Skip to content

TScert v0.7

Choose a tag to compare

@cristianoc cristianoc released this 02 Nov 21:26
· 557 commits to main since this release

TScert is a research verifier for type claims about TypeScript functions. A certificate is accepted when it passes an independent checker whose soundness is proved in Lean.

v0.7 meets the 70% front-end function coverage threshold. The complete tools/ census verifies 489 of 679 functions (72.0%), compared with 411 of 644 (63.8%) at v0.6. Counted statement coverage is separately 5,487 of 15,715 (34.9%), compared with 4,441 of 15,235 (29.1%) at v0.6.

The measurement is the clean main census of 9bf8c463, taken on 2026-11-02 with a 300-second stage limit. Release commit d4eaa54e retains exactly that commit's tools/, verifier, producer and pipeline code; its subsequent changes document the results and clarify report scopes.

The headline counts named top-level function declarations with bodies in tools/, once each. Unsupported, unverified and timed-out functions remain in the denominator. Module initializers, class members and functions held by module bindings are outside this metric. See the function and release metric, statement metric, and dated work ledger.

Changes

  • Checked callback footprints and stable member-read facts support more front-end code while retaining the distinction between normal and exceptional returns.
  • Ordinary-call spread snapshots use the existing dense-copy checker, preserving evaluation order when later arguments have effects. Negative and forged-evidence controls exercise this boundary.
  • Shared complete effect constructors retain every native contract field. The rollout preserves a previously verified dependent; two catalog migrations that caused a timeout are excluded.
  • Request-local sharing of strictly validated types, proved-equivalent normalization indexes, and an audited search shortcut reduce repeated work. On one identical saved request, CPU falls 92.9 to 8.5 seconds and peak memory 2.63 to 1.04 GB, with byte-identical refusal output. This is a fixed-input result.
  • Explicit project data contracts, live callback-option views and a shared semantic printer remove duplicated or weaker representations.

The latest twenty-case batch, compared with the complete main census at ca4e3c8a, gains 29 verified functions and loses none. One unverified forwarding helper is deleted. All 2,845 pipeline cases, 21 suites and 214 controls pass. The latest batch and the representation patch received independent Astra review.

Cost and remaining limits

Recorded unique producing-search CPU across the two main censuses increases 2,909 to 6,088 seconds. Most of the increase is in 20 refused type-term slices that reach a later predicate obligation; their cost rises 51 to 3,355 seconds. The 460 functions verified in both censuses have recorded search cost 1,447 to 1,346 seconds. Inputs, dependency contracts, verifier implementations and cache histories differ; these totals do not establish a same-work project speed ratio. Timing and shared-work accounting defines the scopes.

One previously unverified function, classStmt, crosses the 300-second wall limit. Its cause remains unknown. The supported fragment, audited host boundaries and unproved source functions remain explicit.

What acceptance guarantees

For inputs meeting their types and a heap meeting its storage contracts, an accepted function's translation into the calculus does not get stuck or return a value outside its result type. It may throw or run forever. The theorem SourceClaimJson.checkJson_safe connects successful checking to this property.

Applying that result to source relies on faithful translation and annotation interpretation, library and host contracts, conforming inputs, and the trusted tools and runtime assumptions. The coverage percentage is a measure of the specified front-end functions, rather than all TypeScript behavior.

Try it and license

Follow the setup instructions, then run:

make check FILE=pipeline/examples/array-last-element.ts

TScert uses the MIT NON-AI License.