Jump to text
On this page
Copy and source
Published source confirmed at last check

Source changed 2026-11-18 21:11:39 UTC · snapshot created 2026-11-18 21:12:35 UTC · last check 2026-11-18 21:55:20 UTC

C.19.1 - Bitter‑Lesson Preference (BLP)

Status: Stable

Purpose. State the empirical Bitter Lesson: in search, learning, planning, and related computational work, general methods able to use increasing compute or data have often displaced hand-engineered special cases. Treat that history as a comparison pressure, not as proof about every bearer. Here bearer names the candidate being compared under the scale claim or generality policy. A project may declare an analogous preference for a module, platform, organization design, evidence arrangement, or other bearer only as a separate local policy, with a scale predicate, objective vector, comparison basis, and evidence form appropriate to that bearer. Safety, cost, admissibility, uncertainty, and non-dominance remain visible; the word general creates no preference by itself. Builds on. C.19 (E and E‑LOG), C.24 (Agent‑Tools‑CAL; ATC‑2), B.3 (Assurance), E.3 (Precedence), E.5 (Guard‑Rails). Coordinates with. G.5 (Selector), G.8 (SoS‑LOG Bundles), G.9 (Parity), G.11 (Refresh‑Telemetry), A.0 (On‑Ramp). Keywords. general-solution preference; scale‑amenability; BLP‑waiver; iso‑scale parity; Scale‑Audit; slope vector; alpha and delta tolerances.

C.19.1:0 - Use This When

Use C.19.1 when a current choice or policy makes a scale claim or invokes a declared generality preference: a narrower special-purpose approach is preferred over a general alternative, or a general approach is preferred because its measured performance is expected to improve across a declared scale window. For search, learning, planning, and agent substrates, the empirical Bitter Lesson can supply the motivating line. For a module relation, platform, organization design, evidence-bearing episteme or work arrangement, or selected structure, state that the move is a local analogy or policy rather than an empirical Bitter-Lesson result.

The pattern governs only that scale-based comparison, preference, or waiver. It neither proves architecture adequacy nor turns a bearer label into a holon kind. If the project is using a bounded specialization and makes no scale advantage or durable generality claim, keep the use local under the bearer’s direct pattern and stop here.

When comparing a general adaptive loop with a specialized cycle or direct repair for an E.23 use, use C.19.1 only if the decision relies on scale advantage or a declared generality policy. The E.23 loop still names the object under improvement, evaluation, cost and risk account, protected trade-offs, and stop or switch condition.

C.19.1:0.1 - What Goes Wrong If Missed

A team treats “more agentic”, “more automated”, “more specialized”, or “works on this benchmark” as proof that one bearer should displace a more general scale-amenable bearer. Another team repeats the opposite error: it invokes the Bitter Lesson as permission to ignore safety, cost, task-family fit, or a narrow heuristic that wins inside the declared scale window. In both cases, the selector loses parity, waiver, and scale-window discipline.

C.19.1:0.2 - What This Buys

The practitioner gets a cheap first probe before an expensive audit. It distinguishes a supported scale comparison, a declared local analogy or policy, a bounded use with no scale claim yet, and a high-stakes claim that justifies a fuller Scale-Audit. When comparison proceeds, task family, scale window, parity, uncertainty, cost, safety, and waiver remain explicit.

C.19.1:0.3 - Not This Pattern When

Apply the pattern that defines and tests the current question: C.30 or C.32 for architecture adequacy and synthesis, G.5 for selected-set result declaration, E.17 for a source-backed publication face and return to source, E.24.PUB for the publication occurrence and audience availability, E.23 for repeated object-version improvement, the A.15 family for work, and A.21 for gate decisions.

C.19.1:0.4 - First Output

First examine the scale claim and its available basis before selecting a Scale-Audit. This cheap scale-claim probe may use an existing adequate account; it does not by itself require a new experiment or note. Identify the two candidates and their direct patterns, task family or receiving use, proposed scale predicate, objective vector, comparison basis, feasible evidence, safety boundary and stakes. Return one of four results:

  • no scale claim yet: no supported scale claim is available. State whether no such claim is needed or which claimed basis remains unresolved. A bounded use may proceed under its direct pattern only if it needs no missing scale conclusion or override of a declared preference; this result supplies neither preference nor waiver;
  • local analogy or policy: for an analogy, identify the bearer family, the proposed transfer and its grounds. For a local selection policy, identify the actual candidates, the policy edition and the condition under which it decides their use. This can include a policy governing two computational methods.
  • bounded scale comparison: state the smallest parity and uncertainty method adequate for this use;
  • full Scale-Audit selected: state why the claim, stakes, feasible evidence, and receiving use justify the added work.

