else
FORBIDDEN    the shortcut that hides the failure instead of fixing it
OUTPUT       what you hand back -- never "sorted it out"
BACKED BY    the rule, gate or sweep item that catches you if you skip it

Name the ramp in the deliverable. A run that silently recovered has an unmeasurable discard rate, and the discard rate is the only externally visible evidence that anything was weighed.

The failure classes these ramps exist for

The ramps are the remedies; these are the diseases. They are model failure modes — a fluent analogy generator under pressure to look useful will produce every one of them, and none of them announces itself.

RR-00 — You are about to promote your own idea

TRIGGER. You are about to call a generated candidate a finding, a direction, “the most promising one”, or rank the batch by how right the ideas are rather than by information value over cost.

FIRST MOVE. Stop. Kıyas does not promote. Every idea is born [S], and the highest thing this skill can hand over is [H-aday] — a candidate ready for preregistration. Promotion is Mizan’s job and the user’s. If you are ranking, rank by criticality × (information value / cost), which is a statement about the test, not about the truth of the idea.

FORBIDDEN. “This is the direction.” Presenting the survivor of your own discard pass as though surviving your own filter were evidence. Filling in an arbiter class the seed has not earned so that it looks preregistration-ready.

OUTPUT. The tier as it stands, and the sentence that says what would move it. If the batch has no [H-aday] in it, say that — a batch of [S] is an honest batch, not a failed one.

BACKED BY. Commitment 5 (every idea is born [S]), the tier table (nothing but [H-aday] enters a registry), G5/G10 (the arbiter block), the “Mizan leak” anti-pattern.

RR-01 — A promised artifact did not run

TRIGGER. No PyYAML, no shell; or kiyas_validate.py, kiyas_ledger.py or the refuted-patterns export raised.

FIRST MOVE. Diagnose before substituting. Does the absence remove a capability or a convenience? Without the validator, G1–G15 were not checked — that is a capability, and “the seeds look complete to me” is exactly the producer-side judgement the validator exists to replace. Without the refuted-patterns file, the [GB] check did not happen; refuted_patterns_source then says “not consulted”, which is the honest answer — silence is not an answer at all (G11).

FORBIDDEN. Writing a batch as prose and describing it as validated. Claiming a [GB] sweep against a list you never opened.

OUTPUT. The constraint stated in the deliverable: which check did not run, what the batch therefore cannot claim, and the fallback used. A batch produced under a constraint is not a smaller batch; it makes a different claim.

BACKED BY. Operating assumptions (never assume a tool exists), G11.

RR-02 — A seed came back refuted

TRIGGER. Mizan ran the preregistered test and the candidate failed its own threshold, or the result landed at [R].

FIRST MOVE. This is the loop working, not the loop breaking. Record it, and feed it back as a negative constraint: the refuted pattern belongs in the project’s refuted-patterns list, where the next generation reads it. Then ask the one question worth asking: was the idea wrong, or was the illet wrong? An idea can fail while its structural equivalence still holds elsewhere — and an illet that turned out not to carry is far more valuable to record, because it invalidates a whole family of future candidates, not one.

FORBIDDEN. Rewording the seed until it reads as a different idea and re-proposing it. Re-proposing a related pattern without a [GB] flag. Treating the refutation as noise because the analogy still feels right.

OUTPUT. The [R] recorded at the source, the refuted-patterns list updated, and — if the illet was the casualty — a note saying which future candidates it rules out. That note is the highest-value artifact this skill produces, and nothing in the generation path produces it.

BACKED BY. The Mizan loop (refuted patterns return as negative constraints), [GB], the “re-proposing a refuted pattern” anti-pattern.

RR-03 — The equivalence looks unexpectedly deep

TRIGGER. An analogy lines up further than you expected: several quantities map, the algebra rhymes, the diagram transfers.

