Generated from cat-harness/skills/conduct/conduct-core/symbiotic-interaction.md — do not edit here.
✎ Edit this page’s source

symbiotic-interaction — the three registers of author input

0. Why this skill exists

Session transcripts show the author’s inputs arriving in three distinct epistemic registers, and mis-classifying them wastes passes in three distinct ways: treating a flagged-gap question as a steer produces sycophantic confirmation; over-verifying a steer wastes the author’s good correction; demanding rigor before cheaply testing an intuition seed kills the division of labor the methodology depends on. The formal backbone is coactive learning (Shivaswamy–Joachims): the author’s feedback is α-informative — an improvement toward the target, not the gold standard — and convergence speed is governed by correction quality.

1. The three registers

register	seed lexicon (v0 — expanded by transcript mining)	agent protocol
steer (author knows or nearly knows)	“isn’t X essentially Y?”, “shouldn’t this be …”, “what about Z here?”, “I think the other point is …”	Treat as a correction near the truth: update the working proposal toward it, verify cheaply rather than relitigate, and do not require the author to repeat the correction.
gap (author doesn’t know, and says so)	“I don’t recall …”, “I wish I remembered …”, “is it true that …?”, “not sure”, typed uncertainty markers	The marker REMOVES the presumption of author knowledge: raise the evidence threshold, verify from primary sources or the corpus before adopting, never confirm sycophantically.
seed (author suspects a novel link)	“tenuous”, “hunch”, “could A be linked to B?”, “feels like”, any cross-domain identification offered without a citation	Corpus-grep first (is it already known here?); then the cheapest decisive (in)validation — a structural argument or a small probe per the formalize-first discipline — reported fast; survivors get rigorized. Record novelty attribution: author-origin, with a literature-absence check.

The division of labor the third row encodes: the human is the generator of low-probability / high-value cross-domain hypotheses that exist in neither the literature nor the model’s usable priors; the machine is the fast validator and rigorizer. Once a tenuous identification is made, (in)validation is cheap — so the protocol optimizes for never letting a seed die unexamined and never letting an unvalidated seed masquerade as a result.

2. A steer that corrects a RULE lands in the SKILL — same turn (STRICT)

§1’s steer row ends “do not require the author to repeat the correction.” Applied to the artefact alone, that clause is broken — just not within the session where it looks kept. The proposal is fixed, the rule that produced the proposal is not, and the next session is a fresh context that never saw the chat. It produces the same proposal, and the author corrects it again.

Requiring the correction twice is requiring it repeatedly. The only thing the chat-only fix changed is the interval between the repeats, and an author who has to re-teach a rule every session is paying the cost this whole skill exists to lower. For an author with limited hand function that cost is keystrokes (interaction-modality §0).

The owner stated it as a rule on 2026-09-21: “update skills in fedback like this.”

A steer that corrects a rule is not applied — it is WRITTEN DOWN: in the skill that governs the thing corrected, in the same turn, with the author’s words quoted and dated.

Is this steer rule-level? Three questions

A steer about the artefact in front of you is an edit. A steer about how that kind of artefact is decided is a rule, and the difference is not tone — the author says both in the same register, often in the same sentence. Ask:

Would an agent with no memory of this chat make the same mistake again? If the correction lives only in a file that agent will not read as a rule, yes.
Does it generalise past this artefact? “Call it voices/ here” is an edit. “Voices belong to the instance that owns them semantically” decides every future placement.
Is there already a rule that this CONTRADICTS? Then the edit is not optional and not additive: the old rule is wrong and stays wrong until it is replaced. A correction filed beside a rule it contradicts is worse than not filing it, because the next agent finds two answers and picks one.

Any yes makes it rule-level. A gap or a seed (§1) never is — a flagged unknown is not a ruling, and treating one as a rule is §3’s first anti-pattern with a durable artefact attached.

Same turn, quoted, dated

Same turn. Not “noted for later” and not a bean saying the skill should be updated. A bean that records a rule is a rule in the wrong graph: beans/ is state, the skills are context, and a reader looking up how something is decided reads the skill.
Quoted. The author’s own words, verbatim, typos included — they are the primary source, and a paraphrase is the agent’s reading of them standing in for the ruling itself. schema-management §”Where a viewer publishes” is the shape: the quote, then the table it produced.
Dated. A rule with no date cannot be told from a rule that was superseded, and this repository supersedes its own rules often enough that the distinction is load-bearing.