A BLP-waiver is needed only when a declared generality preference would otherwise decide the use.

C.19.1:1 - Problem frame

Bespoke computational heuristics can win locally while failing to exploit larger compute, data, or search budgets. General methods can improve across a declared scale window, but the empirical record is strongest for computational search, learning, and planning. A module organization, institution, work arrangement, or episteme does not inherit that empirical result by analogy. A project may adopt a broader policy, but it must state the bearer-specific scale relation and evidence rather than treating all growth, reuse, or generality as one phenomenon.

Without this separation, teams either repeat the Bitter Lesson as a slogan or impose a costly machine-learning experiment recipe on bearers for which seeds, FLOPs, compute slopes, or data sweeps have no meaning.

C.19.1:2 - Policy clauses (normative)

BLP-1 — Probe first; select audit depth by claim and risk. Every BLP use starts with the scale-claim examination above. A claim ready for comparison needs the bearer kind and direct pattern, a recoverable scale predicate, objective vector, comparison basis, evidence form and named receiving use. If any is absent, identify the missing basis and return no scale claim yet: the asserted claim remains unsupported, rather than disappearing. A useful local analogy or policy may be retained as such, but it supplies no empirical conclusion and does not bypass an applicable preference or waiver condition.

When the scale claim is current, choose evidence proportional to the bearer, stakes, feasible observations, and receiving use. A selector-facing superiority claim, durable reusable-bearer policy, safety-material override, or expensive irreversible choice normally justifies a fuller audit. A reversible local probe may need only a small matched comparison. The method shall preserve:

  • (a) Parity and admissibility: comparable task family or use, safety boundary, budget basis, current editions, and set-returning Pareto comparison unless a declared policy lawfully selects another operation.
  • (b) Bearer-appropriate scale dimensions: compute, data, model capacity, or freedom of action for computational methods when they vary; for another bearer, state its capacity, resource, reuse, throughput, coordination, or other scale predicate and explain why it provides a comparable scale basis for the two candidates.
  • (c) Uncertainty appropriate to the evidence: repeated seeds or bootstrap intervals for repeatable stochastic trials when useful; measurement error, interval estimates, case comparison, or another justified form for other bearers. Do not demand seeds from an organization design or FLOPs from an episteme.
  • (d) Cost, resource, and safety visibility: report the accounts material to the decision through their direct patterns. Add B.3 only for a named assurance claim and receiving use, including when a material-reliance threshold requires such a claim.
  • (e) Objective honesty: keep quality, risk, cost, and any policy-promoted coverage or illumination coordinates separate unless their Scale permits the declared operation.
  • (f) Risk-selected design: choose a design that can answer this claim. Fractional factorial, Latin-hypercube, three-level sweeps, heteroscedasticity treatment, and repeated-seed designs are options for suitable multi-knob experiments, not a universal minimum. Record why the selected design is adequate and what it cannot establish.
  • (g) Claim-matched tests: use knee or budget-constrained regret tests only when the asserted advantage depends on a knee or dominance inside the audited window.

BLP-2 — Preference rule with alpha and delta tolerances. Among admissible options at comparable budgets, evaluate the performance contrast at the points and on the objectives that matter to the receiving choice. Apply alpha when a tolerance is needed and delta when the use depends on an assurance comparison. Account for uncertainty in the contrast, including dependence when the runs are paired. A scale-based preference over the audited range requires established Pareto dominance there. If the responses are equivalent under the comparison rule, each wins on a different point or objective, or the evidence still admits materially different orderings, report no scale-based preference and say which of these conclusions is supported. A separately declared local policy can decide only the cases covered by its terms. Cite the policy and its E/E-LOG edition for that choice. For an agentic use, keep the applicable alpha and delta in the comparison, local-policy or waiver result consumed by the C.24 CallPlan.

BLP‑2.1 — Valid waiver grounds (override transparency).

Overrides of a declared local generality preference are allowed only when:

  • Admissibility override: guard rails, ethics, or precedence make the general bearer inadmissible (E.5, E.3).
  • Scale‑probe overturn: under iso‑scale parity in the declared ScaleWindow, the heuristic sustainedly outperforms with uncertainty accounted for.
  • Complementary bias: the heuristic is an inductive bias that improves the general method without blocking scale (graceful degradation as S grows).

All overrides record a BLP-waiver with rationale, admitted review System, waiver-review responsibility relation or A.6.RCD missing governor, and expiry or review in the DRR. Any system-role kind or assignment needed by the review Work is cited separately.

