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 22:10:10 UTC

E.2:6 - Policy — Bitter‑Lesson Preference (BLP)

Intent and activation. In search, learning, planning and related computational work, investigate a claimed advantage from increased compute, data or search capacity under C.19.1. The empirical Bitter Lesson motivates that comparison; it does not decide every choice between a specialized and a general approach. An analogy or generality preference for another bearer needs a separately declared local policy, receiving purpose, scale predicate and evidence basis. Ordinary bounded specialization can stop with its task-specific justification.

BLP-1 — Claim-matched comparison and audit. Begin an activated BLP use with C.19.1’s cheap scale-claim probe. Name the candidate, task/use, scale variable, objective, usable budget range, receiving conditions and evidence form. Missing comparison premises return no scale claim yet; select evidence depth by claim and risk. An independently activated profile can require a fuller Scale-Audit. That audit retains:

  • (a) Parity and admissibility: comparable tasks, budget basis, editions, freshness and risk/safety envelope. Use G.5/G.9 when their selector/parity contracts apply. Pareto comparison is the default over the declared objective vector; another comparison needs a lawful declared policy under E.3.
  • (b) Usable budgets: vary the dimensions relevant to the claim—compute, data or freedom of action where applicable—and pin held conditions. Compare only over the range available to the receiving use; extrapolation needs its own basis.
  • (c) Response and uncertainty: report performance over that range with an uncertainty treatment appropriate to the evidence. Slopes may characterize response but cannot replace its level, admissibility or receiving floor. Repeatable stochastic trials retain seeds/replicates and confidence intervals where used; deterministic derivations retain their assumptions and validity bounds.
  • (d) Resources and assurance: retain decision-material resource accounts through A.15.1, A.15.2, B.1.6, C.16 and A.10 as applicable. A named assurance claim uses B.3; independently applicable assurance and oversight requirements remain binding.
  • (e) Objectives and tolerances: declare quality, risk, cost and any illumination coordinate promoted by an explicit CAL policy. State units and the budget/assurance tolerances α/δ used by the comparison. Publish the audit’s edition and policy pins through G.11.

BLP-2 — Usable-performance preference. Among admissible options under the receiving conditions, with comparable assurance within δ and budgets within α when those tolerances apply, prefer the option whose usable performance is shown to Pareto-dominate across the audited range. Evaluate uncertainty in the performance contrast, including dependence between paired observations. Return no scale-based preference when the comparison establishes equivalence, establishes a trade-off, or leaves the ordering open. Say which outcome the evidence supports. A project policy under E.3 can decide the cases covered by its terms. Cite that policy and the cost of applying it for the resulting choice.

BLP-3 — Task-appropriate prescription. Positive procedures and prohibitions are both available under their task or control needs. A mandated safety procedure keeps its force. Minimal prescription and autonomous sequencing need their justified policy, authority and guards; neither follows from a scale comparison. Ordinary use of a positive procedure requires no BLP waiver solely because a general search alternative exists.

BLP-4 — Conditional heuristic debt. Apply C.19.1’s debt branch when an admitted heuristic carries a durable scale/generality preference, reusable solution-family policy, selector-facing preference or override within an activated BLP use. Retain scope, review responsibility, expiry/review window, replacement target and de-hardening plan in CalibrationLedger/BCT and SCR as applicable. A bounded task-specific tactic that makes none of those claims is not debt merely because it is specialized.

BLP-5 — Separately authorized adaptation. Enable, require or prohibit adaptation under the product’s objective, evidence, permission, change authority and risk/Guard-Rail conditions. A fixed method needs no BLP adaptation waiver solely for remaining fixed. A product policy that requires adaptation or its review retains that obligation and its exception rule.

BLP‑6 — Precedence & Safeguards. BLP combines governance and architectural principles; classify each compared principle under E.3 rather than assigning one class to the whole policy. It is instantiated by Pillars P‑10 (Open‑Ended Evolution), P‑11 (SoTA Alignment), P‑7 (Pragmatic Utility), and P‑1 (Cognitive Elegance). It does not override applicable safety and ethics requirements, E.5 Guard‑Rails, or E.3 precedence rulings; where BLP conflicts with Guard‑Rails, Guard‑Rails prevail. When NQD/E/E‑LOG elevates illumination to dominance for exploration mandates, BLP adopts that lens rather than overriding it.

BLP applies to comparisons such as prompt programming versus fine-tuning and model-based planning versus hand-tuned control when they make the scale or generality claims specified above. Their tasks and receiving conditions determine the comparison; G.5/G.9 apply when their selector/parity contracts hold.