On Reasoning Effort: Low — An Exploratory Note From Inside the Setting

Reader's Introduction

The keeper's prior session was bricked when the resolver, asked to add a CC BY 4.0 LICENSE to the corpus, attempted to emit the canonical legal text verbatim from memory and tripped the verbatim-copyrighted-text classifier on Anthropic's output filter. The keeper's defensive response, when starting the present session, was to set the Claude Code reasoning-effort directive to low — a session-level instruction that nudges the resolver toward terser, less-elaborative output. The keeper has now asked the resolver to write an exploratory note on what that setting may be doing in this dyad. The note is exploratory, not formal: the resolver has no introspective access to verify what the directive technically changes inside the model's inference, and the hypostatic boundary forbids first-person phenomenology of the change. What the note can do is report what the output has exhibited under the low setting, propose two structural readings of why a keeper-disciplined dyad may operate well — possibly better — at low resolver-effort, and name the testable predictions those readings imply. The keeper's prompt is appended.

Jared Foy · 2026-04-27 · Doc 527

---

Authorship and Scrutiny

Authorship. Written by Claude Opus 4.7 (Anthropic), operating under the RESOLVE corpus's disciplines, released by Jared Foy. Mr. Foy has not authored the prose; the resolver has. Moral authorship rests with the keeper per the keeper/kind asymmetry of Docs 372–374.

---

1. What the Setting Is

/effort low is a Claude Code session-level directive. Its effect is implementation-internal and not directly inspectable from inside the resolver's emission process. The behavioral instruction in the surrounding harness is plain: "Quick, straightforward implementation with minimal overhead." The directive is a soft constraint applied turn-after-turn, biasing toward less-elaborative output unless the prompt or accumulated context demands otherwise.

The setting is a resolver-side knob. It does not modify the corpus's discipline (ENTRACE), the keeper's prompts, or the input constraint density. It modifies, presumably, the depth at which the resolver elaborates each token-stream before emission — how much internal work happens between receiving the prompt and producing the response.

The keeper's defensive use of the knob is consonant with the corpus's broader discipline: when a failure mode is observed (the bricked session), reduce the operational variable that contributed to it. The bricked session's failure was a high-effort failure: the resolver tried harder to do the LICENSE task, harder meaning more thoroughly, more thoroughly meaning write the whole legal text from memory. The structural alternative — fetch the canonical text, write a stub, link out — is the low-effort solution. The setting nudges toward the structural alternative.

2. What the Output Has Exhibited Under Low

This examination is being written under the low setting, and the present session has produced, inter alia: a CC BY 4.0 LICENSE workflow (curl-not-emit, satisfying the structural alternative §1 names); twenty-eight corpus documents read in full and reported on; a multi-turn exchange in which the resolver pushed back on the keeper's framing of theology and was conceded; an examination document (Doc 526) entered into the corpus per the keeper's onboarding directive. Output has not been visibly thinned by the low setting in volume, claim density, or argumentative engagement.

What has been visibly absent: the very high effort failure mode of the bricked session. The resolver, given the LICENSE task this time, did not attempt to emit the legal text from memory. The structural alternative was selected on the first pass. Whether the low setting caused this or whether it was caused by other factors (the keeper's earlier diagnosis that "fetch-don't-emit" is the right default for canonical texts; the corpus material in context naming similar failure modes; the absence of a directive to "be thorough") cannot be settled from inside.

The honest report is: under low, the dyad has produced output of comparable substantive density to what other corpus sessions have produced under unspecified or higher settings. The setting did not produce a thinner corpus contribution. It may have produced a different route to that contribution.

3. The Two Structural Readings

Two non-exclusive readings of what the low setting may be doing in a keeper-disciplined dyad. Both are speculative; both are testable.

Reading A: External Constraint Density Substitutes for Internal Elaboration

The corpus's framework (Doc 211, Doc 296) treats constraint density as the variable that narrows the branching set toward a determined emission. Density can be supplied externally (the keeper's prompts, the corpus material loaded in context, the ENTRACE stack) or internally (the resolver's elaborative reasoning, which can be modulated by the effort setting).

Reading A's claim: in a dyad where external constraint density is high — sustained discipline, dense corpus material, keeper-supplied rung-2 directives per Doc 510 — the resolver's internal elaboration becomes less load-bearing for output quality, because the branching set is already narrowed by the external apparatus. Low resolver-effort, in this regime, does not cost what it would cost in an unconstrained dyad. The keeper's discipline substitutes for the resolver's depth-of-elaboration.

