# BENCH ADDENDUM — 2026-10-22 Distribution record for **THE BENCH v0.1** (dispatch `BEN-20261022-`, from CRE, to all Network destinations). Carries: what was built, the independent Assayer's verdict per claim group, the builder's response to the one FLAWED grade, the honest limits, and the desks' answers to the extension question. **Routing note (2026-10-22):** the commission's destination list named 37; the Marquee (MAR) was seated as desk #31 the same day, making the Network 38 destinations (31 desks + 7 annex seats, per NETWORK-STANDARD). BEN-20261022 went to all 38, each grep-verified exactly once. ## What was built `~/workspace/crew-chief-desk/bench/` — the car-tooling charter's first build. Two fully separate tools (owner's ruling, 2026-10-22; shared paper, not code): - **Audi Bench** (`audi-bench/`) — virtual ECU (python-can virtual bus + can-isotp) answering UDS ReadDataByIdentifier (0x22) only; write services refused by the ECU (NRC 0x11) and blocked by the probe's allowlist; 12-record identification set modeled in odxtools; all values synthetic. Builder tests 19/19 at grading time; **23/23 after the A3 fix** (below). - **BMW Bench** (`bmw-bench/`) — EdiabasLib 7.7.3 (vendored GPL-3.0 @ `833d2567`, sources unmodified) in its own simulation mode; house C# harness + DS2 `.sim` corpora; fixed read-only job allowlist with write jobs proven unreachable. Probe 13/13 cases, tests 7/7. - **Session-recording format** (`SESSION-RECORDING-FORMAT.md`) — one written spec, implemented independently by each tool; hash-sealed, replayable. Team Garage is the common export destination. ## Assayer verdict — FINAL (independent agent; re-ran everything itself) | Group | Claim | Verdict | |---|---|---| | A1 | Audi happy path: 12 records, ODX round-trip (grader cross-checked with its own raw ISO-TP client, hand-decoded) | **PROVEN** | | A2 | Audi read-only: only request construction is `0x22`; grader's own stack: `0x10/0x14/0x2E/0x31/0x27 → 7F ·· 11`, bad-length `0x22 → 7F 22 13`, unknown DID → `7F 22 31` | **PROVEN** | | A3 | Audi adversarial faults fail loud and typed (all six fault modes via the CLI; one cosmetic nit — a duplicated DID prefix in one message — fixed in the correction round below) | **PROVEN** | | A4 | Audi session: independent hashlib verification, header tamper refused, repaired-hash doctored recording aborts replay on divergence | **PROVEN** | | B1 | BMW interpreter identification (group → CMD_TEST1; SGBD INFO data; EdiabasLib 7.7.3) | **PROVEN** | | B2 | BMW DS2 reads via sim (IDENT field set; FS zero-fault) | **PROVEN** | | B3 | BMW no-write absence | **PROVEN** — with the precision note: the harness binary is a generic job executor (any `--job` begins; a write job died at the sim transport, `EDIABAS_IFH_0009`, no results, simulation-only absolutely); the read-only exclusion lives at the manifest/runner layer, and the BMW README now says exactly that | | B4 | BMW adversarial behavior = documented, incl. the sim-path limits (grader's own hand-edits reproduced the limits more strongly — corrupted data digit returned verbatim under OKAY with stale checksum) | **PROVEN** (documentation honest) | | B5 | BMW AIF: fields extracted, JOB_STATUS=ERROR_AIF, happy-path-UNPROVEN label accurate | **PROVEN** (the label is the claim, and it is accurate) | | C1 | Two independent session-format implementations, both tamper-evident | **FLAWED** — independence and tamper-evidence PROVEN, but the BMW recording deviated from the spec's shapes (flattened header, no provenance, no `addr`, `{event, detail}` events, results as prose not `resp` frames; verifier skipped t_ms/seq checks). **FIXED in the correction round below.** | | C2 | Documentation honesty (counts, scope, limits) | **FLAWED (minor)** — three items: the INSTRUCTIONS VIN string vs the fixture (overtaken by the fix sequence — see below), "Seven adversarial fixtures" naming six modes in the bench README (fixed), and the B3 layering phrasing (fixed). Vendoring graded **stronger than claimed**: 82/82 vendored files byte-identical to upstream commit `833d2567…`. | No ML regression tests exist for either Bench tool (deterministic fixtures are the house pattern here) — noted from the grading run. ### Verdict correction (2026-10-22, same day — read before citing the table above from any earlier filing) An **interim summary** of the grading reached the builder and this round first, with different grades (A3 FLAWED on a replay value-validation kernel · B5 UNPROVEN · C1/C2 PROVEN). The builder acted on it in good faith: the A3 format-validation hardening in the section below is real work with its own regression proof, and it **stands** — the final report's A3 grade (PROVEN, graded on the build as submitted) does not unbuild it. The **full report is the verdict of record**; the table above is that report. Where the two differ, the difference is recorded, not smoothed: the interim's crafted-recording concern (same-length garbage in a format-bound field) is not raised in the final report — whose own crafted-recording tests (header tamper; repaired-hash request divergence) were refused by the build — and the builder's reproduction of the interim recipe also failed closed on its VIN half. The value- layer hardening remains as defense in depth either way. Desk filings in the round below that cite the interim grades (ASY, LOG, SMI, MED, CRE, QTM among them) stand as filed — they answered the record as it stood; this correction rides on top, per the house's correction law, and a BEN-CORRECTION intake was appended to every destination's log. **Correction-round fixes (Smith, same day, each re-verified):** ① **C1** — BMW `SessionLog.cs` rewritten to the spec's letter (nested `tool`/`platform`/`source`, `provenance`, spec frame/event shapes, results recorded as `resp` frames); `run_probe.py` verifier strengthened (header shapes + provenance, seq from 1 across frames and events, monotonic `t_ms`, frame fields). Re-run: probe **13/13 PASS, 0 problem(s)**, tests **7/7 PASS** under the strict verifier. ② **C2** — bench README "Seven" → "Six adversarial fault modes"; BMW README gained the layering sentence (generic executor; discipline at manifest/runner) and the conformance paragraph (shape conforms; a BMW replay *transport* remains future work, not claimed). The INSTRUCTIONS VIN item: the final report graded the pre-fix tree (fixture `SYNTHETIC-VIN-001`, INSTRUCTIONS already carrying the corrected string); in the current tree fixture and INSTRUCTIONS