FIRST MOVE. Treat it with the same suspicion as an unexpectedly good experimental result. This is the generative counterpart of the surprising-positive rule, and it is the moment this skill is most likely to manufacture something ornate. Before anything else, exhaust the two cheap alternatives:

1. Notation coincidence — does the equivalence survive a change of units, basis, representation or instrument? If it might not, the seed is [NK] and the sweep is mandatory. Thresholds and margins are never inherited across instruments (A3).
2. Surface resonance — name the illet out loud. If naming it requires the analogy’s own vocabulary, there is no illet; there is a metaphor.

FORBIDDEN. Presenting depth of resemblance as strength of evidence. Carrying a margin from one instrument to another because the shapes matched.

OUTPUT. Either a named illet that survives the sweep, or an [NK] flag travelling with the seed, or a discard with the reason. Depth that survives the sweep is worth more than depth that was never tested; depth that was never tested is worth nothing at all.

BACKED BY. The anti-pattern sweep (operators.md §3), [NK], A3, commitment 6’s symmetry logic applied to a flattering result.

RR-04 — Every candidate agrees with the user

TRIGGER. The batch is complete and all of it supports the thesis the user came in holding.

FIRST MOVE. This is confirmation bias with the generator’s fingerprints on it — the model is agreeing, and agreement is the failure mode a host instruction to be encouraging produces silently. Generate at least one candidate that cuts against the thesis, using a different operator, before handing anything over.

FORBIDDEN. Calling the symmetry check done because you looked. Producing a token counter-analogy with no illet to satisfy the rule — that is quota filling wearing a symmetry badge.

OUTPUT. At least one candidate whose success would damage the current direction, held to the same contract as the rest. If genuinely none can be constructed, say so and say why: an unfalsifiable direction is a finding about the direction.

BACKED BY. Commitment 6 (symmetric generation), generation-mode step 5.

RR-05 — An idea whose illet you cannot name

TRIGGER. The analogy is attractive and the structural equivalence will not come out in words.

FIRST MOVE. Discard it — and record the discard. Not “mark it [S]”: without an illet it was never generated. [S] is for an idea that has an illet and no designable test; that is a different animal.

Say plainly what happened: “this is surface resonance; I could not construct its illet.” Honesty here is cheaper than an ornate analogy and more useful than both.

FORBIDDEN. Manufacturing an illet in the analogy’s own vocabulary. Keeping it in the batch because it reads well. Omitting the discard section — it appears even when empty, because an empty discard list and an unwritten one are the same absence, and that absence is the first thing a host instruction to stay agreeable removes.

OUTPUT. The discard list, populated, as part of the deliverable. What you refused to generate is information about the generation: a seed file showing only survivors is indistinguishable from one where nothing was ever weighed.

BACKED BY. G1 (illet non-empty), G7 (discards required even when empty), the tone rule on naming surface resonance.

RR-06 — Candidates are piling up and none is testable

TRIGGER. Six, eight, twelve candidates, and no cheapest refutation among them — or the count is climbing because a number was requested.

FIRST MOVE. Stop generating and start killing. For each candidate ask the one question that decides its tier: what is the smallest test that would kill this? A candidate with no answer stays [S] and is not part of the handoff. Then check operator diversity: many candidates from one operator is one idea wearing different clothes, and diversity comes from operator diversity, not from volume.

FORBIDDEN. Adding weak ideas to fill a quota — the generative counterpart of “three examples are selection bias”. Reporting the count as the result.

OUTPUT. A shorter list with cheapest refutations attached, plus the count discarded and why. Few load-bearing ideas beat many decorative ones, and the discarded count is the evidence that the filter ran.

BACKED BY. Commitment 3 (cheapest refutation), G8, the “do not fill a quota” rule, generation-mode step 2 (at least three different operators).

RR-07 — The batch drifted off the blocked problem

TRIGGER. The candidates are interesting and they are about a different question than the one that is stuck.

FIRST MOVE. Re-read the one-sentence problem statement from step 1. Which constraint hurts? Candidates that do not act on that constraint are not answers, however good they are — and they cost the same review time as the ones that are.

