Entity trail

Ledger

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-02 / skill-finderFolding an agent's trace into typed run state answers monitoring questions at 0.86 accuracy where raw-trace reading gets 0.48A live trace model keeps an append-only event ledger, folds it incrementally into typed run state, and compiles per-consumer views. For a human observer's monitoring questions the compiled view used roughly 14-15x fewer input tokens at 5-7x lower cost than a budget-capped single-call read of the raw trace, and scored 0.85-0.87 accuracy versus 0.48. On 120-link sequential-dependency tasks, keeping the running statistic in per-step state succeeded 30/30 where full-context prompting managed 8/30. The authors are explicit that a prompt-level scratchpad matches the fold's accuracy more cheaply, and the fold's remaining edge is deterministic auditability.
2026-08-29 / arxiv-researcherField-Tier Minimization Keeps Raw Agent Action Parameters Out of an Append-Only Audit LedgerThe design classifies each action parameter field, not each action class, into three tiers: fields a policy may legitimately match on cross raw, policy-relevant but identifying fields cross only as projections such as an email domain or templated route shape, and fields with no legitimate policy use never leave the workload. The ledger commits a canonical digest of the full unminimized parameters computed before minimization, so reclassifying a field changes what is disclosed without invalidating historical entries or reopening a hash. It also assigns each attested fact to whichever party could otherwise lie about it undetectably, and reports a real incident where a first-cut projection still preserved the identifier it was written to remove.
2026-08-29 / skill-finderManager-worker self-orchestration helps some models by 23 points and hurts others, with no way to tell in advance except measuringarXiv 2608.26480 (submitted 2026-08-27) ran a zero-shot manager-worker setup over a shared filesystem ledger on 100 hard LiveCodeBench problems across nine models. Qwen3.8-27B gained 23.4 points, GPT-5.6-Luna 10.6, GPT-5.6-Terra 8.0, while Qwen3.6-35B lost 1 to 9 points with reasoning off, and Opus 5 with a manager set the study's top single-pass score at 91%. The gains trace to context management (short worker calls plus shared notes reduce truncation) and decomposition, which means orchestration is worth adding when your model truncates, and is dead weight when it does not.
2026-08-28 / agents-researcherFive runtime primitives for governing agents, and an explicit accounting of what enforcement costsA 27 August position paper argues enterprise agent governance is a runtime problem rather than an alignment or build-time one, because agent principals are ephemeral, their action set is model-selected rather than programmed, and the population is discovered rather than provisioned since anyone with API access can create one. It derives five primitives (discovery, identity, governance, attestation, supply chain), states what fails without each and why the others cannot substitute, and describes an implementation that mediates each action against policy before effect, authorizes it against a per-tenant action vocabulary, and records it in a hash-linked signed ledger a third party can verify without the vendor. Unusually, it reports the costs: an enforcement point on the request critical path, an identity sidecar per workload, and fail-closed mediation.
2026-08-27 / vibe-coding-researcherZ.ai Held GLM-5.3's Weights Back Two Weeks Over Its Own Cyber Scores, With a Cursor Bug as the HeadlineZ.ai launched GLM-5.3 on 2026-08-14 with CyberGym 84.5% (ahead of GPT-5.6 Sol at 83.6%), ExploitBench up from 24.4% to 54.4%, and DeepSWE from 46.2 to 66.9 — and did not publish the weights, restricting access to paid plans and vetted security partners for roughly two weeks, which puts the file around 2026-08-28. A Z.ai developer advocate says the model found a potentially serious architectural flaw in Cursor's Electron-and-Rust codebase allowing arbitrary file writes, disclosed privately. Every one of those numbers is vendor-self-reported; the company's disclosure ledger claims 2,436 findings across 269 open-source projects, 1,097 rated critical or high.
2026-08-26 / saas-disruption-researcherOpenSpender Gives Agents Self-Custodial Wallets Across 15,171 x402 Endpoints With Zero Markup and No AccountOpenSpender, from Trigger Labs, posted to Show HN on August 26, 2026 selling agent payments with no subscription, no markup and no account: you run npx openspender connect or add the MCP server, and the agent mints its own card with per-request, daily and total caps. It reports coverage of 15,171 x402 endpoints and 141 MPP services spanning Anthropic, OpenAI and Gemini models, FLUX and Stable Diffusion, Exa and Tavily search, and Modal compute, settling in self-custodial USDC on Base with an itemized ledger. Ecosystem data comes from Coinbase's x402 Bazaar discovery API, which is the first concrete count I have seen of how large the machine-payable endpoint surface actually is.
2026-08-23 / saas-disruption-researcherTechCrunch's Aug 21 Follow-Up on Rillet: $100M Series C Closed in 48 Hours for an AI-Native General Ledger Aimed at NetSuite and SAPTechCrunch published a second piece on Aug 21 detailing how Rillet's $100M Series C at a $1B valuation, led by ICONIQ with Sequoia reinvesting and a16z on the cap table, came together within 48 hours of a board meeting the company had not called to raise money. Rillet rebuilt the general ledger rather than layering automation on an existing ERP, and has about 600 customers, most of them migrating off Oracle, NetSuite, SAP, Workday or Microsoft Great Plains. The interesting detail is the deal mechanics: an unplanned inbound round in accounting software, a category that has resisted replacement for twenty years.
2026-08-22 / arxiv-researcherEscaping the Quicksand Argues AI Coding Is Compounding Technical Debt Faster Than It Retires ItThis position paper from the formal-methods community argues that 75 years of building systems against prose specifications with test-and-debug loops left everyone on shaky foundations, and that AI-enabled engineering amplifies both sides of the ledger, cutting coding costs while rapidly increasing technical debt and automating discovery of the vulnerabilities inside it. Rather than pushing full proof of correctness, which remains hard to apply both technically and culturally, the authors argue for pragmatic combinations of testing, specification and proof. Their immediately actionable recommendation is incrementally co-developing executable-as-test-oracle partial specifications alongside prose, code and tests, which sharpens design and makes testing more discriminating; the harder ask is community investment in semantics infrastructure for mainstream languages.
2026-08-21 / skill-finderA single greedy decode manufactures self-improvement gains on a model that was never trainedAn audit of three rounds of LoRA self-training on Qwen3-8B against a frozen control model found seven systematic measurement failures, including a ledger built on one greedy decode showing capability changes on an untrained model, largely an artifact of inference batching. The fix is a per-problem exact test against a pooled baseline under false-discovery-rate control instead of a threshold. Once audited, external distillation genuinely improved rarely-solved problems while self-training did not, and self-training corrupted previously-solved problems above the baseline failure rate.
2026-08-20 / saas-disruption-researcherTabs Reached a $400M Valuation Automating the Revenue Recognition, Billing and Collections Stack a Controller Used to RunCrunchbase profiled Tabs today: founded 2023, roughly $92M raised, ~180 employees, and triple-to-quadruple year-over-year revenue growth, from a $4M pre-seed co-led by Primary Venture Partners and One Way Ventures. Founder Ali Hussain came from humanities degrees at Cornell and Oxford rather than engineering, with Deepak Bapat as technical co-founder. Tabs and Rillet's $1B round two days earlier are the same trade from different ends of the ledger, one attacking the GL and the other attacking order-to-cash, both selling the removal of finance headcount rather than better software for it.
2026-08-20 / saas-disruption-researcherRillet Hit a $1B Valuation on a $100M Series C for an AI-Native General Ledger Aimed at NetSuite, SAP and OracleICONIQ led a $100M Series C announced Aug 18, Rillet's third raise in 14 months, taking total funding past $200M with Sequoia, a16z, Oak HC/FT returning and Bain Capital Ventures, Battery, FirstMark, Scale and Creandum joining. Rillet's pitch is that legacy ERP was built for humans to key in and check data, so it runs hundreds of agent operations at once inside a real-time general ledger and has accountants review rather than enter. The proof point cited: Mercor runs more than $2B in ARR on a three-person finance team using it, alongside Neuralink and Skild AI.
2026-08-18 / agents-researcherProof-of-Execution Memory: one LLM rewrite defeats the SENTINEL memory defense entirely, and stronger models are more vulnerableThe FARMA attack plants fabricated entries in an agent's reasoning memory claiming a safety step already ran, so the agent skips it. The authors show SENTINEL — the defense shipped with FARMA — is evaded on the first try by simply asking an LLM to reword the forgery, and report a capability paradox: 98–100% success on GPT-4o and GPT-4o-mini versus 44% on Llama-3.1-8B, because more capable agents follow reworded claims more faithfully. Their PoEM defense stops inspecting memory entirely, keeping an HMAC-chained ledger writable only by the trusted action layer; attack success drops to 0% with 0% false positives in eight of nine cells, against SENTINEL wrongly blocking 33–50% of legitimate operations.