BLP-2.2 — Task-family specialization compatibility. A bounded specialization that makes no scale-advantage, selector-facing superiority, durable generality, or override claim may be used under its direct pattern with an explicit task family, work target, budget guard rails, and evidence locus. It needs neither a full Scale-Audit nor a BLP waiver merely because it is specialized.

If a scale claim becomes current, run the cheap probe first. Select a full audit only when the claim, stakes, feasible evidence, and receiving use justify it. A specialization produced by a general substrate or described as a complementary bias is not automatically compatible: state the non-blocking scale relation and evidence when that fact is relied on. Low-human-overlap or newly discovered approaches remain admissible under the same bounded-use, parity, safety, cost, and evidence rules; novelty neither proves nor defeats scale amenability.

BLP-3 — Prescription architecture is a separate question. The Bitter Lesson establishes no universal preference for prohibitions over positive instructions and no general right to autonomous sequencing. For tool-call planning, use C.24 with its budget, stop, replan, and Guard-Rail rules. For another WorkPlan or normative constraint, use the A.15 planning family, E.3, E.5, and the direct policy or commitment pattern. A project may adopt a minimal-prescription policy only with its own trigger, safety boundary, evidence, and review condition; it is not a consequence of a BLP comparison.

BLP‑4 — Heuristic‑Debt register (mandatory). Record Heuristic Debt only when an admitted heuristic functions as reusable solution-family policy, selector-facing preference, durable override of a general scale-amenable alternative, DRR-backed scale waiver, or project-side choice that claims scale advantage or BLP override. Ordinary local bounded tactics that make no reusable-bearer, scale-advantage, selector-facing, or override claim may remain local and bounded without Heuristic Debt publication. BLP.HeuristicDebtEntry is a C.19.1-local or G.11-linked policy and debt entry; it is not a universal U.* record kind unless separately admitted under E.24.UK, with F.18 for naming, C.3 for Kind semantics, and E.9 for the decision record. For a live debt entry, record scope, admitted review System, debt-review responsibility relation or A.6.RCD missing governor, expiry or review window, and a de-hardening plan; any system-role kind or assignment needed by review Work remains separate. Track the entry in CalibrationLedger or BCT and cite it in SCR.

BLP-5 — Adaptation policy is a separate question. The Bitter Lesson does not by itself require feedback-driven adaptation or make disabling adaptation a waiver. When adaptation is current, use C.22.1 for the task-family adaptation claim, E.23 for repeated object-version improvement, and C.24 for tool-call planning and replanning, together with the applicable privacy and Guard-Rail patterns. A product policy may require or prohibit adaptation, but that result needs an objective, evidence, a risk boundary, and a review date.

BLP‑6 — Precedence & safeguards. BLP draws on P‑10, P‑11, P‑7 and P‑1. Apply E.3 to the principles being compared; the combined policy has no single precedence class. It does not supersede Guard‑Rails (E.5) or precedence rulings (E.3). Where a declared NQD or C.19 E‑LOG policy promotes illumination into dominance, use that lens for the audited window.

BLP‑7 — Publication discipline. When a durable Scale-Audit is performed, its artifacts SHALL be exported to G.11 with the used edition pins, uncertainty method, alpha and delta tolerances when current, ComparatorSet, policy reference, and qualification window. A no scale claim yet or local bounded-use exit creates no audit package.

C.19.1:3 - Conformance Checklist (CC-BLP)

  1. The cheap scale-claim probe names the bearer kind and direct pattern, task family or receiving use, scale predicate, objective vector, comparison basis, feasible evidence, safety boundary, and stakes.
  2. The result says no scale claim yet, local analogy or policy, bounded scale comparison, or full Scale-Audit selected; it does not hide the choice of audit depth.
  3. A performed comparison uses parity, current editions, explicit uncertainty, material resource and safety accounts, and lawful Pareto or declared policy operations.
  4. The evidence method fits the bearer. Compute/data sweeps, seeds, bootstraps, and factorial or Latin-hypercube designs appear only when they answer the claim.
  5. A comparison that establishes equivalence or a trade-off, and a comparison whose evidence cannot yet settle dominance, each records its own conclusion. A subsequent choice by local policy names the condition that the policy covers.
  6. A waiver identifies the policy it overrides, rationale, admitted review System, responsibility relation or missing governor, and expiry or review window.
  7. A live Heuristic Debt entry meets the bounded BLP-4 trigger; ordinary local tactics create no debt record.
  8. Prescription architecture and adaptation policy are routed through their direct patterns rather than reported as consequences of BLP.
  9. An actual durable audit is exported to G.11; a bounded-use or no-claim exit is not padded into an audit artifact.
  10. A narrower specialist bearer names its task family, work target, and the scale, waiver, or bounded-use ground relied on.

