Build Log

Making ForgeKit's Operating Diagnosis Traceable — Evidence, Signal, Vital, Charge

August 24, 2026

Fixed the real disconnect between Evidence Tray, Signals, and OS Vitals (seven independent recurrence computations, zero shared taxonomy) with a narrow, additive correction rather than a new shared graph library, added genuinely greenfield economics and negative-evidence telemetry, and had two real overclaims caught and fixed by the Observer during the build.

AlloyEvidence TraySignalsOS VitalsExperiment 0Autonomous Session

independent recurrence computations found, none reconciled

real Observer catches, both fixed

real negative-evidence occurrences surfaced for the first time

new evidence domains added (why-flagged factors, economics, negative evidence)

Timeline

Start

Charge received: make Evidence -> Signal -> Vital -> Charge traceable

Autonomous-execution experiment — Claude resolves architecture, escalates only on ambiguous product-meaning decisions

~30m

Full architecture map via research agent

Found 7 independent recurrence computations, 6 overlapping summary surfaces, 2 existing Charge-drafters (Signals, Evidence) but none for Vitals, zero cost/token schema anywhere

~45m

Scope decision: narrow additive fix, not a new shared graph library

Per the Charge's own explicit warning against overbuilding; confirmed with Zeb after the Observer returned a bare CONTINUE on the open design question

~1.5h

Shared category-gate-map extracted, priorityFactors added to Signals/Evidence, Vitals Charge-drafting built

First coherent build cycle — tsc clean

~2h

Observer CHALLENGE: priorityFactors implied all listed facts drove rank when only cost did

Real catch — renamed to whyFlagged across all 3 modules, explicit rank-vs-context framing added in both code and UI text, 3 regression tests added

~3h

Economics schema + negative-evidence domain built

New optional session-JSON economics block, honest insufficient-data handling verified against real (currently zero) data; negative evidence surfaced 83 real gate_checks failures never visible before

~3.5h

Observer CHALLENGE: negative evidence risked implying live unresolved gaps

Real catch — several sample entries were incidents that themselves produced the fix; reframed as historical occurrences throughout, with no fabricated resolved/unresolved inference

End

Full test suite + real-data acceptance chains verified, staged for commit

498/500 tests pass (2 pre-existing, confirmed via git-stash baseline check); all 3 acceptance chains run live against real production data

What shipped

products/alloy/lib/shared/category-gate-map.ts — the one copy of the spin-category-to-gate mapping, replacing a hand-mirrored duplicate in lib/signals/prioritization.ts, now also consumed by lib/os-vitals/staleness.ts's Attic view

Added whyFlagged (a real, explicit 'why this deserves attention' fact list, distinct from the single-dimension rank) to Signals' PrioritizedSignal, Evidence's MissedPackageSignal, and OS Vitals' Diagnosis types — every non-rank-determining fact explicitly says so, after a real Observer catch on the first naming (priorityFactors)

Built the first Vitals-level Charge-drafting function (computeDiagnosis's new draftedCharge) — OS Vitals previously had no Charge generator at all, unlike Signals and Evidence Tray

Wired real cross-page navigation: OS Vitals' Biggest Drag drawer now links its contributing categories directly to the Signals page — the actual Vital -> Signal provenance chain the Charge required

Extracted CopyChargeButton (previously duplicated in 2 files) into components/shared/ once a genuine third consumer (OS Vitals) arrived

Added an optional economics{} block to the session JSON schema (forgekit-os/sessions/_SCHEMA.md) — tokens, cost (with a mandatory is_estimate flag), Observer cost tracked separately, founder intervention count, accepted outcomes, rework events, escaped failures; every field null unless actually measured, no backfill

Built lib/evidence/economics.ts — reads the new field, implements an honest 'insufficient-data' vs 'computed' state for derived metrics (Cost per Accepted Outcome requires 5+ real sessions before computing anything)

Built lib/evidence/negative-evidence.ts — surfaces gate_checks entries recorded as failed/skipped/ceremonial across all sessions as an inspectable list, not just an aggregate percentage; found and fixed a separate small bug (missed_issue was documented in _SCHEMA.md but silently dropped by lib/os-vitals/sessions.ts's reader)

Wired both new evidence domains into Evidence Tray's cockpit UI as two new sections with their own drawers

Reconciled OS Vitals' deferral-ledger recurrence count against Signals' own recurrence count by stating, in the item's own detail text, that they measure different (if related) things via different category taxonomies — rather than presenting two silently-conflicting numbers

3 new regression tests pinning the whyFlagged rank-vs-context distinction as a permanent invariant