Polylogue Proof Artifacts
This page maps public-facing claims to inspectable artifacts. It is deliberately stricter than a feature list: if a claim has no artifact, phrase it as a capability under construction instead of as a proven result.
Current Claim Map
| Claim | Status | Artifact | What It Proves | Caveat |
|---|---|---|---|---|
| A deterministic private-data-free corpus exercises the advertised demo constructs. | Fact | Demo Corpus Construct Audit | A fresh no-daemon seed/verify run produces non-empty coverage for declared source families: five origins, acquired attachments, browser-capture coalescing, lineage, subagent runs, terminal states, synthetic embeddings, and user overlays. | The corpus proves product behavior and construct coverage, not private archive scale. |
| Polylogue reads tool outcomes from structure, not assistant prose. | Fact | Claim-vs-Evidence summary and devtools workspace claim-vs-evidence |
The failure predicate is normalized tool-result evidence: is_error=1 or non-zero exit_code. Assistant prose is only a follow-up acknowledgment signal. |
The live aggregate is private-archive evidence; reproduce the method on the demo archive. |
| Codex token accounting bills disjoint lanes instead of double-counting cached input. | Fact | Cost accounting example and Cost Model | The synthetic run uses the real writer and pricing logic to show fresh_input + cache_read = provider input, avoiding the old cached-token double bill. |
Provider prices and subscription equivalence are curated estimates, not billing authority. |
| Polylogue can analyze agent affordance usage across MCP and tool-call spellings. | Fact | Agent affordance usage summary and devtools workspace affordance-usage |
Family-normalized counts group tool aliases such as Serena, Lynchpin, Polylogue MCP, Context7, and codebase tools across captured action evidence. | Counts measure captured usage and failure signals, not independent utility or user benefit. |
| Agents with reviewed memory should recover better. | Promise | Pending uplift experiment (polylogue-cfk) |
The substrate can compile context bundles, judged notes, and resume briefs. | Do not claim measured performance uplift until the paired experiment is complete. |
Demo Corpus
Regenerate the construct audit:
devtools render demo-corpus-datasheet
devtools render all --check
The current tracked datasheet reports:
- 13 sessions
- 55 indexed messages
- 99 blocks
- 13 session profiles
- 14 run rows
- 45 observed-event rows
- 14 context-snapshot rows
- five origins:
aistudio-drive,chatgpt-export,claude-ai-export,claude-code-session,codex-session - no residual declared-construct gaps
Claim-vs-Evidence Failure Follow-Up
The current private-archive demo shelf carries a public-safe aggregate summary generated on 2026-07-04 against archive schema v24. It inspected a bounded, origin-stratified sample of 5,000 structured failures from a frame of 42,033 structured failures:
- acknowledged next turn: 420
- silent-proceed next turn: 1,205
- ambiguous next turn: 3,375
- silent lower bound: 24.1%
- acknowledged within the next three assistant turns: 722
- marker calibration rows: 50 labeled
- acknowledged-marker precision: 100.0%
- acknowledged-marker recall: 84.2%
Regenerate the local packet:
devtools workspace claim-vs-evidence \
--limit 5000 \
--out-dir .agent/demos/claim-vs-evidence \
--json
devtools workspace demo-shelf
Reproduce the method without private data:
export POLYLOGUE_ARCHIVE_ROOT=/realm/tmp/polylogue-claim-vs-evidence-demo
polylogue demo tour --out-dir /realm/tmp/polylogue-demo-tour --force
polylogue demo seed --root "$POLYLOGUE_ARCHIVE_ROOT" --force --with-overlays --format json
polylogue demo verify --root "$POLYLOGUE_ARCHIVE_ROOT" --require-overlays --format json
devtools workspace claim-vs-evidence \
--archive-root "$POLYLOGUE_ARCHIVE_ROOT" \
--limit 5000 \
--out-dir /realm/tmp/polylogue-claim-vs-evidence-repro \
--json
The current one-command tour packet is committed under
docs/examples/demo-tour/: report.md,
transcript.txt, command outputs, recording.tape, demo-tour.gif, and the
uvx --from <repo> install-path timing note.
Agent Affordance Usage
The current demo shelf includes a recent-session-window affordance report. Top families in that window:
lynchpin: 32 actions across 9 raw tool names; 3 distinct sessions; errors 0.polylogue: 21 actions across 10 raw tool names; 6 distinct sessions; errors 3.serena: 15 actions across 3 raw tool names; 4 distinct sessions; errors 0.
Regenerate:
devtools workspace affordance-usage \
--out-dir .agent/demos/agent-affordance-usage \
--json
devtools workspace demo-shelf
Use this as product-development evidence, not as a leaderboard. It answers "which affordances are actually being used, with what structured failure signals?" A utility judgment still requires reading the session context around the samples.
Release Readiness
The release gate is tracked separately at Release Readiness Gate. Before a public release, the README claims, generated docs, demo corpus, visual evidence, and packaging checks must all agree with that gate.