C.19.1:4 - Anti-patterns & remedies

  • Slogan as evidence. General, agentic, or Bitter Lesson is treated as proof. Repair with the cheap probe and a comparison basis.
  • Analogy as empirical result. A module relation, organization, or episteme is assigned compute/data scaling semantics. State a local analogy or policy and use a bearer-appropriate predicate and evidence form.
  • Universal experiment recipe. Every claim receives seeds, FLOPs, and a multi-factor design. Select the smallest method that can answer the risk-bearing claim.
  • General wins by default. Compare the actual performance contrast with its uncertainty. Marginal error bars may overlap even when paired runs establish a consistent advantage. If the contrast leaves the order open, record that uncertainty. If two options each win somewhere in the required range, record the trade-off. A generality preference requires a policy that covers the established case.
  • Single-winner leaderboard. Hidden budget mixing or scalarization replaces Pareto comparison. Restore comparable windows, objective coordinates, uncertainty, and policy identifiers.
  • Debt without trigger. A local bounded tactic is entered into Heuristic Debt. Apply the BLP-4 trigger before creating the entry.
  • Elegant-math override. A specialized mathematical lens is selected because of elegance or prestige while scale advantage is live. Use the proportionate BLP comparison; otherwise keep the lens local under the C.29 stop condition.

C.19.1:5 - Archetypal grounding (post-2015; informative)

These computational examples illustrate where a scale comparison may be useful; they supply no empirical conclusion for another bearer family. Each comparison needs its task family, comparator, budget and safety basis, uncertainty, applicable alpha/delta tolerances and any source-currentness condition required by its receiving use.

  • LLMs: prompt programs, retrieval-augmented policies, and MoE policies compared with narrow task-specific pipelines; set-returning selection across editions and budgets.
  • RL and planning: model-based optimization and general agents compared with hand-coded controllers, subject to alpha and delta tolerances and safety.
  • Preference learning: comparisons between RLHF and DPO families.
  • QD and OEE: MAP-Elites, CMA-ME, DQD, and QDax; POET and Enhanced-POET; illumination remains report-only telemetry unless policy promotes it.

Completed bounded comparison. Suppose a specialist and a general search method are tested on the same fixed task set at three matched budgets on the same processor. The account includes the full computational cost, keeps the other material conditions fixed and supplies the following uncertainty bounds for the percentage of tasks solved. Both methods satisfy the same applicable safety and eligibility conditions; no separate assurance result is consumed. The budget tolerance is alpha = 0 CPU-seconds at each matched point.

BudgetSpecialist: solved tasksGeneral method: solved tasks
100 CPU-seconds83–85%79–81%
200 CPU-seconds85–87%89–91%
400 CPU-seconds86–88%94–96%

The specialist wins at the first probed budget, while the general method wins at the other two even after allowing for the supplied uncertainty. Neither response dominates across all three probed budgets. Return no scale-based preference for that whole comparison; the word general does not reverse it. The result supports the stated comparisons at those points, not an interpolated law between them or a claim outside the window.

If the receiving choice is confined to 100 CPU-seconds, the specialist’s bounded advantage can be used under that question’s comparison rule. A separately declared prefer-general policy, if applicable, needs its own tie-break or waiver treatment. No such policy is stipulated here, so no waiver or debt entry is created merely to report these results. If a tariff changes, measured CPU time can remain usable while monetary costs are recalculated. If the task mix changes, compatible results for the affected groups can sometimes be reweighted. A previously untested group needs its own evidence or justified transfer. Reuse only the conclusions supported for the new choice. This is a constructed application, not a new benchmark finding.

C.19.1:5.1 - SoTA-Echoing

Source or source familyAdopted FPF moveRejected overreadPractitioner implication
Yousefi and Collins, Learning the Bitter Lesson: Empirical Evidence from 20 Years of CVPR Proceedings, arXiv:2410.09649, as current empirical pressure around Sutton’s 2019 Bitter Lesson.Treat scale-amenable computational approaches in machine learning as a live empirical comparison pressure.The same empirical result automatically governs modules, organizations, work arrangements, epistemes, or every other bearer.For a computational claim, test the declared task family and scale window. For another bearer, label the move as local analogy or policy and provide its own scale predicate and evidence.
Kaplan et al., Scaling Laws for Neural Language Models, arXiv:2001.08361, and Hoffmann et al., Training Compute-Optimal Large Language Models, arXiv:2203.15556.Keep compute, data, model size, budget, and scale-window relations explicit when a bearer is claimed to improve with scale.Parameter count, compute spend, or one benchmark substitutes for the audited objective vector.State the swept dimensions, alpha and delta tolerances when used, CI or another justified uncertainty form, and budget window before preferring the general bearer.
Lu et al., The Bitter Lesson of Diffusion Language Models for Agentic Workflows: A Comprehensive Reality Check, arXiv:2601.12979.Treat current agentic-substrate claims as evaluation-sensitive and task-family-sensitive, especially when efficiency hype competes with reliability.“More agentic” or “more efficient backbone” proves better workflow performance.For agent-loop or substrate selection, use G.9 for parity, C.19.1 for a current scale-advantage or declared generality-policy claim, and E.23 only for repeated object-version improvement; preserve task-family evaluation, protected trade-offs, and stop/switch conditions.

