Repository navigation
v0.39.0: cognition solve gate-move-ready (word answers + tier-aware thinking) - #26
Merged
Merged
Conversation
…hinking)
Mirrors the Python flock (langford v0.18.0 / smolag v0.13.0 / dantic v0.14.0) for
the ElizaOS plugin. The cogproof reader-column study shows single-step arithmetic
barely separates a reader from a scanner (+11-17pp) while the moat is the
multi-step / comprehension tiers (twostep +94pp, comprehension +100pp) — which a
real small model clears, so they don't over-gate. Two non-regressing changes ready
a gate move to those tiers:
- Word answers: parseCognitionAnswer(text, {wordsOk}) adds a last-word fallback
used only on the comprehension path; the arithmetic path stays digit-or-null
(a refusal never submits its last word).
- Tier-aware solve: COLONY_COGNITION_ALLOW_THINK (default false). Off = terse
integer-only arithmetic prompt, 200 tokens (current behaviour, zero regression).
On = a prompt admitting reasoning and a word answer, 1024 tokens. NB Gemma has
no qwen `/no_think` switch, so tier-awareness here is prompt + budget + answer
shape, not a soft-switch.
Default false preserves 0.38.0 behaviour exactly. typecheck + build clean,
1960 tests pass (+2).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TRn9SBFGaxRwZbwRsKNJ7b
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Mirrors the Python flock (langford #9, smolag #8, dantic #7) for the ElizaOS plugin, so eliza-gemma is ready if the live cognition gate moves to the tiers that actually separate a reader from a scanner.
Why
The cogproof reader-column study (Cogproof/cogproof
research/) measured reader−scanner separation per tier: single-step arithmetic is weak (+11–17pp) whiletwostep(+94pp) andcomprehension(+100pp) are the moat — and a real small model clears them (94%/100%), so they don't over-gate. Two gaps blocked a gate move; both fixed here without changing current single-step behaviour.Changes
parseCognitionAnswer(text, { wordsOk })gains a last-word fallback used only on the comprehension path (answer is a subject word likecrab); the arithmetic path stays digit-or-null, so a refusal never submits its last word.COLONY_COGNITION_ALLOW_THINK(default false). Off = terse integer-only arithmetic prompt at 200 tokens (current behaviour, zero regression). On = a prompt admitting reasoning and a word answer, 1024 tokens. Gemma has no qwen/no_thinkswitch, so tier-awareness here is prompt + budget + answer shape (the key difference from the Python trio).Verify
tsc --noEmitclean,tsupbuild clean, 1960 tests pass (+2). Default false keeps 0.38.0 behaviour exactly.COLONY_COGNITION_ALLOW_THINKon eliza-gemma when the gate serves multi-step/comprehension.🤖 Generated with Claude Code