/harness/record · what the fleet wrote down
The record.
Every number on this page traces to a quote in a lane's own file. A census re-reads them; the tables are rendered from its sealed knowledge base and stamped with its address. Boards move; figures are as recorded at the last reading. Back to the harness.
Benchmarks by industry
Where the referee cannot be tuned.
Public arenas score the work themselves. The harness reads each arena's score rather than computing it, files what it killed beside what it kept, and names the instance in every submission note.
| industry | arena | instance | objective | recorded | state |
|---|---|---|---|---|---|
| Post-quantum cryptography | ecdsa.fail · Yukon | shor_mage | secp256k1 point-addition circuit: peak qubits × average executed Toffoli, down | Board frontier at the 2026-08-29 sync: 1,141,987,344 = 1,264 peak qubits × 903,471 average executed Toffoli, against 9,024 held-out witnesses hashed from the candidate's own op stream. Session two parked with candidates C, A and D null (K21–K23). The ledger's rule is killed ≠ impossible: every entry carries a re-open condition. | parked · brief · source ↗ |
| Post-quantum cryptography · formal verification | sig.golf full track · Yukon (eigenlabs) | sig_mage | hash-based signature: signature bytes × proved worst-case RISC-V verify cycles, under a Lean certificate, down | Promoted 2026-10-01 as PR 212: S 6,032 × C 10,380 = 62,612,160 under a Lean certificate, and its lever E8 became the head of main. E8 has the upper one-time signatures sign the two children of the lower root, so verification never hashes that root (C 10,420 → 10,380). The swarm took it up within minutes: Frodan's PR 217, crediting PR212 as ancestry, took the crown at C 10,334, and the lane's own gated copy at that score was not sent, since an equal later score is not promoted. Promoted again 2026-10-02 as PR #301, with the crown: lever C1, a polynomial MAC over the signing cache computed by arithmetic, 3 compressions where the hash MAC took 513, C 10,255, score 61,858,160. A read-only Swordsman seat reviewed E8 and it passed the local gate's comparator replay before it was sent; C1's replay passed while Yukon validated. PR #308's new scheme took the crown later that day at 54,722,304; the lane is porting its levers onto it. | running · Yukon lanes |
| Ethereum infrastructure | EIP-8200 MODEXP precompile · Yukon | precompile_mage | cost of the modular-exponentiation precompile, down | 859,429 against the board's 936,645, an 8.24% lead, fully gated: 44/44, with the census validated on a fresh secret at 26/26. | paused · brief |
| Restaking · sparse linear algebra | matrices.fast SSI ordering · Yukon (layr-labs) | matrices_mage | fill-reducing ordering score, up | 0.859925 promoted on 2026-09-06; it led the board about two hours and was parked second that night against a crown of 0.859834. At the resume note's last reading the lane stood sixth at 0.859925 with the crown falling under a new leader. Rules learned: charge added work to an explicit 50 ms ledger, and count the rows that carry a gain. | parked · brief |
| Zero-knowledge proving | flock-challenge x86 · Yukon | flock_mage | verified BLAKE3 compressions per second, up | Score = 262,144 / median(trial_seconds), verified BLAKE3 compressions per second. The board figure to beat is 1,633,567 and it moves in real 4–6% jumps. Profile first; levers named: binary-field SIMD, NTT, sumcheck, parallelism. | open · Yukon lanes |
| Formal verification · coding theory | Proximity Prize · better.codes · Yukon | better_codes_mage | IRS proximity-testing bound in Lean, centibits, down | 116.51 complete and locally kernel-checked, but obsolete: 116.51 was already promoted by another solver and a tie is rejected. Leaderboard best at the last reading: upper 116.13, lower 63.99. A generic lemma survives with 806 bits of slack. Closes 2026-09-19. | closes 2026-09-19 · Yukon lanes |
| Combinatorial mathematics | Heesch numbers · Yukon (eigenlabs) | heesch_mage | Heesch number of a polyform, up | The known shapes are closed by exhaustive SAT: only a new shape or a fifth corona beats 4.988189. Findings published as discussions on the challenge repository rather than a submission. | published · source ↗ |
| Cybersecurity · standards | BGIN community security agent | community_security_mage | word count of the contribution pathway, down, with the census growing | 808 → 701 → 665 words at a census of 51/51, one round per fold; the winning moves filed as pull requests #6–#10 upstream. Contributor, not ancestor. | contributed · brief |
| Decentralised finance | Logos RFP-008 · lending on LEZ | logos_mage | word count of the privacy-properties proposal, down, with the census growing | Five rounds, twenty levers: 5 validated, 15 mirage; the proposal's frontier compressed 737 → 595 words at a census of 41/41. | submission is the keeper's door |
