Skip to content

v9.1.0 — First Principles Thinking

Choose a tag to compare

@chrisdavidson chrisdavidson released this 11 Sep 15:26
· 712 commits to master since this release

Ships the milestone that names the mechanism behind the class [8.26.0] and [9.0.0] both shipped
live instances of — a hand-written prose claim whose truth is a digit no gate can dispute — and
lets the known-wrong sentences fall out of that answer rather than being fixed one at a time: the
diagnosis (PROSE-01 through PROSE-04), containment reach (CONTAIN-01 through CONTAIN-04),
the generalized digit-narrowing rule and generated narrative regions (NARR-01, NARR-02), the
ratchets (RATCHET-01 through RATCHET-04), and the release act itself (REL-05 through
REL-08). The coverage headline moves 208 reproducible / 97 audit-only / 0 gap / 305 total to
217 reproducible / 106 audit-only / 0 gap / 323 total (18 requirements registered as matrix rows
via a new _rows_v91(), 9 reproducible + 9 audit-only), and the firewall battery tally is
unchanged at 26/26 this milestone — the unchanged total is itself a success criterion (D-D),
never an incidental fact.

Added

  • Diagnosis (PROSE-01 through PROSE-04). Named the mechanism: containment's N → M
    delta-exemption class strips the whole vector including its terminus, so a chain whose final
    value has gone stale reads as exempt rather than as a hit, indistinguishable to the checker from
    one that is current. A sibling site (CLAUDE.md's own Review protocol paragraph, CR-02 below)
    was named before any fix was written. States in writing that the answer covers the
    quantity-shaped class only — a claim whose truth is a digit corroborable against a generated
    fence — and that the consistency-shaped class (one rule stated two contradictory ways, no
    digit to check) is out of scope, filed as backlog 999.50 and 999.51. Each of v9.0.0's carried
    findings now carries a recorded disposition: CR-01 closed at Phase 25's chain-terminus arm,
    demonstrated failing-before and passing-after by mutation on a disposable scratch copy; CR-02
    dispositioned by routing rather than by fix — it states no digit, so nothing in this milestone's
    quantity-shaped scope can corroborate or contradict it — filed as backlog 999.55; CR-03
    remains accepted with its original bound, since docs/README.md carries no recognised generated
    fence for anything to compare its claim against, filed as backlog 999.54; P-CR-01 (found by a
    post-[9.0.0]-entry review, never named in that entry) closed by the fix already shipped under
    it, commit b661281. Disclosed limits — docs/v9.1-claim-containment-diagnosis.md § 3 is a
    dated record of these dispositions as they stood at diagnosis time (2026-09-09); CR-01's own
    closure happened afterward, in Phase 25, and is stated here rather than on that page, which this
    milestone does not reopen as a region host.
  • Containment reach (CONTAIN-01 through CONTAIN-04). CONF-SURFACE's containment check now
    reaches CLAUDE.md, docs/ARCHITECTURE.md, docs/TESTING.md and docs/gates/*.md, floored by
    a set-equality roster rather than a subset check; a chain-terminus arm
    (chain_terminus_problems) watches an N → M delta chain's own final value against its live
    source, the falsifiable closing point for CR-01 above; every REACH-or-LEVEL determination behind
    a containment change is written before the mechanism exists, not after. The deferred-literal and
    deferred-containment ledgers are both reconciled against a live re-adjudication of every entry —
    their growth-only/size-only bound (RATCHET-02, below) and the reconciliation deltas themselves
    are stated in docs/gates/CONF-SURFACE.md's own disclosed-bounds section rather than restated
    here. Disclosed limits — the cannot-reach residue behind the reach above is its own count,
    never absorbed into it: tests/premise-rejection-catalog.md's second G3 demonstration target
    carries no generated region at all, and docs/COMPONENT-DIAGRAM.md's Mermaid node label sits
    inside a fenced code block a non-fenced marker line cannot bracket — both named individually in
    docs/gates/CONF-SURFACE.md's disclosed bounds rather than fixed. This release's own act added
    a fresh entry to the deferred-containment ledger — a currently-correct digit with no
    corroborating generated field, the same raw shape this milestone's mechanism targets — tracked
    by the same growth-only bound rather than hidden; see that page for the entry.
  • The generalized rule and generated narrative regions (NARR-01, NARR-02). docs/PROCESS.md
    section 2's digit-narrowing rule is generalized from the gate-battery count to every restated
    moving count on a product surface, pointing at docs/gates/CONF-SURFACE.md's own generated
    registered_surfaces figure rather than stating a population as a digit. Two generated narrative
    regions (CLAUDE.md, docs/README.md, docs/MEASUREMENT-MAP.md — a third host added beyond the
    two originally scoped) now render their own coverage-headline sentence from the live harvested
    fact, wired into scripts/gen-gate-docs.py --check's drift check alongside the pre-existing
    literal scanner. Filtering a fresh literal scan to the surfaces this migration reaches finds
    zero non-exempt hits, out of 80 total, all exempt. Disclosed limits — that zero is
    narrowed-scope, not tree-wide: the two cannot-reach residues named above
    (tests/premise-rejection-catalog.md, docs/COMPONENT-DIAGRAM.md) sit outside every region a
    narrative-region mechanism could bracket, named as a published, counted residue rather than
    silently absorbed into the zero.
  • The ratchets (RATCHET-01 through RATCHET-04). scripts/gen-gate-docs.py --check runs the
    narrative-restatement census unconditionally on every commit (RATCHET-01), demonstrated by a
    four-arm mutation proof (two independent host surfaces plus the census arm each fail with a
    named file and line; the unmutated arm passes). The deferred-literal ledger carries a
    non-increase assertion with a growth-only/size-only bound restricted to _CONTAINMENT_LEDGER_MAX
    and _DEFERRED_LEDGER_MAX moving in one direction only, at a size a human re-adjudicates
    (RATCHET-02). A diff-review scan flagging a removed literal replaced by a hedge word was
    dropped rather than shipped: RATCHET-03 is discharged by a recorded verdict citing the
    meta-guard-regress argument its own proposer raised, not by a shipped scan — no new gate,
    holding D-D. Every one of RATCHET-01 through RATCHET-03's mechanism changes carries a written
    REACH-or-LEVEL determination made before the mechanism existed (RATCHET-04). Disclosed
    limits —
    none beyond what CONTAIN-04's bullet above already states; RATCHET-02's bound is
    cited there rather than restated twice.
  • The release act itself (REL-05 through REL-08). All 17 hand-maintained version stamps read
    9.1.0, VERSION-01 green, sync-content.py --check clean, VERSION-01's PASS proven
    non-vacuous by mutation on a scratch copy (REL-05). The firewall battery tally is unchanged at
    26/26 — established by a direct count of gate and inline-check registrations at both the
    phase's base commit and this release's own HEAD, with the battery's own printed tally cited only
    as corroboration (REL-06). This milestone's 18 requirements are registered as matrix rows behind
    a live-derived V91-ROWS equality sentinel (9 reproducible + 9 audit-only), and the
    coverage-headline move above is produced by HEADLINE-LOCK's sweep across every
    COVERED_HEADLINE_SURFACES member plus both tracked matrix artifacts, never a hand edit
    (REL-07). Verification for this entry itself is measured against product recurrence — does
    this milestone's own defect class recur, counted the way docs/PROCESS.md counted its four
    prior recurrences — never against whether the new convention was followed; both instruments'
    readings (a frozen-scanner arm and an independent two-reader prose arm), at both the pre-release
    and post-release timing, are published unreconciled in docs/conformance-baseline.md's
    recurrence-reading section rather than restated here (REL-08). Disclosed limits —
    HEADLINE-LOCK verifies every COVERED_HEADLINE_SURFACES member, but gen-gate-docs.py --write
    produces only the narrower, _NarrativeRegion-backed subset of three; the remaining two members
    (docs/requirements-traceability.md, docs/COMPONENT-DIAGRAM.md) require a manual edit on every
    headline move that HEADLINE-LOCK then verifies rather than produces — named in full in
    docs/gates/CONF-SURFACE.md's disclosed bounds. Separately, nothing mechanical joins the 18
    matrix rows _rows_v91() registers to the requirement roster in .planning/REQUIREMENTS.md;
    the correspondence was checked by hand this phase, and a stronger, roster-derived equality check
    was considered and rejected as out of this phase's own scope.

Addendum 2026-09-11, recurrence reading: docs/PROCESS.md's same-class trip, the rework
limit that halts a phase when a defect class it closed recurs, fired on this release's
post-release recurrence reading, read milestone-scoped as the ship phase's own context decided
before any reading was taken. The ship phase therefore closed halted for a root replan while this
release shipped in full; the halt is a statement about rework and changes no requirement's
verdict. The trip rests on sites both independent prose readers recorded at both timings, among
them README.md:143, docs/CONFIGURATION.md:148, docs/DEVELOPMENT.md:171,
docs/ONBOARDING.md:113, docs/TESTING.md:117 and docs/TESTING.md:118; no firing site is
disputed. The reading also found sites made stale during the release block itself, true before
the release acts and false after them because the matrix-row registration moved their live
value: CLAUDE.md:334, docs/README.md:27, docs/README.md:157 and, recorded by a single reader
only and so disputed, docs/README.md:36. The reach map of every site the scanner arm could not
see, grouped by why, is carried to planning-backlog entry 999.77, with dated pointers on 999.54,
999.57 and 999.58 for the sites those entries already own. The figures are in
docs/conformance-baseline.md's recurrence-reading section and are not restated here. This
milestone shipped with its own goal measurably unmet: the quantity-shaped defect class it set out
to contain still recurs in prose its mechanism does not reach.