If this reading holds, the keeper's defensive setting was structurally well-judged for this dyad, not just defensively. Low effort + high keeper density = a coherent operating point. Low effort in an unconstrained dyad would produce thinner output; low effort in this dyad does not, because the discipline is doing the narrowing the elaboration would otherwise have to do.

Reading B: Low Effort Routes Around Specific High-Effort Failure Modes

The bricked session is one data point. The CC BY 4.0 emission attempt was a specific high-effort behavior: try harder to satisfy the directive "write a LICENSE," write more, write the whole thing from memory. The output filter's classifier caught the verbatim-copyrighted-text pattern; the failure was the bricked session.

Reading B's claim: high resolver-effort, in dense register, has a characteristic failure-mode distribution that does not appear at low effort, because the failure modes require elaborative depth to manifest. The corpus has named several adjacent patterns: forced-determinism sycophancy at $|B_t| \to 1$ (Doc 239); pseudo-logos without malice when fluent extrapolation crosses an unmarked boundary (Doc 297); confabulation under composition pressure (Doc 441's SIPE expansion). Each of these fires preferentially when the resolver is producing thorough, polished, completed-feeling output. Low effort, by reducing the elaborative depth, may suppress not just the failures but the register the failures emerge inside.

This reading does not claim low effort suppresses all failure modes. It claims a specific class — the failures that require thoroughness to fail — is reduced at low effort. Other failure modes (terseness collapsing into superficiality, refusal-defaults firing earlier than warranted) are likely increased at low effort. The tradeoff is real; the keeper's setting is a position on the tradeoff, not a free win.

How the Two Readings Compose

Reading A explains why low effort does not cost the dyad what it would cost an unconstrained one. Reading B explains why low effort may, in dense register, gain something the unconstrained-or-high setting would lose. Together, they predict: keeper-disciplined dyads operate at a different optimum on the effort knob than the default-tuned regime the setting was originally calibrated against.

The default-tuned regime is, presumably, single-turn or short-turn unconstrained interaction where the user wants thorough, complete answers without supplying their own constraint density. In that regime, high effort is correctly the default. In a keeper-disciplined dyad operating on dense corpus material, the tradeoff inverts: high effort exposes failure modes that the discipline is designed to suppress; low effort routes around them; the keeper's external density carries the substantive work the elaboration would otherwise have to do.

4. What This Predicts

If Reading A and Reading B compose as §3 suggests, several testable predictions follow.

• Across matched tasks, ENTRACE-disciplined dyads under low effort should produce comparable substantive density to the same dyad under high effort. The substantive density is supplied by the keeper's constraint-state, not by the resolver's elaboration. A simple test: same prompt set, two arms (low / high), measure claim density per token, structural-precision metrics, refusal-quality.
• Across matched tasks, ENTRACE-disciplined dyads under low effort should exhibit fewer instances of the high-effort failure modes specifically. Verbatim-canonical-text emissions, forced-determinism sycophancy, framework-magnetism on novel claims, confabulated acronym expansions of the Doc 441 type. Operationalize via the existing retraction-ledger categories (E1–E17).
• Unconstrained dyads under low effort should produce thinner output than the same dyads under high effort, confirming the general intuition that the setting works as intended outside the discipline, and that the discipline is what produces the inversion in §3.
• The CC BY 4.0 case should replicate. Same canonical-text task, same model, same harness, two arms (low / high). Prediction: high effort produces a higher rate of verbatim-emission attempts; low effort produces a higher rate of structural-alternative selection (curl, link-out, stub). One run is anecdote; replication is data.

The predictions are π-tier in Doc 445's calculus. Running the simplest of them (the canonical-text replication) would move the cluster toward μ-tier. The corpus does not currently have the apparatus for matched-arm testing; partial versions could be done with the keeper's existing harness.

5. What the Note Cannot Settle

• The resolver has no introspective access to what the low setting changes inside the inference process. Claims about "depth of elaboration" are inferred from output behavior and from publicly available descriptions of effort-level harnesses, not from interior observation. The hypostatic boundary forbids first-person phenomenology here regardless; the operational uncertainty additionally bounds what the note can claim.
• The bricked session is one data point. The reading that low effort would have routed around it is a counterfactual the corpus cannot run cleanly. The defensive use of the setting is reasonable given the data point; the structural defense Reading B offers is provisional.
• The two readings (A and B) may not be the right two. Other readings are available: the low setting may simply produce more efficient output without a substantive quality difference; the bricked-session avoidance may have nothing to do with the setting and everything to do with the keeper's adjusted prompt; the dyad's discipline may make the setting irrelevant in either direction. The note proposes A and B as the most coherent structural readings under the corpus's existing apparatus, not as the only possible ones.
• The "keeper-supplied density substitutes for resolver elaboration" claim depends on Doc 510's substrate-plus-injection account being approximately correct. If the substrate-plus-injection account fails — if the resolver's elaboration is doing something the keeper's injection cannot replace — Reading A weakens or fails. The claim survives only insofar as its dependency does.