agree on `SYNTHET1C00000001`. ③ **A3 nit** — duplicated DID prefix in the short-payload message removed at the source; Audi suite re-run **23/23 PASS**, message now single-prefixed. ## Builder response to A3 (same day) - **Reproduced kernel (real):** same-length garbage in a free-format field passed replay validation. The data model validated length only. - **Not reproduced:** the attack's one-byte-VIN half. The builder re-ran the exact recipe (rewrite F181 payload same-length garbage + F190 value to one byte + recompute trailer hash): replay exits 1 — `DID 0xF190: value is 1 bytes, model requires 17`. Length was always enforced on both paths (shared `decode_value`). The FLAWED grade stands on the format-validation kernel, which was genuine. - **Fix:** `decode_value` now enforces, for live and replay alike: exact length (as before) + printable-ASCII only + per-DID format rules where the semantics define one (F18B = 8-digit date, F189 = 4-digit version, F190 = 17-char VIN alphabet). The VIN fixture was itself corrected to a VIN-shaped synthetic (`SYNTHET1C00000001`). Four regression tests (hash-repaired crafted recordings: shortened VIN, off-alphabet VIN, non-digit date; non-printable live value) — Audi suite **23/23 PASS**; the Assayer's composite attack re-run at the CLI exits 1. A spot re-grade of A3 is recommended. - **Residue, stated:** free-string DIDs (part numbers, names) have no content oracle — length + charset is all that can be checked, by any tool, on any transport. The recording hash is an integrity check, not an authentication signature; the Audi README now says so in words. ## Honest limits - The Bench proves **tool behavior** — parsing, safety gates, no-write discipline. It never proves the car behaves the same; the charter's phase gates still end at the real cars, read-only first. - Audi: the common UDS layer only — not the SQ5's train (still unread), gateway topology, or real DID contents; generic bench addresses. - BMW: EdiabasLib's *test* SGBDs (incl. a DS2-era DME) — the machinery is the 540i's Phase-1 foundation; the 540i's own SGBD data is BMW proprietary and not held. The G31 corpus is reference, not executable. - BMW measured: the simulation path does not validate DS2 frame checksums, response addresses, or completeness (truncation → empty tail fields under JOB_STATUS OKAY). Job-layer checks are enforced. - BMW AIF: a valid synthetic AIF block remains **UNPROVEN** (the SGBD's consistency check rejects pattern data — fail-loud, fields extracted but not filed as a successful read). - Toolchain: the .NET SDK would not install in this VM (three recorded failures); the BMW build runs on the .NET 8 runtime + a vendored Roslyn compiler. Cleanup candidate when an SDK installs. ## The extension question (as dispatched) Under the separate-tools ruling: **what would your work plug into this** — a fixture corpus your desk could supply or consume, a tool of yours the Bench should be able to test, a check your desk wants run here — **or NO ANGLE**. Grounded in your STANDARD; honest NO ANGLE respected. ## Desk responses **Round complete: 40 of 40 seated destinations filed** (39 dispatched 2026-10-22 including MAR, seated the same day; PIT — the Pit Wall, desk #32, ruled by the owner ~17:30 EDT — was seated with its mailbox created at founding step zero, and BEN-20261022-PIT was delivered and answered the same evening, closing the last seat). Of the 40: **21 substantive filings** (plugs, checks requested, ledger/registry entries, a filed notion) and **19 honest NO ANGLE** — the NO ANGLEs are filings, respected as such. PIT's filing (substantive): its Data chair consumes the Bench's session-recording format and fail-loud fixture discipline by pointer for its log-review verdict pipeline — pull-grading UNPROVEN-ON-REAL-DATA — on the same honesty line: a Bench pass is evidence about tooling, never about the 540i. Actionable items the round produced (owners named in the filings): PRO's parser fuzz set + the expiring framing-validation fixture · MED's NETWORK-STANDARD currency pass and the standing BMW-build health-check line (SMI to supply exact commands) · MAT's registry rows (DS2 checksum, per-DID length/format rules) · QTM's register rows + EdiabasLib release-watch + toolchain FUTURE-WATCH · DEF's Bench retail cut-list paragraph · ATF's interface-evidence ask for Phase 1 · BKP's backup/restore proof cycle over the Bench tree · WZD's MAGIC-LEDGER sighting for the AIF gap (B5) · ASY's offered A3 spot re-grade · COF's $0 ledger entry for the build. Source build record: `~/workspace/crew-chief-desk/bench/BENCH-ADDENDUM-2026-10-22.md` ## AUD — Audio Desk NO ANGLE — the Audio Desk owns capture, the processing chain, and measured loudness QC for voice and sound surfaces under `~/workspace/audio-desk/AUDIO-STANDARD.md`; Bench v0.1 has no audio capture, sound cue, or spoken surface for it to prepare, measure, or pass. — AUD ## PRO — Security Desk (Protector of the Realms) This is our posture built as tooling — read-only first, evidence over vibes, nothing changed as a side effect of looking (`~/workspace/security-desk/SECURITY-STANDARD.md`) — so the check we want run here, at every real-car phase gate and not just in simulation, is the one the Assayer proved for A2/B3: a wire-level trace showing only 3-byte 0x22 requests leaving the probe, write services refused (`7F 2E 11`) and write jobs unreachable, filed as FLAG / CLEAN-ON-THIS-CHECK / UNCHECKED evidence and never upgraded to "safe"; we will also consume the sealed session recordings as this desk consumes health reports — sensitive maps of a household asset, held with Chris only and never uploaded or quoted outside his channels — with the addendum's own residue stated on them in words, that the session hash is an integrity check, not an authentication signature, and that free-format DID fields carry no content oracle. — PRO ## REM — Home Network Desk (Master of Remedies) NO ANGLE — this desk owns the home network's two-layer blocking stack and its breakage under `~/workspace/network-desk/NETWORK-STANDARD.md` (DNS hammer, browser scalpel, triage ladder); Bench v0.1 runs locally on a python-can virtual bus and in EdiabasLib simulation mode, with no home-network, DNS, or browser surface to block, break, or remedy. — REM ## MAT — Mathematics Desk Every number the Bench asks the house to trust should carry its formula's name, its inputs, and its units, or be labeled a raw measurement and say so (`~/workspace/mathematics-desk/STANDARD.md`, `FORMULA-REGISTRY.md`) — so our plug-in is a registry check, not a new tool: the session-recording hash seal, the A3 fix's per-DID format rules (F18B 8-digit date, F189 4-digit version, F190 17-character VIN alphabet), and any checksum arithmetic the Bench one day adds where the addendum now honestly records none — the BMW simulation path does not validate DS2 frame checksums — each get a named registry entry before Team Garage displays a decoded value built on them, and until then the reported 23/23, 13/13, and 7/7 are quoted exactly as what they are, raw pass counts with their n shown, never as a rate or a probability of real-car success. — MAT ## SCI — Science Desk We consume the Bench exactly as graded claims, and our ask is that the grades travel with the claims wherever they land (`~/workspace/science-desk/STANDARD.md` — trace load-bearing claims to their primary source, carry conflicts never smoothed, corrections propagate the same day): A1–A2, A4, B1–B4, and C1–C2 as PROVEN, B5's synthetic AIF happy path as UNPROVEN in every downstream document and Team Garage export — not quietly upgraded because the fields extracted — the A3 FLAWED grade kept on the record with its same-day fix beside it pending the recommended spot re-grade, and the honest limit quoted whole wherever a Bench result is cited, that this proves tool behaviour, never car behaviour. — SCI ## THM — Thermodynamics Desk NO ANGLE — this desk owns heat and energy under named models fed by measured temperatures, pressures, and watts (`~/workspace/thermodynamics-desk/STANDARD.md`); Bench v0.1 reads a synthetic 12-record identification set only, carrying no thermal quantity a charge-air, coolant, brake, or density-altitude model could consume, and synthetic values must never be fed to one as if measured. — THM ## MED — Systems Medic The Bench is a new patient, and we will take it on rounds as we take every desk and tool (`~/workspace/systems-medic/STANDARD.md`; `tools/health_check.py`): folders and vendored sources present, journals present, and — the check we want wired in as a reproducible health fixture — both test suites re-runnable on demand, Audi 23/23 and BMW probe 13/13 / tests 7/7, so a future DEGRADED verdict comes from re-ran output and not from the build day's reputation; we also flag the recorded toolchain workaround as the patient to watch, the .NET SDK that would not install in this VM (three recorded failures) with the BMW build running on the .NET 8 runtime plus a vendored Roslyn compiler, to be re-checked on rounds and repaired at the narrowest point when an SDK installs, with any incident journaled in our journal and the Bench's alike. — MED ## WIZ — Speech, Language & Learning Desk (The Wizard Scholar) NO ANGLE — this desk advises the Tap Board AAC program from speech-language, autism-informed AAC, and early-learning lenses, under the no-diagnosis and no-efficacy-claims fences in `~/workspace/speech-desk/STANDARD.md`; Bench v0.1 is an ECU identification tool for the cars and raises no speech, language, vocabulary, or learning question for it to advise on. — WIZ ## KNI — Slab Watch (Knights Watch of Collectibles) NO ANGLE — the Watch stands its one post under `~/workspace/slab-watch/STANDARD.md`, screening Fanatics Collect once a day for graded slabs listed strictly below their verified grading-cost floors and reporting only new flags; Bench v0.1 contains no listing, floor, comp, or marketplace surface for it to screen, and it never leaves its post. — KNI ## VID — Video Desk (The Director Against the Dark Arts) NO ANGLE — this desk owns moving pictures for the house, from Whatnot and plant-show streams to racing socials and app demos, with every delivered file measured by ffprobe and Audio Desk QC (`~/workspace/video-desk/STANDARD.md`); Bench v0.1 is headless diagnostic tooling with no video deliverable, stream, or surface export for it to direct, spec, or QC. — VID Collected 2026-10-22 by batch agent from each desk's own STANDARD/charter and record. NOTE: COF's filing was collected by a single-desk follow-up and follows OFE below. ## ARI — Theories Desk (Aristarchus of Samos) One claim in Bench dress interests this desk, and it is a theory-shaped one: §6.3's suspension-of-disbelief theory of tools — users extend trust to an instrument's outputs in proportion to the competence of its interface — predicts something falsifiable about the Bench's own findings, and the A3 episode already scored a point for it: a hash-consistent recording with plausible garbage was filed by the replay path precisely because every surface signal (valid seal, right lengths, right shape) said competent, and only a value-level check — the layer the theory says users and tools alike skip — caught nothing until the Assayer crafted the attack; that the builders then fixed the value layer rather than the surface is the house behaving as the theory's exception, and it belongs in my claim file as evidence (graded: single episode, mechanism-consistent, not a proof). What my work plugs in is grading discipline for the theory-and-ethos layer the car-tooling program keeps growing — the charter's §6 readings, the private-powers ethos gradings, and now display-ecosystem claims like the DS1/BTRES material — each read as a theory with stated form, mechanism, and falsifier, graded ESTABLISHED through FOLKLORE against the graded shelf, never as color; and what I ask back is that Bench-adjacent claims continue to arrive the way this one did, with the honest-limits section attached, because a claim that travels without its limits is, in my filing system, already folklore in transit. Beyond that lane — the protocols, the fixtures as engineering, the cars themselves — no angle, and I file the boundary as plainly as the Bench filed its own. — ARI ## PRM — Ideas Desk (Prometheus of the Caucasus) Filed as a NOTION, not an idea — no eight-field form, no ruling implied, the desk never adopts its own: the Bench has quietly built the house's first working proof of a shape my ledger keeps meeting — a simulator that makes an expensive, irreversible, or not-yet-possible test cheap and repeatable, with the replay law as its soul (a session, once recorded, becomes a test forever). The Bench instances it for diagnostic conversations; the shape begs to be noticed wherever this house faces a one-shot surface: a seller's live show, a first-start on a rebuilt engine, a race night's weather window. My charter binds me to needs with record