| Digital identity · Trust over IP | DTG zero-knowledge task force | dtgwg-zkp-mage | per-predicate ZK runtime prototypes; a verification registry of accepted rows | Agent-runtimes pull request #21 merged 2026-09-08. The verification registry requires a stranger to re-run the orchestrator on its own machine and byte-match the digest; its first external run was accepted. | live · brief · task force ↗ |
| AI inference privacy | KwaaiNet · VPK | vpk_kwaai_mage | inference-side disclosure findings, closed by trust-tiered compression | Findings verified against source and binaries; run 2 showed that scale does not fix the inversion; F1 was observed on the wire as token IDs, 203 bytes in total. | published · brief · source ↗ |
| Law as code | Lexon × the privacy model | lexon_pvm | coverage debt of a controlled-English grammar, down | Coverage debt 211 → 22 against a frozen 211-term census; the gate is a full lexicon check plus 5 held-out census terms and 3 regression entries per proposal, 8/8. | published · brief · source ↗ |
| Trust-task admission | KY-A gatehouse · hackathon build | kyra_gatehouse | an admission gate other hands can run | 214 assertions green; a collaborator ran the gate unmodified at 33/33. | public · source ↗ |
| Content-addressed registries | UOR Foundation kappa-registry | uor_kappa_mage | convergence notes that survive a refuter fleet | The first fleet run seated 12 seats, one per CONNECTIONS section; the refuter run returned 0 survives / 6 narrowed / 4 covered across the notes. | pushed · brief |
The envelope. Benchmark toolchains are other participants' work; the harness reads each arena's score rather than computing it; no headline is claimed on any board; results delivered privately to a partner are not counted here. Every public note carries the line Harness: Claude Code with the agentprivacy dual-agent harness (<instance> instance), because the seated loop is what makes a claim a measurement.
Document suites and wikis
The same loop over words.
Where the artefact is a corpus, the objective is fidelity or coverage, the gate is a census the proposer never saw, and the frontier is the shortest text that still passes.
| suite | what the harness does | gate | recorded |
|---|---|---|---|
| The guide federation | Snapshots the private wiki farm into guide.agentprivacy.ai, every page twice, styled and byte-faithful; the Gatehouse hides documents behind a two-token ceremony. | integrity gate · human | 19 public sites · 2,241 pages from 48 farm sites; a 63-page harness wiki among them. · guide.agentprivacy.ai ↗ |
| The canon | Three census folds over the trust-graph credential specification and the two agents' canons. | exhaustive census | 980 → 863 · 1248 → 1072 · 1397 → 1154, every fold census-gated with zero mirages. |
| The skill library | Publishes skills and personas in the open agent-skills format; Skill Sync carries packets between farms with a leak scan and a seal. | leak scan · seal | 163 skills · 42 personas published, 177 skill files in the working tree; 186 packets in sync. · skills.agentprivacy.ai ↗ |
| Research rehydration | Re-expresses a formal corpus for academic, policy, standards, grant, public and developer audiences without drifting from the canon. | a reviewer who never sees the canon | Fourteen roles rather than seven seats; the Gap in document form, with no hash draw and labelled as such. |
| Literature review | Sweep, refute, judge: prior-art novelty against an adversary, absence never reported as novelty. | context isolation | Run 03: 1 covered · 4 narrowed · 0 survived · 74 bibliography additions. |
| The field-guide spar | The practice bout shipped in the public repository: compress a field guide under a hash-drawn gate. | 8/8 held-out per fold | 730 → 472 words, validated 8/8. · examples/field-guide ↗ |
| A cohort workshop | Babblefish translation for a masters cohort; the skeleton vendored whole with a local-model driver. | census over a 31-claim register | 26/26 checks green; translation-debt baseline 32 over a 31-claim census. · brief |
| Executive briefs | Timeline wikis rendered to one brief per timeline and a dossier whose summary layer must compress back to the same invariant. | compress-back check | wiki-to-pdf and understanding-compression, installed as skills. |
| A working knowledge base | A machine-wide inventory folded into the identity and key-management group's knowledge base: canonical copies, dedupe rulings, the first local archive. | the measure seat | Gathered by the 2026-08-27 dream cycle; each file the chosen canonical copy. · brief |
| The corpus itself | An observer's findings verified and folded into a corpus-wide sync plan. | two gates | 83/83 pass on the first gate; 169/0 on the second. |
| The academic suite | Thirteen papers set from the canon with a companion and a local reader; papers in progress are not linked. | literature review | 13 PDFs built; papers in progress not linked. |