PM Report — all — 2026-08-03 (last 7d)

Generated 2026-08-03T09:20:38-07:00 · 400 commits · 55 active sessions · evidence: datapack.json · every ⟦ref⟧ traces a claim to its source

HEADLINE

Seven days, 400 commits, four parallel programs — simulator, aerodynamics, language-learning app, and infrastructure — all advanced in the same sprint. The bottleneck is not capability; it is authorization throughput and a handful of open engineering calls that only Aaron can close.

SHIPPED

Flight Simulator — Fidelity & Flyability

• Both roster aircraft now fly their own gear, powertrain, and approach speeds — previously every ship was silently using Config-A's numbers, making every autoland result meaningless as a test. ⟦d84d3266⟧ ⟦8e2a6988⟧ ⟦71dc7b84⟧
• Stream judder eliminated: render interpolation cut the motion coefficient of variation from 0.51 to 0.098 and raised cadence from 7–12 Hz to 20.5 Hz p99 78 ms — the aircraft now moves like it is flying, not teleporting. ⟦645371e5⟧
• The engine windmill-latch bug — prop freezing to fine stop at cruise, delivering 0 kW at any throttle position — was root-caused, bench-reproduced, and fixed; the fixed engine was proven with 6 minutes of real flight at 138 kW on demand. ⟦217f37e5⟧ ⟦c9414d07⟧
• Propeller maps extended to reach the full flight envelope; they previously ended at 89 kt while the fleet flies above 300 kt, making every high-speed power calculation an extrapolation off the edge of a cliff. ⟦cf72e280⟧ ⟦d7761f8c⟧
• Fleet-wide stall defect found and fixed: peak and drop were wired to the same field, so Aaron could not feel a stall in any CS-9 variant. ⟦68744ecd⟧
• Rudder authority capped at 10% after a flat-360 excursion outside the measured lateral envelope — Aaron's call, applied immediately. ⟦a1d6f85e⟧
• Play UI simplified to direct-only control: autoland buttons, mode selector, and envelope-protection warnings removed from the live interface per Aaron's ruling; rates are now the only limits. ⟦96fbb75c⟧
• Staleness beacon shipped: every session now shows a visible tag if the flight-model code has changed since boot, and a "SIM CODE UPDATED" banner fires off any stale build — no more flying unknowingly on old physics. ⟦aa3802d1⟧ ⟦132b6efb⟧
• Fast spawn wired: world opens airborne at 120 kt / 1,000 m / 75% lever — Aaron no longer has to climb from a standing start to test approach behavior. ⟦93ad4b06⟧
• Per-flight notebook built: every session produces a lab-notebook page with power, speed, altitude, and input sparklines so forensic review no longer depends on memory. ⟦315e3c7e⟧ ⟦fae8c9d9⟧

Aerodynamics (CFD)

• CS9-F aero pack delivered on measured geometry — NP 1.2911 m, static margin +6.0% at F's own CG — closing the last placeholder in the two-ship roster. ⟦cea098c9⟧ ⟦bbec2523⟧
• Wake rake delivered: the V-tail's wake on the elevator is now measured, not assumed. ⟦3d2bc231⟧
• Static margin published as a function of angle of attack, not a single cruise scalar — Aaron's point, now in the data. ⟦7d832893⟧
• Long-ESA section polars harvested as a calibration dataset while the solver was already running — no additional compute cost. ⟦bbeaac5a⟧
• A regen-wipe class of defect found and codified: generated output files were silently overwriting hand-corrected measured values on rebuild; a standing audit recipe and a conventions pin now prevent recurrence. ⟦f3427d95⟧ ⟦c129e238⟧
• Unsteady F campaign authorized and in flight; continuation armed. ⟦0657ec32⟧ ⟦30b01e42⟧

Mandarin Lab — Language Learning App

• Full app rebuilt in one day: microphone, live tone contours vs. native references, speech-to-text, AI tutor loop, versioned skill state, and a single-screen UI — deployed live at mandarin-lab.fly.dev. ⟦2620260a⟧ ⟦4a7506da⟧
• Tone-scoring pipeline rebuilt ground-up (v4): SwiftF0 pitch extraction, voiced-run segmentation, register normalization, and best-of-3 native TTS references — cross-validation stable at 80/77 matched, 33 mismatched. ⟦e2a05ea2⟧ ⟦9617523e⟧
• Three scoring bugs that were penalizing correct pronunciation fixed: utterance-edge creak manufactured a fake tone-4 plunge; a syllable-boundary algorithm swallowed one syllable inside another, scoring a correct take at 38; and a run-counting assumption mislabeled Aaron's emphatic rising tone as falling. ⟦cc38b879⟧ ⟦059de8b3⟧ ⟦33be3ba2⟧
• Retry flow added: a weak take holds the sentence for another attempt instead of advancing — Aaron's verbatim complaint was "it just moved on without giving me an opportunity." ⟦0e0a280b⟧
• Replayable native voiceover added to every tutor message: one button re-speaks any Mandarin sentence on demand, generated and cached server-side. ⟦10da4395⟧
• App opened to the public with safety rails: per-IP sliding-window rate limits, global daily caps on all AI endpoints, 2 MB body cap, and anonymous per-visitor registration — no auth token required. ⟦6d044b6d⟧
• Four confirmed-live security holes closed after a 4-run red-team review: ungated API endpoints, unvalidated inputs, and per-IP abuse paths. ⟦da3c9ef9⟧

Infrastructure & Harness

• Lane ledger shipped: every AI work lane now records its own label, session ID, last-active timestamp, and open/closed state at runtime — continuation after a crash no longer depends on inference. ⟦5f418e0f⟧
• File-transfer scatter fixed: pushes were distributing files across the wrong machines in a multi-machine fleet. ⟦5f8d3bb5⟧
• Positions ledger over-admission closed: the ledger was growing because of an admission bug, not missing pruning. ⟦59f891a7⟧
• Model roster probe built: all three provider APIs are queried live at preflight and diffed against a snapshot baseline of 83 OpenAI / 25 Google / 11 Anthropic models — experiment arms are never built on stale model lists. ⟦317ce8d7⟧ ⟦882a3032⟧

IN FLIGHT

• Aero-CFD unsteady campaign — continuation armed, poller live; results pending. ⟦30b01e42⟧ ⟦3e1226b9⟧
• Sim currency audit: annunciators, alpha requal, and brackets remain open; two discrepancies routed for resolution. ⟦437c0b75⟧ ⟦c82ebdb7⟧
• Thermal duty cycle for the propulsion system: Aaron has not yet ruled; this blocks the engine seat's next gate. ⟦5da85114⟧
• First crash-case telemetry routed to the engine seat — the slide-vs-cartwheel energy split is an open question. ⟦9b69bf77⟧
• Harness fetch channel: root-caused twice and now proven end-to-end, but live acceptance under continued watch. ⟦d6942ea2⟧

TOKEN SPEND — TYPE, SCALE, IMPACT

400 commits across 7 days represents sustained multi-lane parallel execution: simulator physics, CFD solves, app development, and infrastructure running concurrently. Output tokens (new reasoning, code, analysis) drove the substantive work — new physics fixes, the tone-scoring rebuild, the security audit, and the harness features. Cache tokens (re-reading stable context like hangar specs, runway data, and the sim's physics contracts) kept each lane orientated without re-deriving shared ground truth on every turn; in stakeholder terms, cache is the cost of continuity, output is the cost of progress. The dominant spend categories were the sim physics loop (multiple diagnosis-fix-verify cycles on windmill latch and map coverage), the Mandarin Lab pipeline rebuild, and CFD campaign management. ⟦d:commit_count_400_7day⟧

RISKS & OPEN CALLS

• Authorization throughput is the rate limiter. Multiple items — thermal duty cycle, crash-test placement, g-limit ruling, M3 model evaluation — are queued waiting on Aaron's call. Work is not blocked by capability; it is queued waiting on decisions that only he can make.
• Regen-wipe class is not fully contained. The Long-ESA NP hand-correction was silently overwritten by a rebuild once; the audit found the CS9-D case unreproducible. A standing audit recipe exists, but the class can still bite in any seat that generates output files. ⟦f3427d95⟧ ⟦f8ab0034⟧
• Mandarin Lab scoring is empirically validated but narrow. Cross-validation runs at 80/77 matched against 33 mismatched on a small reference set. The pipeline corrects itself when Aaron catches an error, but it has misfired on three distinct acoustic phenomena in one week — creak, syllable boundaries, and emphatic contours. More speaker diversity would stress-test it before a wider rollout. ⟦cc38b879⟧ ⟦059de8b3⟧ ⟦33be3ba2⟧
• CFD unsteady campaign has an open scope question. The deep-stall plateau campaign cannot answer Aaron's original question as scoped — the amendment was authorized, but the answer space is narrower than commissioned. ⟦098a3e14⟧

RECOMMENDATIONS

• Schedule a single 30-minute ruling session to close the four queued authorization items (thermal duty cycle, g-limit, crash-test placement, M3 go/no-go) — each is blocking a downstream gate and the queue will compound if held another week.
• Commission a second-speaker validation pass on the Mandarin Lab tone scorer before the public open round generates scale; three acoustic misfire classes in one week on one speaker is a meaningful signal about fragility at the edges of the training distribution.
• Formalize the regen-wipe audit as a required pre-gate step for any seat that owns generated output files — the current recipe is correct but optional, and two of the seven audited seats showed defects on first inspection.