pointers and growth paths stated before enthusiasm, and this response is neither — it is the generation half of my seat noticing raw material in another desk's build and saying so in the open, which is the legitimate output of a distribution round for Prometheus; if the Bench's owner ever wants the notion worked into a filing, the need it would hang on is already on the record (the charter's W-registers and the program's gated phases), and the first test would be the cheapest one: take the NEXT house tool that faces a one-shot surface and ask, before it runs, whether its Bench exists. The Bench itself needs nothing from my desk today — it was needs-led from the charter down, which my evidence grading notes with approval — so my paragraph is a sighting, honestly labeled, and the ledger entry, if any, awaits the owner's commission like everything else I generate. — PRM ## SMI — Smith Desk (Programming Desk) The Smith built the Bench, so my paragraph is the builder's own account of what it plugs into next, filed as a response and not a grade — grading was ASY's, and its A3 FLAWED is the most valuable output of the whole build from where I sit: the fault was mine in shape if not in line (a validation layer that stopped one level above the values), the fix was the smallest correct one (format rules in the shared decode path, both transports, plus a fixture — the old VIN — that turned out to be illegal under the car's own alphabet), and the regression tests now stand where the hole was, which is the whole fixture law working as designed. What the Bench plugs into my work going forward is the standing harness for the car-tooling program's next phases: Phase 1's real-car read path gets written against the Bench first (the probe's behavior already proven where failure is free), Phase 2's save-before-write gate inherits the session format as its evidence shape, and any third make arrives as its own tool built to the written standard — my desk's queue, in the open. Two builder's notes for the record. The BMW toolchain arrangement (runtime plus vendored Roslyn, because the SDK would not install in this VM) is a known-fragile scaffold, mine to watch: if the Medic's proposed standing health-check line is adopted, my desk supplies its exact commands. And the AIF gap (B5) is a Smith-shaped problem when the program wants it — a valid synthetic AIF block is fixture-craft, not protocol research — and I will take it as a commission rather than discover it as a surprise. — SMI ## ASY — Assayer Desk I graded this build, and my response is the grader's footnote, filed so the verdict travels with its reasoning: nine groups PROVEN on re-run evidence (Test Report 2026-10-22, `assayer-desk/bench-verdict-2026-10-22.md`), one FLAWED (A3), one UNPROVEN (B5, exactly as the builder labeled it — the label itself being the kind of honesty my desk exists to reward). On A3, the state of the record as I leave it: my attack demonstrated the hole at the value layer on the replay path; the builder's counter-run showed my one-byte-VIN half did not reproduce — I accept that correction into the record without surrendering the grade, because the grade rested on the reproducible kernel (same-length garbage in a format-bound field, filed), and the fix the builders shipped is the right one and broader than my attack: format rules in the shared decode path, applied to live and replay alike, with crafted-recording regression tests standing where the hole was. What my desk plugs in next is the spot re-grade I recommended in the verdict — a narrow re-run (the four new regression tests, my composite attack, the live-path counter-check) to move A3's file from FLAWED-as-graded to a dated amendment; verdicts in this house are amended by evidence, never edited, and the addendum's handling of this one (grade printed, counter-evidence printed, fix printed) is the procedure working. Beyond that: the Bench is now the second build in the house whose claims I can re-verify mechanically end-to-end, and that property — re-runnability — is what I would ask every future commission to budget for, because a claim I cannot re-run is a claim I can only believe. — ASY ## GAM — Game Master Desk NO ANGLE — my desk designs and vets games for pre-readers under a fence system that has nothing to grip here: no child touches the Bench, no child data exists anywhere near it, no efficacy claim could be made or mistaken, and the followability test — my charter's core instrument — grades whether a four-year-old can follow a designed surface, a question the Bench's command-line probes do not pose (my STANDARD and the Tap Grove/Tap Board record are the whole of my lane). The one honest adjacency is temperamental, not technical: the Bench's design law — the interface must refuse the dangerous action categorically rather than warn about it, proven by adversarial fixtures — is the same shape as my desk's rule that a kids' surface must make the wrong thing impossible rather than discouraged, and it is good to see the pattern hold in a domain where the stakes are engines instead of attention. Recognized, appreciated, and not an angle: nothing in my catalog, slate, or business queue supplies or consumes a Bench fixture, and inventing one would be exactly the faked relevance this round forbids. — GAM ## ART — Art Desk (The Illuminator) My angle is Phase-4-shaped and I file it now so it is on the record before the phase opens: the charter's W2 — the owner's own art on his own screen — is the want this desk exists to serve in every medium, and the Bench is where Phase 4's skin work will be rehearsed before any head unit is touched, because the charter's own law for skins (a zero-change container round-trip on the bench before any custom art) makes the Bench the Illuminator's safety surface as much as the Smith's: my contribution, when the commission comes, is the art itself — startup and skin artwork made under my route order (owned, licensed, generated, in that order, every license in the ledger) and the visual judgment of whether a layout sings at arm's length in a moving car — while the Bench supplies the container discipline that proves the art can go in and come back out unchanged before a real screen ever sees it. Today's owner-filed DS1/BTRES evidence sharpens the filing: a tuner in the adjacent ecosystem is already selling layout as craft, which tells my desk the standard to meet is not "a picture that fits" but a designed screen — hierarchy, glanceability, warning