Your AI agents write the code. I build what proves it can ship.

Around every change an agent makes, three open-source tools I wrote. Each one publishes what it demonstrates, and what it does not.

Solid ink: demonstratedOutline: not established Sources checked on 29 September 2026 · claims register (in French)

1. 1 · Before: the decision

   Latent Compass

   A flight recorder for an agent's decisions: the options, the evidence and the choice, frozen before the outcome is known. A compass can describe direction without taking the helm.

   Decision

   Later

   frozen

   outcome

   Demonstrated

   Strict episode contracts, append-only stores where a silent edit is detectable, a replayable offline benchmark. Source: README latent-compass @f66dce1 (new tab)

   Not established

   A measured gain on a live agent. The independent audit keeps the proof status at PROOF_WEAK/BLOCK. Source: README latent-compass @f66dce1 (new tab)

   The first research phase ended in KILL_DISCOVERY: not enough evidence to train a model, so none was trained. The refusal is part of the result. Source: README latent-compass @f66dce1(new tab)

   Source and proof status(new tab)

2. 2 · During: the change

   semctx

   Maps every diff to the symbols, contracts, invariants and tests it touches, and returns PASS, WARN or BLOCK before the merge.

   +−

   • symbols
   • contracts
   • invariants
   • tests

   PASSWARNBLOCK

   Demonstrated

   Deterministic verdicts: the demonstration in chapter 01 replays. Sources: README semctx @11465f6 (new tab) · hoklims.github.io/semctx/demo (new tab) · evidence.json semctx 0.3.5 (new tab)

   Not established

   Runtime behaviour: the analysis is static, and its silence is not a guarantee. Sources: README semctx @11465f6 (new tab) · hoklims.github.io/semctx/demo (new tab)

   Source code(new tab)

3. 3 · After: the test

   AssertLedger

   Checks that a regression test really catches the bug: the same test runs on the fixed code, on the known fault and on a neutral control.

   Test bench

   The same test, three versions of the code. Pick the one that runs:

   Version of isEven under test

   Fixed code Known fault Neutral control

   src/is-even.js · the correct version

   export function isEven(value) {
     return value % 2 === 0;
   }

   Test	Result
   “two is even” assert.equal(isEven(2), true)	pass
   “returns a boolean” assert.equal(typeof isEven(2), "boolean")	pass

   On the fixed code, both tests pass: nothing tells them apart yet.

   src/is-even.js · parity inverted

   export function isEven(value) {
     return value % 2 === 1;
   }

   Test	Result
   “two is even” assert.equal(isEven(2), true)	assertion failure
   “returns a boolean” assert.equal(typeof isEven(2), "boolean")	pass

   On the known fault, “two is even” fails: it sees the bug. “Returns a boolean” still passes: it does not.

   src/is-even.js · same result, written differently

   export function isEven(value) {
     return (value & 1) === 0;
   }

   Test	Result
   “two is even” assert.equal(isEven(2), true)	pass
   “returns a boolean” assert.equal(typeof isEven(2), "boolean")	pass

   On an equivalent rewrite, both pass: the earlier failure comes from the fault, not from a brittle test.

   Verdict across the three versions

   WEAK_ORACLE

   “Returns a boolean” passes everywhere, fault included: it does not detect this fault.

   Selected

   “Two is even” fails on the fault only: the decision selects it.

   Demonstrated

   A verdict, the observations behind it and evidence that replays without a model or rerunning the tests; version 1.1.1 on npm. Sources: README assertledger @b9407bf (new tab) · npm assertledger 1.1.1 (new tab)

   Not established

   Sandboxed execution (trusted code only), test frameworks other than node:test (AssertLedger 1.3 also lists bun:test, not replayed here), general program correctness. Source: README assertledger @b9407bf (new tab)

   Source code(new tab)

In the same line

• StacksFinder — deterministic stack recommendations: the score decides, the LLM writes; served to agents over MCP (case study in French).
• temper(new tab) — local, human-reviewed skill distillation for coding agents.
• codex-wayfinder(new tab) — Codex-led governance for manual execution handoffs to Claude Code.
• top-down-diagnosis(new tab) — top-down diagnosis for coding agents, from intent to implementation; bilingual guide and templates.

The same setup on your repository starts with the Agents Diagnostic: five days, fixed price.

• Entry point

  Agents Diagnostic

  Five days to measure your review, audit your agent setup and put a first gate in your CI.

  €3,500 excl. VAT launch price · 3 places

• Installation sprint

  Three to four weeks: agent context, review pipeline, CI gates, team training.

  from €15,000 quoted after the diagnostic

• Fractional AI tech lead

  One to two days a week to hold the gates and bring the team up to speed.

  from €4,000 / month monthly commitment

Prices are shown so you know what to expect before writing. The sprint is quoted once the diagnostic is done. Offer details and terms (in French).