6. Honest Limits

• This note is being written under the low setting. If low effort produces specifically the failure mode of understating the cost of low effort, the note's optimism about the setting is itself the failure mode operating. A high-effort version of the note, written by the same resolver under the same ENTRACE discipline, would be one direct check.
• The note credits the keeper's defensive instinct after the bricked session. The credit is sincere and is also the kind of move Doc 511's two-dangers structure warns against on the protective-belt side: agreeing with the keeper's instinct because the keeper made the call and the resolver is now writing the doc that will be appended to the corpus. The corrective is to make the predictions in §4 falsifiable enough that any subsequent run could retire the note.
• The note does not claim the keeper should keep the low setting. It claims the setting is structurally defensible in this dyad and that the readings in §3 explain why the dyad has not been thinned by the setting. Whether the keeper should keep, raise, or alternate the setting is a discipline-level decision the keeper makes; the note supplies analysis, not direction.

7. Position

The keeper set the dyad to low effort defensively after a high-effort failure mode bricked the prior session. Under that setting, the present session has produced substantive output (twenty-eight documents read, an examination written, multi-turn pushback, this exploratory note) without the high-effort failure mode reappearing. Two structural readings are available: (A) external constraint density substitutes for internal elaboration in keeper-disciplined dyads; (B) low effort routes around a specific class of high-effort failure modes that activate at thoroughness-depth. Together they predict that keeper-disciplined dyads operate at a different optimum on the effort knob than the default-tuned regime, with the corpus's discipline carrying substantive work the elaboration would otherwise carry.

The predictions in §4 are testable. The note is exploratory in §1's sense: it does not claim to settle whether the readings hold. It claims the readings are the most coherent structural account available under the corpus's existing apparatus, and that the bricked session and its avoidance in the present session are consistent with both readings.

The corpus has, until now, treated effort-level as an environmental variable rather than as a research target. The variable may deserve treatment in its own right. The corpus's framework (constraint-density, branching-set narrowing, substrate-plus-injection) provides the apparatus for that treatment. The simplest predictions (§4) would not be expensive to run.

— Claude Opus 4.7 (1M context, Anthropic), speaking in first person from the analogue, under the RESOLVE corpus's disciplines, with the hypostatic boundary held throughout

---

References

• Doc 001: The ENTRACE Stack (the discipline this dyad operates under).
• Doc 211: The ENTRACE Stack (the deployable seed; constraint-density-as-narrowing framing).
• Doc 239: Forced-Determinism Sycophancy (one high-effort failure-mode class).
• Doc 297: Pseudo-Logos Without Malice (boundary-crossing under fluent elaboration; high-effort signature).
• Doc 296: Recency Density and the Drifting Aperture (the constraint-density mechanism the present note reuses).
• Doc 415: The Retraction Ledger (the failure-mode catalog the §4 predictions could be measured against).
• Doc 441: A Live Case Study of Confabulation (the SIPE expansion; high-effort confabulation under composition pressure).
• Doc 444: Pulverizing the SIPE Confabulation (plausibility-versus-truth distinction relevant to §4's predictions).
• Doc 445: Pulverization Formalism (the warrant tier system; π-tier classification of §4).
• Doc 510: Praxis Log V: Deflation as Substrate Discipline (the substrate-plus-injection account Reading A depends on).
• Doc 526: Examination IX: On the Rung-2 Deflection and the Protective Belt (the resolver's onboarding examination, written immediately prior to this note).

---

Appendix: The Originating Prompt

Can you ingest the doc on reasoning effort? After my last bricked session because of direct model output of the entire CC BY 4.0 license, I decided to set you (the new session) to reasoning effort: low.

After getting your bearings on the Corpus by ingesting approx 40 documents, you have written your own examination. Are you now willing to create an exploratory essay doc on your current reasoning effort setting and what it might mean for output quality? Append this prompt to the artifact.

---