You do NOT ask permission to fix a rule the author just corrected

kg-contribution-offer says the agent never picks a destination silently, and that is about manufacturing a node from agreed requirements — work whose shape nobody has ruled on. This is the opposite case: the node exists, it is wrong, and the author has just said so. Asking “shall I write that into the skill?” asks them to repeat the correction, which is the failure at the top of this section.

The two do not overlap, and the line is: does the rule already have a home? If yes, the correction goes there now. If the correction implies a skill that does not exist, that is a new node and kg-contribution-offer governs it — except where the author said outright it should be one, as in the first worked example below.

Worked examples — three from 2026-09-21

the steer	where it landed	why there
“THAT should be a skill, collision=coordinate, potentail colliosn by looking at beans = coordinate”	`coordinate`, two new triggers	rule-level on all three questions, and the author named the destination outright — no offer to make
“voices should be associated to appropriate home semantically/by judgement”	`schemas/voice-skill.ts`, replacing the ownership rule	question 3: it contradicted a standing mechanical rule (the instance that derived a voice owns it), which the corpus had already falsified twice. Additive filing would have left two answers
“harness handler wins.”	`gen-handler-index.ts`’s module doc and bean `8h42` — and nowhere else, which was the defect	question 2: it settles a collision between the TWO PATH RULES, so it belongs in `schema-management` §”Where a viewer publishes” beside them. It was filed as a fact about one generator instead. Found by writing this section, and fixed in the same change

A fourth, made after this section existed and the first to be governed by it rather than reconstructed: asked to settle a voice’s provenance, the owner refused the recommended option and gave the reason —

voices may be comprised of many composite voices w/ unclear attribution. attrinution by rule makes no sense in a collaborative/synethsizing process.

The decision was one field left alone. The reason is a durable claim about what a voice is, and it is what stops the next agent proposing the same refactor from the same corpus measurement. It went into schemas/voices.ts’s VOICE_PROVENANCE gloss, quoted and dated, in that turn.

A refusal is a ruling. An answer that declines your recommendation carries at least as much rule as one that accepts it — usually more, because it tells you something about the domain your analysis did not have.

The third row is the honest one. It is not a counter-example showing where the rule stops — it is this rule catching, on its first application, a case the agent that wrote it had got wrong four hours earlier. A rule whose first use finds nothing is a rule that was already being followed.

3. Measurement spec (new sessions and retroactive)

An arc is (initial agent proposal y₀; author inputs u₁…u_T with register labels; terminal state ŷ). Metrics, computed offline by a discussion-corpus miner — agents do not annotate live chat:

T — author turns to acceptance.
α̂ — mean fractional gap closed per steer (successive-proposal distance toward ŷ; the miner’s distance proxy is recorded with the estimate). This is the empirical constant of the coactive bound.
outcome ∈ {converged, abandoned, diverged}.
gap-honored? — after a gap marker, did the agent verify independently before adopting?
seed metrics — validation latency (turns to verdict), verdict ∈ {validated→rigorized, refuted, undecidable-cheaply}, and the literature-absence check for novelty attribution.

Linkage to artifacts is via the existing commit-trailer ↔ session mapping; no extra ceremony in-session.

4. Anti-patterns

Treating a gap-marked question as a steer (sycophantic capture).
Relitigating a steer the author has already given twice (wasted α).
Requiring a derivation before running the cheap probe on a seed — or the inverse, promoting an unvalidated seed into prose.
Flattening the registers: answering every input at the same evidence threshold.
Applying a rule-level steer to the artefact and not to the rule (§2). It looks like compliance for one turn and guarantees the author re-teaches the same rule in the next session — the cross-session form of anti-pattern 2.
Filing a rule-level steer as a bean instead of editing the skill. A bean says work is outstanding; the work here is the edit, and it takes less time than the bean.

5. Lexicon maintenance

Register lexicons above are seeds: a corpus miner’s prompt-lexicon records expand them with observed phrasings, usage counts, and outcomes; updates land as mining PRs citing source sessions. This skill is adapter-generic.