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 22:20:20 UTC

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.