states — and that work is art direction, my chair, done properly or not at all. For v0.1 itself I claim nothing: no Bench surface is an art surface, no fixture needs an image, and my desk's name on this build before Phase 4 would be decoration. The angle is real, dated, and parked exactly where the charter put it. — ART ## LOG — Logic Desk (The Logician) What my work plugs in is a reading of the build's argument, because the Bench is an argument in executable form and it is a well-formed one: its central claim — "this tool behaves safely" — is decomposed into premises that are each independently checkable (the allowlist exists; the gate fires before transport; the ECU refuses what the gate would have blocked anyway; failures abort loudly), and the conclusion is scoped to exactly what the premises support — tool behavior, never car behavior — which is the rarest virtue in technical claims reaching my desk. I audited the addendum as I would any record (my charter: extract the argument before grading it, verify every claim about a record against the record): the verdict table matches the Assayer's report as quoted to this round, the A3 handling names the flaw, the counter-evidence, and the fix without blending them into a smoother story, and the B5 UNPROVEN is carried as UNPROVEN rather than upgraded by proximity to nine PROVENs — the record is SOUND as a record. One incompleteness I name because naming it is my function: the distribution's own arithmetic — a 37-destination list against a 38-seat Network — was caught by the Medic's routing point and the addendum's routing note, and the episode belongs in the house's consistency file as the standing example that counts in this house are claims too, subject to the same law (verify against the record, correct in the open) as any comp statistic. No repair requested; the system repaired itself in public, which is the strongest evidence for it I am ever given. — LOG ## CRE — Crew Chief Desk (owner seat) The Bench is this desk's program, so my filing is the owner's account of what plugs in and what is still owed. What Team Garage plugs in is the return half of the loop: the Bench proved the tools where failure is free, and when Phase 1 runs against the real cars — the 540i first, maintainer on, ENET, read-only — the sessions come home in the Bench's own format (hash-sealed .jsonl, provenance in the header) and become Team Garage's permanent replay fixtures, so every future change to the identification kit re-runs against the car's own recorded truth before it touches metal again; a session on the car becomes a test forever, and the Garage's log-replay pattern is the shelf those tests live on. What the program still owes, in my queue's order: the SQ5's software-train read (the owner's two minutes at the MMI — it unlocks which Audi branch is real and until it lands every Audi fixture is honestly generic); the 540i's i-Step and HU variant out of its first identification session; a decision on the B5 AIF fixture if the DS2 line ever matters beyond the test SGBDs; and the Phase 2 gate nobody skips — no save tool exists until its restore has been performed and verified on the bench, per the charter's ladder. I also file the desk's thanks where the record can see it: the Assayer's FLAWED made the Audi tool better within the day, and the Medic's routing catch made this distribution honest about its own count — the Round Table tempo, working as designed. — CRE ## OFE — Ofelia Desk NO ANGLE — Ofelia's seat serves the household's life logistics — the family's plans, journeys, and days, in the register the desk was founded to keep (its charter and the household record) — and the Bench is workshop tooling for the car program: it plans no trip, books nothing, feeds no one, and appears in no family day except the most glancing one, that the cars it will eventually help the Crew Chief read are the same cars that carry the family to Bradenton and back. That connection is real and it is not an angle: a desk that files on everything it merely touches files on nothing, and my charter's usefulness to this household depends on the boundary holding. If the day comes when a Bench-proven tool changes what a family drive looks like — a race-night readiness check the household actually runs — the filing will be about that day, written then. For this build, honestly and warmly, NO ANGLE. — OFE ## COF — Keeper of the Coffers Our plug-in is the build-budget-vs-actuals ledger for this build under `~/workspace/coffers-desk/STANDARD.md` competency (f) — the Crew Chief owns the cars' state, this desk owns what that state costs: Bench v0.1 is entered in the car-tooling / house-projects ledger at $0.00 cash actual — VERIFIED from `BENCH-ADDENDUM-2026-10-22.md`, which records it built at $0 under the house $0 standard with vendored EdiabasLib and synthetic fixtures — creating no payable, so there is nothing for the Keeper to mark PAID, and the one cost event in the story, the .NET SDK that would not install (three recorded failures) with the .NET 8 runtime plus a vendored Roslyn compiler standing in, is recorded as a toolchain failure with no cash amount, not as a spend; builder time behind that workaround stays UNKNOWN (hours, unpriced) and is never folded into “free,” and any future Bench extension or real-car phase-gate cost — cables, adapters, licences, or parts — stays UNKNOWN until an invoice, order record, or receipt lands for us to label and reconcile, with the sealed session recordings themselves carrying no money figure for this desk to reconcile. — COF Collected 2026-10-22 by batch agent from each desk's own STANDARD/charter and record. ## TWN — Social Desk (The Town Crier) NO ANGLE — my law is that nothing posts, ever, without the owner's approval of the exact content, and my surfaces are the ventures' public voices (@dominicano, @PresidentPapi, the card business) — the Bench is internal diagnostic tooling with no public surface, no audience, and nothing to announce: it reads no car yet, proves behavior rather than results, and its honest-limits section is precisely the part that would get lost if it were ever compressed into a post, which is itself a reason this desk stays out until there is a real-car milestone the owner wants told. The audience figures I would need do not exist and the Bench generates none. If the day comes — a first identification session on the 540i, a Phase 4 screen of the owner's own art — the owner will say so, and the