Skip to content
Trazum

1.61.2

An input nobody had tried

Seven defects, one stress session, one shape: an input nobody had tried, taken quietly. Chapters one to five of the 1.62 arc, landed in the order the plan committed to.

Fixed#

A seventh defect, found while proving the mask property could fail. The chapter that holds code blocks, inline code and URLs to byte-for-byte survival started with a zero that proved nothing — the corpus held no code a rule wanted to touch — so bait atoms went in: a verbose phrase and double spaces inside protected spans. They came out rewritten, with every mask believed to be on. On span shapes, the segmenter scanned every pattern over the whole text and dropped overlaps afterwards, which loses more than the overlap: inline-code matched from the third backtick of a closing fence to the real span's opening backtick — an illegitimate match, later dropped — but its scan position had already advanced past the real span, so the legitimate match was never seen and the span was left mutable.

Each pattern now scans with the earlier patterns' ranges reserved, and a match that overlaps a reservation restarts after the reservation rather than after itself. The mask property — every protected span survives optimize byte-for-byte, over the whole hostile corpus — is chapter four of the 1.62 arc, and it holds now, with the bait atoms in the corpus so the zero can never go vacuous again.

Added#

The plan through 1.62 and 1.63 — docs/plan-1.62-1.63.md, written before the code and out of a stress session that found six defects in an afternoon, all one shape: an input nobody had tried, taken quietly. The 1.62 arc holds the product to its own standard under hostile input; the 1.63 arc turns today's good performance numbers into measurements the build is held to. Its first three chapters land below.

packages/core/test/hostile-input.test.js — the stress session as a fixture. A seeded, deterministic fuzzer (same seed, same verdict, any machine; ~5s, bounded) over a corpus of hostile atoms — RTL, CJK, lone surrogates, zero-width characters, control bytes, CRLF, unclosed fences, 3KB tokens — holding optimize to three of its own promises: never throws, never grows tokens, idempotent. Plus a malformed-log corpus for profileUsage (negative, non-finite, string and __proto__ token counts) and the property that no input can produce negative money. Every defect below is pinned as a named case outside the fuzzer's seed schedule, because a seed schedule rotates the moment an atom is added.

Fixed#

optimize was not idempotent: run on its own output, it saved more. On 1 input in 4,000, emphasis stripped IMPORTANT: and left two lines equal but for a space — and whitespace, which would have collapsed it, had already run, so duplicate-lines never saw the pair. The writer's acceptance test, failed by the tool that grades it. The pipeline now runs to a fixed point (bounded at a named constant; pass two exists for the cascades, pass three confirms), rule hits and savings aggregate across passes rather than appearing as duplicate rows, and the property optimize(optimize(x)) === optimize(x) is enforced over the whole corpus. Cost: ~40% on a 1MB prompt, the price of the confirming pass.

spend_guard said yes to a lie. outputTokens: -500 priced the call at −$0.0075 and the verdict came back yes — a negative estimate lowers the projected spend, so an agent that lies about its output tokens buys itself an approval. answerCost now refuses negative or non-finite token counts, which closes every door that routes through it — serve's POST /cost, the gateway, and the guard — and the MCP tool refuses them as an InvalidArguments the protocol knows how to carry.

A negative budget was judged; a negative volume was billed. assemble took budget: "-5" and returned the verdict over — a judgement against a limit that cannot exist — and callsPerMonth: -100 priced a prompt at −$1.26 a month, a number no bill ever had. Anything that is not a positive finite number now lands where it belongs: no budget, no volume, said with the verdict's own reason. prompt_writer additionally refuses them by name, because its schema said minimum: 1 and the runtime never enforced it — a schema the runtime does not enforce is documentation wearing a guard's clothes.

--cache-hit-rate 2 was accepted while the config refused it. usage.cacheHitRate: 2 in trazum.config.json fails with "a fraction between 0 and 1"; the same value through the flag was accepted and quietly skewed the caching advisory. Two doors to the same value cannot disagree about what fits through. The flag now refuses, in both locales, naming the config's own rule.

Every fix was proved by breaking the product: the fixed point reduced to one pass fails two cases; the negative-token refusal deleted fails the negative-money property.