Governance: model + effort tier as a per-dispatch proportionality call#2
Open
cyphhh wants to merge 1 commit into
Open
Governance: model + effort tier as a per-dispatch proportionality call#2cyphhh wants to merge 1 commit into
cyphhh wants to merge 1 commit into
Conversation
The Classification Gate already frames per-task *effort* as the challengeable record behind the global effort dial, but says nothing about *model tier*. Capable models over-provision by default — routing mechanical work to the session's own top tier at high effort because it's the frictionless choice. That is Complexity Accumulation in capability spend: slower and costlier with no gain on bounded work. This adds the missing axis in v8's own idiom, without new machinery: - harness-core.md (Classification Gate): one dense addition extending the effort sentence to model tier — set each worker's tier/effort per task, not at the session default; escalate one worker reactively rather than pre-provisioning. Framed against the Complexity Accumulation anti-pattern. - references/prompt-design.md: a compact "Model & Effort Tier for Sub-Agents" subsection under Context Budget with the concrete mechanics (Agent `model`; Workflow agent() model/effort) and reactive-escalation rules of thumb. Governance-layer only: the firmware supplies the whether/when/how-well judgment; the runtime already executes the how via the model/effort params. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Voxinator
added a commit
that referenced
this pull request
Jul 13, 2026
Validator packaged with installed skills + hardened (hostile fixtures, duplicate-JSON-key rejection); manifest-based uninstall preserving user agents; effects-honest settings example; tiered verified_producer/ independent/adversarial terminal states + five evidence classes; available/configured capability split; A3 escalator narrowed; chatgpt profile rescoped to Projects w/ ChatGPT Work note; metadata bootstrap rerouted; anchored marker regexes + unterminated-block refusal after the off-contract data-loss catch. Includes the judged PLAN-r9-hardening ahead of its execution. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Adds model tier alongside effort as a per-dispatch proportionality judgment, in v8's existing idiom — no new sections in the machinery sense, no new machinery.
core/harness-core.md(Classification Gate): one dense addition extending the existing effort sentence. The gate already frames per-task effort as the challengeable record behind the globalsettings.effortLeveldial; this adds the model-tier axis of the same judgment — set each worker's tier/effort per task, not at the session default, and escalate one worker reactively rather than pre-provisioning.references/prompt-design.md: a compactModel & Effort Tier for Sub-Agentssubsection under Context Budget, with the concrete mechanics (Agenttoolmodel; Workflowagent()model/effort) and reactive-escalation rules of thumb.Why
The Classification Gate governs per-task effort but is silent on model tier. Capable models over-provision by default — routing mechanical work to the session's own top tier at high effort because it's the frictionless choice. That's Complexity Accumulation in capability spend: slower and costlier with no gain on bounded work. The change is framed explicitly against that anti-pattern.
Fit with v8's boundary
Governance-layer only. The firmware supplies the whether / when / how-well judgment (which tier a given worker warrants); the runtime already executes the how via the existing
model/effortparams. Nothing is hand-rolled that the runtime does natively — it just supplies the proportionality judgment the runtime doesn't.Diff
26 insertions, 1 deletion, two files. Based on current
main(v8).🤖 Generated with Claude Code