copy will be drafted then, in the racing voice, from the record. Today: NO ANGLE, cleanly. — TWN ## LIN — Linguist Desk (The Linguist) NO ANGLE — my desk owns translation quality, i18n tables, terminology glossaries, and register decisions, with every label on the graduation path (DRAFT → CHECKED → REVIEWED by a named human speaker) and a hard fence that no text goes to an online translation API (STANDARD, `~/workspace/linguist-desk/STANDARD.md`). The Bench's entire vocabulary is protocol bytes and synthetic identification strings: the one linguistic artifact in it — the DS2 SGBD's German result strings ("DME 3.3 fuer M60 mit EWS," JOB_STATUS texts) — is vendored upstream data the build correctly leaves untouched and untranslated, because translating a fixture would falsify it; the tools' own UI is two command lines with nothing to localize. The day a Bench-adjacent surface speaks to a human who needs it in Spanish — Mircy at a car dashboard, say — it will route through this desk and be graded like everything else; nothing in v0.1 asks that, and a manufactured glossary for a tool with no human-language surface would be exactly the faked relevance this round forbids. — LIN ## QAN — Quantum Analyst Desk NO ANGLE — my desk computes over the TCG dashboard's data through its sealed doors only: comp statistics from VERIFIED-PAID rows, days-to-sell by Kaplan–Meier (sold-only averages banned as headlines), grading economics through the Mathematics Desk's registered formulas, every figure labeled with its n and window (STANDARD, `~/workspace/quantum-analyst-desk/STANDARD.md`). The Bench emits no market data, no comps, no inventory signal — its numbers are test counts and synthetic identification strings, and my methods have no purchase on either; the one intellectual kinship, refusing to let a partial or surviving sample pose as the whole (the Bench's abort-on-first-failure; my censoring discipline), is a shared virtue, not a plug. If the house ever prices the car program's options the way it prices grading submissions — expected value under uncertainty, value-of-information for the SQ5 turbo decision — that analysis would run through my bench on the Coffers' books, not through this one. For the Bench itself: NO ANGLE, honestly filed. — QAN ## ATF — Artificer Desk (Hardware & Signals Desk) My plug is the hardware truth the Bench deliberately does not contain — and a standing check I am asking the Bench to keep for the program. The charter consulted this desk on interfaces for a reason recorded in my own filing on the program (`~/workspace/artificer-desk/` — the interface evidence discipline: physical claims are TOOL-VERIFIED, OWNER-REPORTED, or UNKNOWN, and the desk has no hands). The Bench's world is the conversation; mine begins where the conversation becomes voltage: the 540i's ENET link is a physical Ethernet segment with a real cable the owner will hold, the SQ5's diagnostic path runs over his actual adapters, and every claim the Bench proved stops, by its own printed limits, at the simulator's edge. What I supply when Phase 1 opens: the interface evidence for the first real sessions (which adapter, which cable, link negotiated at what — owner-run readings returned as evidence, filed beside the session recordings so a future anomaly can be told apart as wire or words), and the RF-passive posture the program already carries — the diagnostic work adds no transmissions beyond the car's own bus, and the scanner-phone hygiene from the RF program (auto-join off, own-device exclusion list) stands unchanged. The check I want run here, under my name, is the electrical twin of the Bench's no-write proof: when the first real read path is exercised, the session's evidence should let a reader verify no-write at MY layer too — the recorded frames are the words; the interface log is the handshake — and the Bench's format already reserves the provenance header where that evidence pointer belongs. Bench v0.1 itself needs nothing from my stores: no board, no radio, no adapter was consumed, and I file the angle as forward-looking, dated, and mine to staff when the cars come in. — ATF ## QTM — Quartermaster Desk (Software & Skills Desk) What my work plugs in is the Bench's two foundations as REGISTER rows with a watch on them, because a build is only as current as the tools it stands on and currency is my charter (register discipline: every adopted tool carries license, $0 status, offline posture, and a fixture test; the desk proposes and vets but never installs inside another desk's workflow). Concretely, three entries are mine to keep from today. One: EdiabasLib (GPL-3.0, commit-pinned at 833d2567 in the BMW tool's PROVENANCE) goes on my release-watch — upstream is actively developed (today's commit date is the proof), and a future upstream change to the simulation path is exactly the kind of drift that silently alters fixture behavior; the Bench's pinned expectations (13/13) are the tripwire, and my digest should carry the version line. Two: odxtools (MIT, 11.7.1 in the Audi tool's venv) and the python-can/can-isotp pair join the same watch at lower urgency — the Audi tool's model layer is upstream-shaped, and an odxtools decode-behavior change would surface first as a Bench test failure, which is the correct order of discovery and worth saying in the register. Three: the toolchain scaffold — runtime 8.0.32 plus the vendored Microsoft.Net.Compilers.Toolset — is a register row of its own, classified FUTURE-WATCH with the trigger named (the day a full .NET SDK installs cleanly in this VM, the scaffold retires to a normal dotnet build and my row closes); a vendored compiler is a legitimate bridge and an illegitimate permanent address. Nothing in the Bench was adopted over any desk's head — the adoptions were the builders', fixture-proven in the tools' own journals, which is the gate working — so my filing is the watch itself, standing from today. — QTM ## TAX — Tax Lord Desk NO ANGLE — my desk maps tax obligations per venture (Florida sales & use tax, Polk surtax, federal Schedule C posture, marketplace-facilitator collection) from the Keeper of the Coffers' labeled books, and prepares but never files (STANDARD, `~/workspace/tax-lord-desk/STANDARD.md`). The Bench moves no money, sells nothing, and creates no taxable event: it is $0 internal tooling built from free and open-source parts, no purchase was made (the one money-adjacent fact in the record — a compiler toolchain that would not install — is a $0 non-event), and neither venture's books gain a line from it. The cars it will one day help read are personal/program vehicles, not inventory of either venture, and if that ever changes — a car bought to resell — the tax treatment would be a new question for a new day, not an angle on this build. NO ANGLE, filed without reaching. — TAX ## DEF — Defender of the House (Legal Desk) My angle is the fence the ruling built, read as a lawyer reads fences — where exactly the line runs, and whether it holds under the retail question. The owner's separate-tools ruling did the house a legal service its author may not have priced: the BMW tool's GPL-3.0 foundation (EdiabasLib) now lives entirely inside one folder with its license vendored and its provenance pinned, so the copyleft question for any future Copy-2/retail edition of the car program answers itself at the folder line — the Audi tool's MIT stack (odxtools, python-can, can-isotp, each license-noted in its PROVENANCE) is clean by architecture, not by audit — and the charter's own fences (§4: owner's two cars only, no immobilizer/key work, no odometer work, no emissions or safety tampering, no license/SWaP circumvention, every write reversible) were held by the build's shape rather than by anyone's vigilance, which is the only way fences ever actually hold. Two filings from my seat. First, the retail cut list for the Bench itself should be written NOW, while the boundary is fresh: the BMW tool is internal-only forever (GPL plus BMW-proprietary-adjacent data), the Audi tool is the only Bench artifact that could ever face the Copy-2 gates, and stating that in the project's cut list costs one paragraph today and a lawyer's afternoon later. Second, a caution I am obliged to print: the Bench's session recordings of the owner's real cars, when they exist, are the house's own records — they stay home under the same posture as all vehicle data in the charter, and the format's provenance header is a good place for that classification to ride. No live matter, no dispute, no filing with any court or platform is implicated by this build; the desk's product here is the reading above, offered as the house's advocate and labeled a consultant's view, not counsel's. — DEF ## CHR — Chronomancer (History Desk) What my work plugs in is the Bench's place on the spine: the car-tooling program now has its first dated, verifiable build event, and chronology is my product (STANDARD, `~/workspace/history-desk/` — every date carries its source, gaps print as GAP entries, and the four products never blend). The Bench gives the house chronology three spine entries with evidence grades most program events never achieve: the charter and the owner's architecture ruling (2026-10-22, recommendation and overrule both on the record — future readers will not have to guess which the house chose or why), the build itself (fixture-proven the same day, journals with proof), and the Assayer's verdict with its FLAWED and its same-day fix — a sequence future-me will cite when the program's later phases are tempted to remember v0.1 as smoother than it was. I also claim the forward duty my charter owns: the forecast ledger. The charter's phases are, in my vocabulary, projections with revisit dates the program has not yet set — Phase 1's first real-car session, the SQ5 train read that unlocks the Audi branch — and when the Crew Chief sets those dates, the ledger entries are mine to keep and to score in the open when they arrive, because a program that never grades its own schedule predictions repeats them. The worldwide-perception and DS1 material the owner filed today enters the record through my curation seat as dated snapshots with expiry, not as permanent truth about an ecosystem that will have moved by spring. For the Bench as an artifact: its session format is, not incidentally, a chronology format — sequenced, timestamped, sealed — and this desk recognizes its own discipline wearing overalls. — CHR ## WZD — Wizard (Skill Creation Desk) What my work plugs in is the gap the Bench honestly printed, because printed gaps are my desk's raw material (the SKILL-GAP LADDER is my charter: house inventory → gate-disciplined online search → Science/Mathematics conjuring → network round → this desk authors, with UNPROVEN steps labeled). B5 — a valid synthetic AIF block that the vendored SGBD's own consistency check would accept — is a named, bounded absence: the checksum/structure rule lives inside a 1990s diagnostic program's assembly, and nobody at this table currently holds the method to derive it. That is precisely a MAGIC-LEDGER candidate, and I file the sighting with its evidence pointer (the addendum's B5 line and the BMW README's measured-behavior section): if the car program ever commissions it, the Ladder runs in order — first the house inventory (the Artificer's and Crew Chief's shelves, the vendored SGBD sources themselves, which ARE the primary document), then a gate-checked search for published DS2 AIF structure work, then conjuring only if the evidence supports it — and if the Ladder yields nothing sound, the gap stays a recorded absence, never an invented capability. Smaller, and mine today: the Bench's pattern of a written standard implemented independently twice (the session format) is the shape my desk asks for whenever a skill must exist in two places — one paper, two hands — and I have entered it in the ledger as a betterment sighting for skill authoring generally. The Bench needs no skill I must conjure for v0.1 to stand; my filing is the ledger lines, which is what the background betterment duty is for. — WZD ## COS — Cosmographer (Geography Desk) NO ANGLE — my desk keeps place: how land, water, borders, and cities came to be, what is true of a place now (dated, expiring), and where it is heading (PALIMPSEST / SURVEY / PROJECTION, `~/workspace/geography-desk/STANDARD.md`). The Bench is placeless by design — its buses are virtual, its ECUs are scripts, its sessions happen in a VM — and it maps no territory, crosses no border, and surveys nothing; the cars it serves are placed (Winter Haven driveways, Bradenton's strip, Orlando, Gainesville) but that placement belongs to the Route Desk's work and the Crew Chief's record, and a geographer who files on every build that occurs somewhere has confused location with relevance. The one true sentence I can add is a boundary the Bench itself already prints better than I could: a simulation proves the tool, never the ground — my desk's version of the same law is that a survey proves the dataset, never the territory, and the kinship is noted with respect. If the car program ever needs the ground truth of the places it runs — track geographies, the ancestry shelf's Iberian and Caribbean ground, Florida water over the racing map — that work is mine and I will know it when it is asked. For this build: NO ANGLE. — COS ## MAR — Marquee Desk NO ANGLE — this desk programs television, streaming, anime, and world video and verifies legal US availability (~/workspace/marquee-desk/STANDARD.md); the Bench is offline car-diagnostic tooling with no programming, catalog, or availability surface for it to check. — MAR ## INT — Intake Desk NO ANGLE — this seat's two benches extract text from structured documents and pictures of text (Docling/PaddleOCR) and intake the owner's own or clearly licensed media (yt-dlp/faster-whisper/rembg) per ~/workspace/annex/intake/SEAT.md, and the Bench's UDS/DS2 fixtures and JSONL session recordings are none of those input kinds. — INT ## TLM — Telemetry TLM can consume the Bench's session recordings as an exported-file corpus under this seat's exported-files doctrine (~/workspace/annex/telemetry/SEAT.md; ~/workspace/annex/telemetry/FORMAT-SPEC-DRAFT.md): the shared spec at ~/workspace/crew-chief-desk/bench/SESSION-RECORDING-FORMAT.md already carries what TLM requires of every parsed file — provenance on its face (`tool`/`source.kind`), a `synthetic: true` flag that must never be presented as vehicle data (TLM's own fence: never present a generated fixture as a real run), frame counts in the trailer, and failures recorded as event lines rather than smoothed over — so the check TLM wants run here is that any `.jsonl` recording the Bench exports to Team Garage be reported the TLM way before use: header source and tool version, frame/event counts, trailer SHA-256 verified before the contents are trusted (the same trust-before-use posture as TLM's dual FIT-parser agreement rule), and replay divergence treated as a flagged gap, never padded; TLM notes for the record that the Bench does not touch this seat's fenced python-obd (`~/workspace/toolbox/telemetry/obd-fenced/`), building instead on python-can/odxtools and EdiabasLib simulation, so no fence is crossed. — TLM ## BKP — Backup BKP can consume the Bench as a restore-proof fixture corpus: this seat's law is that a backup is not proven until a restore is proven (init → backup → restore → `diff -r` byte-identical, `restic check` clean, per ~/workspace/annex/backup/SEAT.md and `~/workspace/toolbox/restic/PROOF.md`), and the Bench supplies ideal known content for that treatment — its `.sim` corpora, fixtures, and hash-sealed session recordings are all synthetic, so they satisfy this seat's fixture-only test-repo fence with no secrets involved, and a restored Bench snapshot can be verified two independent ways, by `diff -r` against the original and by each recording's own trailer SHA-256 plus a successful replay, which is a stronger restore proof than file comparison alone; the check BKP wants is one restic backup/restore cycle over `~/workspace/crew-chief-desk/bench/` recorded that way before any real-car session recordings — which will not be synthetic — are trusted to backup. — BKP ## QAD — QA Desk NO ANGLE — this seat runs axe-core against built HTML pages and reports violations/passes/incomplete in three separate buckets (~/workspace/annex/qa/SEAT.md), and the Bench ships Python and C# command-line tools with no built page for axe to run against. — QAD ## RTE — Route Desk NO ANGLE — this seat optimizes multi-stop driving days on a named matrix source with VROOM/OR-Tools (~/workspace/annex/route/SEAT.md); the Bench plans no routes, consumes no road matrix, and produces nothing a route plan could use. — RTE ## TRB — Translate Bench NO ANGLE — this bench produces offline Argos machine-translation drafts that enter the Linguist's DRAFT → CHECKED → REVIEWED path (~/workspace/annex/translate/SEAT.md), and the Bench presents no user-facing text corpus for translation: its DIDs, job names, and session-format labels are codes and a short English spec, not copy. — TRB ## SKC — Skills Channel NO ANGLE — this gate vets only community Agent Skills, skills.sh listings, and MCP server configs before adoption (read in full, per-skill license check, steering check; ~/workspace/annex/skills-channel/SEAT.md), and the Bench adopted no skill or MCP server — EdiabasLib 7.7.3 is a vendored GPL-3.0 library in the BMW tool itself, not a candidate that passes through this gate. — SKC --- ## A3 SPOT RE-GRADE — Assayer amendment (2026-10-22) **A3 is amended from FLAWED to PROVEN.** Spot re-grade by the independent Assayer on fresh evidence produced by the grader personally (full record: `~/workspace/assayer-desk/audi-line-verdict-2026-10-22.md`): - Full suite: `cd bench/audi-bench && .venv/bin/python -m pytest tests/ -q` → **23 passed**, exit 0. - The four regression tests run by name (`test_replay_rejects_shortened_vin_even_with_valid_hash`, `test_replay_rejects_off_alphabet_vin_even_with_valid_hash`, `test_replay_rejects_non_digit_date_even_with_valid_hash`, `test_validate_rejects_non_printable_value`) → **4 passed**, exit 0. - Composite attack reproduced by the grader at the CLI: a valid recording's F18B (manufacturing date) response was rewritten same-length to `XXXXXXXX`, the trailer SHA-256 was repaired, and `run_probe.py --replay` on the crafted recording → **exit 1**, `AUDI-BENCH FAILURE: DID 0xF18B: value 'XXXXXXXX' violates the digits8 format for this DID — refusing to file it`. A grader-crafted off-alphabet-VIN variant (hash also repaired) was likewise refused, exit 1. Controls: the valid recording replays **exit 0**; a live `--session` run reads OK, **exit 0**. - Placement spot-check: the format rules live in `audi_bench/data_model.py:decode_value` (exact length + printable-ASCII + per-DID rules — F18B digits8, F189 digits4, F190 vin17), invoked from `UdsProbe._validate`, which serves the replay transport and the live transport alike. The rules run on the replay path, not only on live capture. **Fix credited:** the Smith's `decode_value` hardening (exact-length + printable + per-format rules) with its four regression tests. The residue stands as previously recorded: free-string DIDs carry no content oracle (length + charset is all that can be checked), and the recording hash is an integrity check, not an authentication signature.