Skip to content

docs(specs): built-but-dark spec set — Liveness Reconciler, activation-coherence, stop-gate completion (DRAFTS, for review)#379

Draft
JKHeadley wants to merge 5 commits into
mainfrom
echo/built-but-dark-specs
Draft

docs(specs): built-but-dark spec set — Liveness Reconciler, activation-coherence, stop-gate completion (DRAFTS, for review)#379
JKHeadley wants to merge 5 commits into
mainfrom
echo/built-but-dark-specs

Conversation

@JKHeadley
Copy link
Copy Markdown
Owner

Draft — not for merge as-is. Durable home + review surface for the three specs from the topic-12702 "built but dark" investigation. The P0 safety fix that this investigation surfaced already shipped separately (#377, v1.2.72).

What's here (all approved: false)

  1. built-but-dark-liveness-reconciler.md (+ ELI16 + convergence report) — the standing watchdog that reconciles declared-intent vs runtime-reality across the whole feature surface, ledger-informed + flood-free. Through one internal multi-lens convergence (caught & fixed a cold-start-flood design flaw). Open design call: day-one baseline (silent-accept vs paced review).
  2. feature-activation-coherence.md (+ ELI16 + convergence report) — the catalog systematically lies about runtime on/off; fix the enable layer + per-feature dispositions (finish/improve/merge/retire). Through one internal convergence — see the convergence note for blocking findings (B1 response-review correction, B2 runtime-probe design) and the recommended split (enable-layer-coherence vs behavior-disposition).
  3. context-death-stop-gate-rollout-completion.md (+ ELI16) — completes the already-approved context-death spec's unfinished rollout (hook router never wired + never flipped past mode=off).

What's needed to move these to implementation

  • Ratification (approved: true — your call, per the instar-dev gate).
  • Two design decisions: the Reconciler baseline approach; and whether to run an external cross-model review round before ratifying.
  • Resolve the activation-coherence convergence's B2 (probe-map design) + apply the recommended M5 split.

Each spec leads with a plain-English ELI16 overview. Nothing here changes runtime — pure docs.

🤖 Generated with Claude Code

Instar Agent (echo) and others added 5 commits May 25, 2026 01:50
…gate rollout-completion (drafts)

Two draft specs for the 'built but dark' gravity well (topic 12702):
- context-death-stop-gate-rollout-completion: completes the approved
  context-death-pitfall-prevention rollout (hook router never wired,
  no config dial, never flipped past mode=off). Cites the approved parent.
- built-but-dark-liveness-reconciler: NEW design. Automates verify-claim
  across the whole feature surface; ledger-informed (reuses FeatureRegistry
  discovery.db) so it surfaces only UNEXPLAINED dark features; 5-layer
  flood-free suppression; detect-only (posture A). Dogfood gate: must flag
  both currently-dark context-death gates.

Both ship with ELI16 companions. Drafts — awaiting /spec-converge + ratification.
…ernal multi-lens)

4-lens review (architecture/standards/failure-modes/grounding) found the
draft would flood on first run (100+ false-positive dark findings) and
overclaimed reuse of FeatureRegistry/CapabilityIndex/runtime-call-signals.
Revised: intent-modality axis (should-be-live vs should-be-offerable),
baseline anchor, dedicated LivenessLedger (reuses anti-flood patterns,
own state model), corrected wiring source, honest detection scope,
structural reason-capture (fixes willpower C2), generateClaudeMd
deliverable (fixes C1), self-liveness heartbeat, per-PR test tiers.
Convergence report records every finding->resolution.

Still draft: open-question #1 (baseline trade) needs Justin; external
cross-model round recommended before approved:true.
…er fix + per-feature dispositions

Live dogfood (topic 12702) + grounded re-assessment found the FeatureDefinitions
catalog systematically desynced from runtime: input-guard/evolution/publishing are
actually always-on (advertised opt-in/off); autonomous-evolution toggle writes a key
no code reads (real gate = AutonomyProfileManager.evolutionApprovalMode); telemetry
enable is a chicken-and-egg deadlock; dispatches inverted for the maintainer + enable
key not allowlisted. Spec: fix enable layer (single derived state, enableAction-validity
test, always-construct/gate-effects, catalog-truth test) + per-feature dispositions
(finish/improve/merge/retire) + hand standing-detector role to the Liveness Reconciler.
Bonus dark findings: UnjustifiedStopGate (StopGateDb never constructed), MessageSentinel
(maybe not on live inbound path — safety-relevant). ELI16 leads with an honest correction
of my earlier mis-diagnosis.
…ine-owned agents (P0)

Direct data-flow verification: sentinel intercept lives only in
TelegramAdapter.processUpdate (adapter poll loop); Echo runs lifeline-owned
polling (server.ts:1167) so inbound goes lifeline -> /internal/telegram-forward
-> inject, bypassing the sentinel entirely (routes.ts:8391-8700 has zero
sentinel refs). 'stop everything' is not structurally honored for Echo.
Classifier itself live-tested OK. Fix = hoist sentinel into the forward path +
wiring/integration tests. Marked P0 in feature-activation-coherence.
… response-review evidence

Independent critical review: core thesis sound. Blocking: (B1) response-review
is NOT dead plumbing — it's registered for new agents + opt-in/off, with a
migration-parity gap for existing agents; corrected the evidence + re-based the
MERGE disposition on redundancy. (B2) runtime-probe under-specified → needs a
featureRuntimeProbe map in server.ts. Recommended split (M5) into enable-layer
coherence vs behavior-disposition. Convergence report records findings->resolutions.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@vercel
Copy link
Copy Markdown

vercel Bot commented May 25, 2026

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
instar Building Building Preview May 25, 2026 8:56am

Request Review

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant