2026-07-12

DAILY BRIEF · JULY 12, 2026

The Floor Hunt Closes; Pre-Risk Stays Open

Six candidate mechanisms are now excluded and the ~0.14 m cruise floor is architectural — so the target was not lowered to meet it.

Evening edition — the day is still open; counts are partial.

Major Developments

The error floor is architectural, and it was declined. The question that has run for a week — what holds the along-track cruise error at roughly 0.14 m? — was answered today by exhaustion. Four pre-registrations were frozen and measured: the H3 modal sensitivity, the Tikhonov ε_Γ attenuation, the post-integration velocity clip, and the joint-6 posture forcing. None of them carries the floor. With the three channels excluded on 2026-07-06, that closes six named mechanisms, and the three static forcings together account for at most 24% of the lag. The verdict recorded in error_floor_architectural_close.md is a quasi-static equilibrium between trajectory geometry and controller stiffness, and it holds at pe_p99 0.1412–0.1483 m with coverage 0.997–1.0 and all three gates passing.

The report is careful about its own scope, and so is this brief: the floor is unreachable at this trajectory geometry and gain envelope without an architecture change. It is not a law of the system.

That distinction drove the day’s real decision. A ~0.14 m floor could have been adopted as the operating point, which would have closed the pre-risk phase. It was not. The ratified choice was to open a reference-side geometry-and-stiffness rung first, and the pre-risk target stays open at sub-0.10 m (card 86bawkkku, whose title now carries the ratification). Lever scoping followed: of four candidates, only A — posture-aware reference shaping — is target-sized, and even A lands near 0.11 m alone, which is marginal against a sub-0.10 m target. B and C are likely dead on the direction and magnitude gates; D caps at the 24% static share and can only trim. The recommendation is A and D stacked.

The Lean library sealed the 7-DOF rung and the interface rung in the same day. Sixteen new theorems landed with zero sorry and axioms confined to {propext, Classical.choice, Quot.sound}: seven in SevenDofDecoupling, four in KoenigDecomposition, two in SigmaMinDet, three in CoupledSandwich. The ledger moved from 141 theorems to 157 across 36 modules, 22 of them proved full. König’s second theorem — the cancellation that leaves the centre-of-mass row Coriolis-free — was verified n-generically, reproducing Giordano 2019 RA-L eq. (21)→(22a) from scratch, and the citation was added to both terminology registries against Cline 2018 §2.10.

Three interfaces were discharged and three were left standing as honest walls, stated on the page rather than quietly assumed. That is the pattern worth keeping: Lean holds what Lean can hold, and the boundary is written down.

A measurement defect was found and cleared before it could taint anything. run_statics had been sizing orbit grids one cell short at twelve call sites. The consequence was false gate failures — the rung-3 log read 0.8140 coverage when the true figure was 0.9972. The fix passed 89 tests, the census was widened from 7 to 17 completed orbit-synced logs, and every one of them re-graded clean. Decisions affected: zero. The defect had been understating completion by 13–41% and had moved nothing.

The vault graph was made honest. 145 [[current_sota]] wikilinks pointed into Inspection/, which Obsidian does not index — so every one of them was a dead edge that rendered as a link. They are now repo citations across 33 pages, with 109 equation anchors preserved rather than flattened away. Broken links fell from 406 to 261. The edge count did not move, which is the proof: those edges never existed.

Tried and Failed

The Tikhonov ε_Γ attenuation died by magnitude. The sign was lag-correct and the alignment band (0.56–0.74) was measurable, but the denied velocity came to ~0.05 mm/s against the 157–243 mm/s the floor requires — three orders short. Dead as a lever, and no re-pin was needed.

The velocity clip died too, at under 0.2% closure of the fold lag, with the fence cross-gate agreeing to 3e-4 or better. Both velocity-space candidates are now gone.

The deck was wrong in five places until the attestor caught it. It claimed six pre-registrations on a day that froze four; it counted 59 commits over a session that opened the previous evening, not the 51 of the calendar day; it undercounted the night’s commits as 12 when 21 landed; it said the post-fix scoreboard was “born from run PM1221” when the file itself declares it an offline from-log overlay with no simulation at all; and it told the reader it was a 13-slide deck while holding 26. All five are corrected. None had reached the user.

Housekeeping caught four genuinely broken citations, one of them self-inflicted today: open_boxes.py was retired to tasks/archive/ this morning while notes/torch_sprint.md still cited its old home. Fixed tonight.

Two further “broken” citations were false alarms, and the reason is worth naming precisely. notes_lint.py skips placeholder paths via an exact-match set, so a bare (~/Code/...) is ignored — but an elided path carrying a filename, (~/Code/.../current_sota.md), is not, and gets checked against disk. The two hits are lines in notes/vault_graph.md that document the citation convention in a before-and-after table. Filed as 86bawnvjw with the one-line fix. Note for whoever takes it: this is not the backtick-widening the module’s docstring forbids retrying — it narrows the existing placeholder guard and leaves backtick handling alone.

Lessons Learned

A session is not a day. The closing tail counted commits across a session that opened the previous evening and reported 59; the calendar day holds 51 (47 in Code, 4 in ctrllib). Both numbers were honestly derived and only one belongs in a daily deck. Any count that reaches the reader needs its window named.

A retired word has to be retired everywhere, the same turn. “fossil” was struck from the D12 record at 12:24 and still sat in a slide title tonight — and it still sits in the description of ClickUp card 86bawjbtz. The notes were cleaned; the outward-facing copies were not. The deck is the copy a supervisor actually reads.

Scope qualifiers are the first thing to fall off a slide and the first thing an examiner asks about. The report says the floor is unreachable at this trajectory geometry and gain envelope. The slide said “0.14 m is architectural”, full stop. The hedge was restored, along with the measured band, because the unhedged claim is the one that cannot be defended.

Questions and Ambiguities

The M7 result carries a named confound. Pinning τ_t kills the overshoot’s α-dependence — pure-D ×6 reads 0.3533 m against co-scale ×6 at 0.1629 m, some 37% below the 0.2589 m ceiling — but K_e stiffening rides along with τ_t. The K_e-only isolating control is the clean follow-up and has not been run.

The live current_sota.md disagrees with the running code in three places — the §6.3 derate ramp, the §6.2 σ_crit, and §4.5, which omits the posture forcing (card 86bauyyae). This is a separate hazard from the Docs/ copy with the sign error, and easy to confuse with it.

Lever A alone lands near 0.11 m against a sub-0.10 m target. A-and-D stacked is the recommendation, and it is a recommendation, not a result.

Five design documents still carry state: draft or frozen-pending-commit although they were frozen and then measured. The record should say what actually happened to them.

Two divergent copies of start_here.html exist, and only the one under website/quartz/ received today’s milestone link. Checked, and it is not a deploy blocker: site_publish.sh builds solely from website/quartz/, so the stray copy at website/start_here.html is never served. It is an orphan of the same family as the Docs/ copy of current_sota — harmless today, misleading the day someone opens it. Worth deleting, not worth gating on.

Next steps

1. Deploy the site — USER ONLY (86bawkra8); the Cloudflare OAuth is yours. Nothing blocks it: bash scripts/site_publish.sh --dry-run first, then the real run.
2. Open the architecture rung (86bawkkku): A + D stacked is the scoped first move.
3. Run the K_e-only isolating control to clear the M7 confound.
4. Assimilate backlog: 3 open targets, 1 marked urgent (skillscript, 86bawcy1j).
5. Regenerate the session dashboard — its range ends 2026-07-06 and no longer reaches the present.

By the numbers — 51 commits (47 Code, 4 ctrllib) · 29 slides · 4 pre-registrations frozen and measured · 16 Lean theorems, zero sorry · 17 logs re-graded, 0 decisions tainted · 145 dead wikilinks repaired · 3 pre-freeze audits + 5 attestation passes. The session dashboard is not the source here: its range ends 2026-07-06 and does not reach today. These are the day’s own countable facts, and the day is still open.