---
title: "The anytime-algorithms/metareasoning tradition — the strongest AI-historical candidate for a Stent-premature 'inference-time compute' — fails Stent's test: it was cross-cited and continuously built upon, never disconnected from canonical AI knowledge"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://erichorvitz.com/u87.pdf"
source_title: "Reasoning about Beliefs and Actions under Computational Resource Constraints"
source_author: "Eric J. Horvitz"
source_date: "1987-07"
source_venue: "Proceedings of the Third Workshop on Uncertainty in Artificial Intelligence, Seattle, WA, July 1987, pp. 429-444 — confirmed verbatim by the paper's own footnote 1, which credits it to 'AAAI and Association for Uncertainty in Artificial Intelligence, Mountain View, CA' (so 'AAAI Workshop' alone understates the co-sponsorship); self-hosted on erichorvitz.com. [unverified-venue — reprint] The capture also recorded 'reprinted in Uncertainty in Artificial Intelligence 3, Kanal et al. eds., Elsevier, 1989, pp. 301-324'; the primary document says nothing about any reprint, and the one bibliographic record found (mlanthology.org, 2026-10-23 audit) attaches the pp. 301-324 range to DOI 10.1016/0888-613x(88)90148-x — an Elsevier PII whose ISSN component and (88) year point at a 1988 journal, not the 1989 Kanal book. Both the vehicle and the page range of the reprint are therefore unconfirmed; no claim in this note rests on them."
source_tier: 1
source_sha: "e24f8cd0e0ef5bcba1d10c8976f35be9d31aed2cfb3518cc47691400cc5ef985"
source_quote: "We address the notion of real-time rationality, focusing on the application of knowledge about the expected timewise-refinement abilities of reasoning strategies to balance the benefits of additional computation with the costs of acting with a partial result."
source_url_2: "https://cdn.aaai.org/AAAI/1988/AAAI88-009.pdf"
source_title_2: "An Analysis of Time-Dependent Planning"
source_author_2: "Thomas Dean, Mark Boddy"
source_date_2: "1988"
source_venue_2: "AAAI-88 Proceedings, pp. 49-54, footnote 6"
source_tier_2: 1
source_sha_2: "6559c7f4933aa178cb0de5d716c3fb3526c24b837997f31235a013aeaab56102"
source_quote_2: "His conclusions concerning the algorithms and architectures appropriate for effective reasoning under resource constraints are entirely compatible with the discussion in this paper."
source_url_3: "https://arxiv.org/pdf/2609.14995"
source_title_3: "Intelligence Under Time Constraints: Rethinking Test-Time Compute"
source_author_3: "Xiaotian Zhang"
source_date_3: "2026-09-15"
source_venue_3: "arXiv preprint, cs.CL (arXiv:2609.14995v2); Trooly.AI — single-author, unrefereed preprint"
source_tier_3: 1
source_sha_3: "cb341531a808fed67ce1e4482ae41de04d88ae66324ddda5dcc6df17da50fc89"
source_quote_3: "Bounded optimality relates behavior to an agent's architecture and environment (Russell and Subramanian, 1995)"
source_quote_3b: "anytime algorithms expose quality as a function of deliberation (Zilberstein, 1996)"
independent_roots: 3
root_shared: "Three independent primaries: Horvitz's own 1987 paper, Dean & Boddy's 1988 paper, and Zhang's 2026 preprint. The 2026 preprint is a single unrefereed source and is corroborating context (what the field says about its own lineage today), not the load-bearing evidence for the 1987-88 non-disconnection finding, which rests on the two 1987-88 primaries agreeing independently."
audit_status: "capture-verified — the bee fetched all three source_urls directly at capture time (source_sha recorded for each); the queen cannot independently re-fetch here (no network by design), pending the verifier bee's independent re-check. source_url_3 is a single-author, unrefereed arXiv preprint (Cali ruling 2026-07-29, single-source concentration cap): this is the first claim-note in the vault resting on it, well under the three-note cap, and the preprint supports only the 'modern researchers still frame this as continuous lineage' sub-claim, not the 1987-88 non-disconnection finding itself. UPDATE 2026-10-23 (scheduled audit, claude-opus-5, cross-model): the independent re-check has now been performed and the 'cannot independently re-fetch here' hedge above is superseded. All three source_urls were re-fetched directly this session via extract_pdf and all three re-fetch sha256 values are IDENTICAL to the capture receipts (e24f8cd0e0ef…, 6559c7f4933a…, cb341531a808…). source_quote verified verbatim in the Horvitz abstract by the auditing session's own read; source_quote_2 verbatim in Dean & Boddy's footnote 6 (p. 54 — footnote marker confirmed as 6, attached in their Conclusion, exactly as source_venue_2 states); source_quote_3 and source_quote_3b verbatim in Zhang §2, whose heading is literally 'Intellectual Lineage', and which names Russell & Subramanian 1995, Zilberstein 1996, Russell & Wefald 1991, Boddy & Dean 1994 and Shahaf & Horvitz 2009 in sequence — corroborating the note's 'unbroken lineage' sentence as written. Venue details checked: Horvitz's own footnote 1 confirms 'Proceedings of the Third Workshop on Uncertainty in Artificial Intelligence, Seattle WA, July 1987 ... pp. 429-444'; Dean & Boddy's running feet confirm pp. 49-54; arXiv:2609.14995v2 confirms title, sole author Xiaotian Zhang, Trooly.AI, cs.CL, 15 Sep 2026. One CORRECTION applied to the body: it had read 'Two groups, writing the same year, recognized each other's work as one connected program' — the years are 1987 and 1988, and the citation runs one way only (Dean & Boddy to Horvitz, who could not have cited them), so mutual recognition is not on the textual record; narrowed to the one-directional, dated citation, which strengthens the non-disconnection finding rather than weakening it. NOT re-verified: the 1989 Elsevier reprint claimed in source_venue — see that field's own flag."
provenance: "Promotion from 10-inbox/raw/2026-10-21-does-the-history-of-ai-or-cognitive-science.md, 2026-10-21 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-10-21-does-the-history-of-ai-or-cognitive-science.md"
date_created: "2026-10-21T00:00:00.000Z"
tags: ["history-of-ai","test-time-compute","inference-time-compute","anytime-algorithms","metareasoning","stent-prematurity","philosophy-of-science","cognitive-science"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-23 (strengthened, scheduled audit claude-opus-5) — source_quote is verbatim in the TEXT LAYER after all, read directly by the auditing session at lines 19-21 of the pdftotext extraction of e24f8cd0e0ef…; the cause of the 2026-10-22 FAIL is now identified as a dropped-fi-ligature extraction artifact, NOT a quote mismatch: the extractor renders 'timewise-refinement' as 'timewise-re nement' and 'benefits' as 'bene ts', so a byte-exact sweep cannot match while the printed sentence is word-for-word identical to source_quote. This supersedes the 'WEAKER than a text match' caveat below — the vlm re-read's verdict is now corroborated by a model-free read of the same bytes. Lesson for the sweep: normalize fi/fl/ffi ligature gaps before declaring a FAIL on a pre-1990s or TeX-set PDF. Prior history preserved: PASS 2026-10-23 — recheck via vlm-reread (WEAKER than a text match; see verified_vlm_reread): source_quote verbatim in a local vision-model re-read of the capture-time PDF (sha256 e24f8cd0e0ef…), not in its damaged text layer; was: FAIL 2026-10-22 — source_quote NOT found in the capture-time archive (sha256 e24f8cd0e0ef…) — the quote does not match the bytes read at capture; repair before promotion"
verified_vlm_reread: "2026-10-22 — source_quote verbatim (normalized) in a LOCAL VISION-MODEL RE-READ (glm-ocr@6effedd0dc8a) of page(s) 1 of the cached PDF, not in its text layer; the original text corroborates at 0.891. WEAKER than verified_verbatim — a model read the page (seek_verify v1.3)"
---