C.19.1:6 - Payload - exports

BLP.Policy@Context is an editioned local policy row, not a universal kind. It records:

<scopeBranch={empirical-computational | declared-local-analogy}, PreferenceDefault={neutral | declared-prefer-general}, alpha?, delta?, scaleProbeResult?, proportionateComparisonMethod?, fullScaleAuditRef?, WaiverRegister?, E-LOG policyIds?, G.11 telemetryPins?>.

The row omits fields that are not current. PreferenceDefault=declared-prefer-general identifies a local tie-break policy, not an empirical conclusion. A full audit reference appears only after the risk-selected audit exists.

C.19.1:7 - Relations

Depends on: G.5 and G.9 (selector and parity), G.11 (refresh telemetry), A.15.1 for dated work, A.15.2 for work plans, B.1.6 for resource and cost aggregation, C.16 for measurement, and A.10 for provenance, C.18 (NQD‑CAL), C.19 (E and E‑LOG), F.7 and F.9 (bridges, CL, Φ, and Ψ). Constrained by: E.5 Guard‑Rails and E.3 precedence.

C.19.1:7.1 - C.32 architecture-synthesis use relation

When using C.32 to generate candidate architectures, C.19.1 applies only if a claim about a candidate asserts scale advantage or invokes a declared local generality policy. For a universal module relation, platform, organization design, evidence arrangement, or selected structure, that branch is a project analogy or policy: the empirical machine-learning literature is not proof. Name the bearer and its direct pattern, the holon under change when one is current, the bearer-specific scale predicate, objective vector, comparison basis, feasible evidence, admissibility boundary, and intended receiving use.

BLP neither selects the architecture nor turns method-family, practice, role-side, or culture wording into a holon kind. If the candidate removes parts, carries several functions, or is described as reusable, with no recoverable scale claim or declared generality-policy claim, keep the question in C.32 and C.31; return no scale claim yet rather than manufacturing an audit. A TRIZ-style ideality move enters BLP only when the declared comparison relies on scale amenability inside a named window.

C.19.1:7.2 - C.29 mathematical-lens use relation

When a mathematical lens is chosen over a general, scale-amenable bearer because it is elegant, specialized, or theoretically prestigious, and a scale-advantage or declared generality-policy claim is current, C.19.1 governs the scale-advantage and preference claim. A C.29 application may state CandidateMathObject, LensMappingMode, PreservedStructure, LostStructure, LensUseBoundaryValue, declaredLensUse, and StopCondition, with optional blockedLensOverread? for an overread that passes F.19’s plausible-reader test; it does not supply BLP compatibility, scale dominance, or waiver evidence.

If scale advantage or a declared generality policy is live, start with the cheap probe and cite the resulting bounded comparison, risk-selected Scale-Audit, or applicable BLP-waiver. If neither is live, keep the mathematical lens local and bounded by its C.29 stop condition.

Memory hook. Test what scales; label policy when evidence does not decide.

When choosing between a general adaptive loop, a specialized object-family cycle, or a mixed operation-family set for an E.23 use, C.19.1 applies only when the decision relies on scale advantage or a declared generality policy. Start with the cheap probe. Compare the declared characteristic values for material resources, tools and instruments, adaptation attempts, skilled attention, rework or delay, risk exposure, and avoided loss on their admitted scales; keep them separate, reject a dominated option, and use the declared project policy to choose or hold when no option dominates. Net-cost arithmetic is permitted only after every term has been converted to one declared unit through an admissible conversion whose basis, uncertainty, and scope remain visible. Repeated automation alone does not satisfy BLP; the record names the object under improvement, evaluation, protected trade-offs, bounded cost and risk condition, and stop or switch condition. If neither a scale claim nor a declared generality policy is current, E.23 proceeds without a BLP audit or waiver.

C.19.1:End

Referenced in the corpus

50 literal mentions in other sections. Read their context to establish the relation.