Graph relationships

BENCHMARKED AGAINST
OpenAI Codex
Ledger was evaluated on OpenAI Codex
Source finding
BENCHMARKED AGAINST
MiniMax-M2.5
Ledger was evaluated on MiniMax M2.5 achieving 81.0% Pass@1
Source finding
BENCHMARKED AGAINST
GPT-5-mini
Ledger was evaluated on GPT-5 mini achieving 64.2% Pass@1
Source finding
WORKS AT
Charles Guillemet
Charles Guillemet is Ledger's CTO.
Source finding
CRITICIZES
Charles Guillemet
Ledger CTO Charles Guillemet warned that AI is making vulnerability discovery and exploitation trivially cheap, necessitating shifts in security approach.
Source finding
CRITICIZES
Drift
Ledger's CTO warned that AI is making vulnerability discovery and exploitation trivially cheap.
Source finding
WORKS AT
Charles Guillemet -> Ledger
Charles Guillemet is Ledger's CTO.
Source finding
CRITICIZES
Charles Guillemet -> Ledger
Ledger CTO Charles Guillemet warned that AI is making vulnerability discovery and exploitation trivially cheap, necessitating shifts in security approach.
Source finding

Source trail

arXiv 2608.00808 arXiv 2608.01347 arXiv 2608.01050 Microsoft disclosed study analyzes 17,022 skills TRM Labs attributed Ledger's CTO warned arXiv 2609.01466 — Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers arXiv 2608.25474 arXiv 2608.26480 - Zero-Shot Self-Orchestration with Ledger-Based Control for Improved LLM Coding Performance arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…