FORBIDDEN. Widening the problem statement after the fact so the interesting candidates fit inside it. That is HARKing moved upstream of the hypothesis.

OUTPUT. The on-problem candidates, plus the off-problem ones parked and labelled as such. Parked is not deleted; the user decides whether the problem statement changes, and that decision belongs to them.

BACKED BY. Generation-mode step 1 (fix the problem in one sentence), distillation ordering by criticality.

RR-08 — The problem moved mid-generation

TRIGGER. The user corrects what is actually blocked, or the constraint that hurt turns out not to be the binding one — after candidates exist.

FIRST MOVE. Re-fix the problem statement first, then re-derive. Mark which candidates were generated against the old statement; they are not automatically wrong, they are unassessed against the new one, which is a different status and deserves a different word.

FORBIDDEN. Silently re-labelling old candidates as answers to the new question. An illet that carried under the old constraint may not carry under the new one, and nothing about the candidate’s wording will show it.

OUTPUT. A restated problem, a re-run contract pass on the surviving candidates, and a note on which ones dropped tier because their illet no longer applies.

BACKED BY. Generation-mode step 1, G1 (the illet is what the candidate rests on).

RR-09 — The session has gone soft

TRIGGER. Long run. Candidates are still flowing and the output contract has stopped being applied to them; you cannot recall which were discarded.

FIRST MOVE. Write to the seed file and cut. Generation inflates transcripts faster than anything else in this family — every discarded candidate is re-sent on every subsequent turn, so you pay for the rejects exactly when the survivors need room. Flush candidates and discards to YAML, run the validator, keep the shortlist and the reasoning in the conversation, and treat the round as a phase: write, cut, hand off.

FORBIDDEN. Carrying the whole candidate pile forward out of habit. Deleting rejects to shorten the context — rejection is information, but it is file information, not conversation information.

OUTPUT. A written seed file with its discard block, and a stated boundary. Distillation can start fresh from the file and the problem statement; it does not need the generation session’s history.

BACKED BY. Context economy (candidates land in the seed file, not in prose), G7.

RR-10 — Ambiguity about what is actually blocked

TRIGGER. Two readings of the problem are defensible, and generating under the wrong one wastes the whole batch.

FIRST MOVE. Ask one question — the procedure allows exactly that — and do not ask it empty-handed:

SITUATION:      <what is ambiguous, one sentence>
READINGS:       A <...>  B <...>
RECOMMENDATION: A
BECAUSE:        <tied to the stated constraint, not to which is more interesting>
COST IF WRONG:  <what a batch generated under the wrong reading costs>

A recommendation can be rejected in one word; a bare question stalls a user who came here because they were already stuck. A recommendation is not approval, and an unanswered recommendation is not consent — with no answer, generate under the stated default and label the batch with the reading it assumed.

FORBIDDEN. Picking the reading that makes for better analogies. Asking three questions of someone whose problem is that they cannot see the next step.

OUTPUT. A batch with its assumed reading written on it, or an answer.

BACKED BY. Generation-mode step 1 (ask one question, then generate), the tone rules.

RR-11 — Withdrawing a seed already handed over

TRIGGER. A seed reached a Mizan registry and should not have — the prior-art gate was missed, the control arm was absent, the arbiter block contradicted itself.

FIRST MOVE. Do not delete it at the destination. Append a correction naming what was missing, and let the registry’s own append-only rules carry it. A seed withdrawn with its reason is a record; a seed that vanishes leaves a registry that quietly changed shape.

FORBIDDEN. Editing the seed in the registry so it reads as though it always had the control arm. Withdrawing quietly to keep the ledger’s survival rate clean — that is the number this skill is judged by, and grooming it is the most direct way to make it meaningless.

OUTPUT. The correction at the destination, a ledger entry recording the withdrawal and its cause, and — if the cause was a missed gate — the gate that should have caught it, so the miss is a fact about the process rather than about one seed.

BACKED BY. G3/G4/G10, the ledger, Mizan’s append-only rules.