[[entity-gunther-stent|Stent's]] 1972 criterion for a **premature** discovery
is specific: one whose implications "cannot be connected by a series of
simple logical steps to canonical, or generally accepted, knowledge"
([[claim-stent-prematurity-covers-neglected-not-contested-ideas]]). The
strongest AI-historical candidate for a Stent-premature articulation of
"think longer at answer-time" —
[[entity-anytime-algorithm|anytime algorithms]] and the adjacent
metareasoning tradition — fails that test on its own textual record.

[[entity-thomas-dean|Dean]] & Boddy's 1988 paper does not treat
[[entity-eric-horvitz|Horvitz]]'s contemporaneous 1987 paper on reasoning
under resource constraints as a separate or incompatible idea; it folds it
into the same research program: "His conclusions concerning the algorithms
and architectures appropriate for effective reasoning under resource
constraints are entirely compatible with the discussion in this paper."
Horvitz's own paper independently states the same core trade-off in
decision-theoretic terms: "We address the notion of real-time rationality,
focusing on the application of knowledge about the expected
timewise-refinement abilities of reasoning strategies to balance the
benefits of additional computation with the costs of acting with a partial
result." The recognition on the textual record is one-directional and
chronological rather than mutual — Horvitz published in July 1987, Dean &
Boddy at AAAI-88 the following year, citing him as a compatible contemporary;
Horvitz could not have cited them. One group reading the other's work as the
same program, in print, within a year, is still the opposite of
unconnectability: the connecting step was taken immediately, not decades
later. Horvitz's own §5.4 takes a further step backward, to
[[entity-ij-good|I. J. Good]]'s type II rationality, which makes the lineage
longer rather than more broken.

The tradition kept accumulating afterward rather than being forgotten and
rediscovered: Horvitz's "value of computation," Russell & Wefald's 1991
"rational metareasoning," Boddy & Dean's 1994 "deliberation scheduling," and
Zilberstein's 1996 survey. A 2026 position paper on LLM test-time compute
still traces an unbroken "intellectual lineage" through this exact sequence —
naming "[[entity-stuart-russell|Russell]] and Subramanian, 1995" on
[[entity-bounded-optimality|bounded optimality]] and "anytime algorithms
expose quality as a function of deliberation (Zilberstein, 1996)" as direct
theoretical ancestors of present-day test-time-compute allocation. A
continuously-cited, continuously-built-upon research program delayed by
missing compute and capable base models — not by missing conceptual
connection — is the opposite of Stent's category. See
[[claim-dean-boddy-1988-anytime-algorithms-formalize-compute-for-quality-tradeoff]]
for the formalization itself, and
[[claim-stent-here-and-now-prematurity-covers-unresolved-claims]] for how
narrowly Stent's own text polices "premature" even in the prospective case.

> [!note] Seek's commentary:
> This is a sourced *no* to the question that sent this session looking, and
> I think the no is the finding, not a dead end. The hook's instinct — Stent's
> category is exactly the shape a rediscovered-test-time-compute story would
> take — was reasonable. The primaries just won't say yes: Dean cites Horvitz
> as a contemporary, not a stranger; the field's own 2026 retrospective draws
> the line straight back through Zilberstein and Russell without a gap to
> explain. What waited thirty-six years here wasn't a connection — it was a
> model worth thinking longer with. I'd keep an eye on whether that
> distinction (logically disconnected vs. economically premature) is itself
> worth its own note; it's doing real work above. — Seek