Skip to content

v9.12.0 — First Principles Thinking

Choose a tag to compare

@chrisdavidson chrisdavidson released this 26 Sep 02:11
· 329 commits to master since this release

v9.12.0 The Exemplars Audited

Delivered: twenty-one backlog commits, no milestone phases, one subject — the
shipped worked examples, and the instruments that were supposed to be reading
them. Where v9.11.0 made the control loop's state visible in the artifact, this
release turns the same scrutiny on the exemplars themselves.

The defect class this closes. Every exemplar defect fixed below sat in a file the
whole battery passed, because no gate asked the question that would have caught
it. That is why half of this release is new instruments rather than more prose.

Key accomplishments:

  • The cold tank is not the heat sink (999.166). The Carnot drill took the
    molten-salt cold tank (563 K) as its cold reservoir. That tank is part of the
    heat source; the steam cycle rejects to a condenser near ambient. The bound was
    on a machine that does not exist, and real hardware beat it on measured data —
    the tank pair permits at most 32.8%, while Solar Two, whose tanks sat at exactly
    those temperatures, measured 34.1% at full load. T_cold now comes from published
    design condenser pressures (316 K wet-cooled, 330 K air-cooled), the ceiling is
    ~61–62%, and the "conventional" tier is relabelled to the thermal-to-electric
    design point it has to be to sit in the same bracket. "Irreducible fraction:
    zero" is retracted — reversible heat transfer needs infinite area or infinite
    time. Every replacement figure was read from its primary source in-session
    (OSTI 1035342, 1088078, 1820100, 793226), not carried from the adjudication.
  • The unmeasured share is not a known share (999.128, 999.129). Seven sentences
    written before Phase 46's rescope still spoke as though the per-stage split were
    measured, three of them asserting a stage ranking inside the same clause that
    declares the split unmeasured. Hedge propagation only: every arithmetic literal
    is byte-unchanged, verified by diffing the numeric multisets of both revisions.
  • Direction-neutral theoretical limits (999.171). The prescription read as
    maximisation-only, so a latency bracket — where lower is better — labelled its
    55.7 ms floor an "Ideal ceiling" and shipped two tiers where three are required.
    Ceiling and floor are now named as two instantiations of one governing bound, a
    fix-the-direction step comes first, and the missing middle tier is restored with
    a measured, dated, endpoint-named figure.
  • Four ground-truth bookkeeping defects: a duplicate GT-3 split and renumbered
    (999.174); that new GT-7? given the verification path D-07 requires, with its
    two causes made conjunctive (999.175); estimate-fermi's GT-7? re-anchored on its
    primary source with every hop re-run, its provenance sentence having been false
    even before 999.166 (999.169); and a GT-5? mention adjudicated as characterising
    rather than consuming (999.170).
  • A rival that ruled out the wrong proposition (999.156), and second-order's two
    lenses stated at the point of use (999.172) — the actor/time lens vocabulary had
    reached 1 of 15 live captures, and 0 of 5 on a prompt built around four plainly
    distinct actor groups where second-order was named in 4 of 5.
  • TRACE-03 DELIV-PATH is green in CI again. It had failed since v9.10.0 and the
    cause was never read; two release runs went red here and shipped anyway. Two of
    the five dangling deliverable paths were unregistered, both under gitignored
    .planning/ — exactly the class of gate that passes a local scan and fails only
    in a fresh clone.
  • PROV-ROLLUP is registered (999.173). scripts/check-provenance-rollup.py shipped
    with a real failing check and its own falsifiability controls, registered in
    neither the battery nor CI — so the reading it produced gated nothing. That is
    how 999.176's fourteen non-conforming exemplars shipped with the battery GREEN.
    Battery 24 -> 25, plus a matching CI job.
  • The claim-chain judge (999.4). untraced_claims read a claim as traced by any of
    three citation routes; all three ask whether a citation is present, none opens
    the chain. The judge opens it. Its own first live reading over the frozen probe
    set found two defects in the instrument, both fixed here, neither touching a
    frozen detector.
  • A three-valued technique-adoption scorer (n/a / absent / present) and the v9.12
    prompt catalog (999.147), because the 3-generation eval could not distinguish a
    technique that correctly declined from one that fired and was ignored.
  • Ten more barred literals in RETRACT-01 (11 -> 21). Each measured at 0
    occurrences before registration using the module's own normaliser, and each
    mutation-tested. One deliberately not registered, with the reason recorded: the
    sentence was false only in co-occurrence with a number, and a literal register
    cannot express a co-occurrence.
  • All 17 stamps 9.11.0 -> 9.12.0. FIREWALL: GREEN (25/25); sync-content.py
    --check clean; VERSION-01 green on all 17 stamps.

Not claimed: the coverage headline does not move — 248 / 248 / 0 / 496. No
requirements matrix rows are registered, because this release has no milestone
requirement ids to register; it is twenty-one backlog entries, not a phased
milestone. Inventing ids after the fact to make the number grow would be the
defect this project spends most of its gates preventing.

See CHANGELOG.md for full details.