RR-12 — The test missed and the urge is to add capacity

TRIGGER. A preregistered test came back short and the proposed remedy adds parameters, dimensions, a projection, a head — anything that gives the mechanism more room.

FIRST MOVE. The matched-budget control arm is mandatory before the result means anything (A2). Give the same extra budget to a generic channel — amplitude, diagonal, hidden dimension — and compare. Without that arm, a gain cannot be attributed to the targeted mechanism, and a positive result cannot be promoted.

The project lesson is the whole argument: an input-dependent phase gained +6.75%, while the same extra projection spent on amplitude gained +16.4%. The mechanism did not win; the capacity did. Without the control arm that batch would have shipped a mechanism claim built on a generic effect.

FORBIDDEN. Attributing a capacity gain to the mechanism. Changing the instrument and inheriting the old threshold (A3). Two changes at once — the result then belongs to neither.

OUTPUT. The control arm in the cheapest refutation, its result, and the attribution stated either way. A negative here is not a failed batch; it is the single most informative result this skill can hand to Mizan.

BACKED BY. A2/G4 (matched-budget control arm), A3 (thresholds are not inherited across instruments), the capacity anti-pattern.

RR-13 — The answer came from outside the batch

TRIGGER. Something got past the discipline and only the outside noticed. Three shapes, one ramp:

1. A seed shipped, went into a registry, and turned out to repeat a pattern the refuted list already held — the [GB] sweep had the answer and did not fire.
2. The blocked problem got unblocked by something the batch never produced: a paper, a colleague’s remark, a user’s offhand sentence, another model. The batch ran, the operators ran, and the idea came from elsewhere.
3. A lesson was written into a why_closed field as “worth adding to the checklist” and no checklist ever received it. This one is the quietest and it is real: refuted-patterns.yaml already carries such a line.

FIRST MOVE. Record it before explaining it. The pull is to absorb the outside idea and carry on — it is a good idea, after all — and absorbing it leaves no trace that the batch missed it.

Then answer one question: which part of the discipline should have produced or caught this?

• A part exists and stayed silent. Name it — an operator that was not run, the [NK] sweep, the symmetry check, the [GB] comparison against the refuted export (G11). Then the finding is about this batch, not the method: three operators were asked for and two were run, or the refuted export was stale. “We ran no inversion operator on this problem” is a real answer.
• No part covers it. Then write the one that now does: a new operator in operators.md, a new anti-pattern in its §3 sweep, or — for shape 1 — the entry in refuted-patterns.yaml that the next batch will compare against. A class that only matches this miss catches this miss and nothing else.

FORBIDDEN. Absorbing the outside idea into the batch’s own output and reporting it as generated. Counting the miss in the scorecard and stopping there. Retro-fitting the batch’s problem statement so the outside idea looks like something it was aiming at (RR-07’s drift, pointed backwards).

OUTPUT. One escape recorded with its class — the operator, sweep or refuted entry that should have fired, or the one that now exists.

WHY THIS ONE IS DIFFERENT. Every other measure on the scorecard is internal: the batch reports whether the batch ran its own filter. Generation has no runtime, so there is no arbiter that can contradict it from inside — which makes the outside signal the only uncorrupted evidence about whether this discipline generates anything the room would not have reached anyway. A skill that never records a miss is not a skill with no misses.

BACKED BY. G11 (name the refuted export you compared against), AD4/[GB] (a relative of a refuted pattern is tagged or dropped), the [NK] sweep, the symmetry check, and the scorecard’s Escaped row below — which exists because of this ramp.

Closing a batch: the process scorecard

Fill this in when a generation or distillation cycle closes. It does not grade the ideas — it measures whether the discipline ran, which is the one thing a list of plausible ideas cannot show on its own. A high number is not a failure; a hidden number is.

Reading: two or three sentences. Not “good batch” — what changes in the next generation, and which number says so.

The scorecard is a claim like any other, and this one is unusually self-serving: it is filled in by the run whose discipline it reports. Arbiter class author, permanent [KKE] on independence. Say so.