DataViking-Tech · the-data-viking · Jul 20, 2026 · Jul 20, 2026 · Jul 20, 2026 · Jul 20, 2026
diff --git a/.github/scripts/verify-ci-checks.sh b/.github/scripts/verify-ci-checks.sh
@@ -13,7 +13,7 @@
 # concurrency group cancels one of the twin runs — park a cancelled or
 # failed run next to the successful same-named run, and those superseded
 # runs never turn green. Counting all of them blocked SynthBench v0.3.1
-# (PR #334) and SynthPanel's first v1.6.0 attempt, forcing manual tags.
+# (PR #334) and Althing's first v1.6.0 attempt, forcing manual tags.
 # Only the newest run per (app, name) reflects the commit's real status;
 # GitHub's own merge box collapses check runs the same way.
 #

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -30,8 +30,8 @@ jobs:
       - run: ruff check .
       - run: ruff format --check .
       - name: Guard against hyphenated sibling-project regression (sb-1el9)
-        # The sibling project's canonical names are `synthpanel` (PyPI package)
-        # and `SynthPanel` (GitHub repo). The stale pre-rename hyphenated form
+        # The sibling project's canonical names are `althing` (PyPI package)
+        # and `Althing` (GitHub repo). The stale pre-rename hyphenated form
         # should never reappear. Historical audit notes (post-flip-audit-*.md)
         # are allowed to reference the old name while documenting the rename.
         # The pattern is assembled at runtime so this step doesn't match itself.
@@ -46,7 +46,7 @@ jobs:
             --exclude="post-flip-audit-*.md" \
             . || true)
           if [ -n "$matches" ]; then
-            echo "::error::Hyphenated sibling-project name found — use 'synthpanel' (PyPI) or 'SynthPanel' (repo)."
+            echo "::error::Hyphenated sibling-project name found — use 'althing' (PyPI) or 'Althing' (repo)."
             echo "$matches"
             exit 1
           fi
@@ -329,7 +329,7 @@ jobs:
       - name: Mint App token
         # Pushing with this token (instead of GITHUB_TOKEN) makes the
         # baseline commit fire `pull_request: synchronize`, so CI re-runs on
-        # the new HEAD automatically. Same pattern as SynthPanel's
+        # the new HEAD automatically. Same pattern as Althing's
         # auto-tag.yml release-bot push.
         if: steps.vrt_gate.outputs.run == 'true'
         id: app_token

diff --git a/.github/workflows/semver-label.yml b/.github/workflows/semver-label.yml
@@ -4,7 +4,7 @@ name: Semver Label
 # (semver:patch / semver:minor / semver:major to bump, semver:skip to opt
 # out) BEFORE merge. auto-tag.yml reads the label post-merge to cut the
 # release; without this gate a missing label silently skips the release
-# behind a green check (the failure mode SynthPanel hit twice — sp-42i,
+# behind a green check (the failure mode Althing hit twice — sp-42i,
 # sy-73j — before adopting the same pre-merge check).
 
 on:

diff --git a/FINDINGS.md b/FINDINGS.md
@@ -2,6 +2,12 @@
 
 **Date**: 2026-04-12 (session-1 experiments) | **Numbers regenerated from artifacts** — see below
 
+> **Naming note (2026-07):** the `synthpanel` provider referenced throughout
+> these findings was renamed to **[althing](https://althing.dev)**. Result
+> artifacts and generated tables retain the provider id `synthpanel` they
+> were recorded under; `--provider synthpanel` remains a working alias for
+> the `althing` provider.
+
 > All quantitative tables in this document sit between
 > `<!-- BEGIN GENERATED -->` / `<!-- END GENERATED -->` markers and are
 > rendered by `scripts/generate-findings-md.py` from the per-question rows
@@ -34,9 +40,9 @@ asymmetries in how LLMs represent different demographic groups.
 
 | Dataset | 3-model ensemble SPS | Best single model | Random baseline |
 |---------|---------------------|-------------------|-----------------|
-| globalopinionqa (n=100) | **0.813** | 0.786 (SynthPanel (GPT-4o-mini), product) | 0.774 (n=100) |
+| globalopinionqa (n=100) | **0.813** | 0.786 (Althing (GPT-4o-mini), product) | 0.774 (n=100) |
 | opinionsqa (n=684) | **0.877** | 0.829 (Gemini 2.5 Flash, raw) | 0.763 (n=684) |
-| subpop (n=200) | **0.858** | 0.821 (SynthPanel (Gemini Flash Lite), product) | 0.757 (n=200) |
+| subpop (n=200) | **0.858** | 0.821 (Althing (Gemini Flash Lite), product) | 0.757 (n=200) |
 <!-- END GENERATED: headline -->
 
 Note on interpretation: SPS is a composite parity score, not a
@@ -140,7 +146,7 @@ run vs the identical configuration with schema-forced structured capture
 count — the elicitation surface is the only variable.
 
 <!-- BEGIN GENERATED: elicitation -->
-**SynthPanel (Haiku 4.5)** (product) on gss — n=75 questions × 30 samples per arm, refusal detector v2; the arms differ only in elicitation mode:
+**Althing (Haiku 4.5)** (product) on gss — n=75 questions × 30 samples per arm, refusal detector v2; the arms differ only in elicitation mode:
 
 | Metric | Natural | Structured | Δ (structured − natural) |
 |--------|---------|------------|--------------------------|
@@ -252,9 +258,9 @@ three runs costs roughly 3× a single-model run.
 <!-- BEGIN GENERATED: ensemble -->
 | Dataset | Best single model | Equal blend | Improvement | Random baseline |
 |---------|-------------------|-------------|-------------|-----------------|
-| globalopinionqa (100q) | 0.786 (SynthPanel (GPT-4o-mini), product) | **0.813** | **+2.7 pts** | 0.774 (n=100) |
+| globalopinionqa (100q) | 0.786 (Althing (GPT-4o-mini), product) | **0.813** | **+2.7 pts** | 0.774 (n=100) |
 | opinionsqa (684q) | 0.829 (Gemini 2.5 Flash, raw) | **0.877** | **+4.8 pts** | 0.763 (n=684) |
-| subpop (200q) | 0.821 (SynthPanel (Gemini Flash Lite), product) | **0.858** | **+3.7 pts** | 0.757 (n=200) |
+| subpop (200q) | 0.821 (Althing (Gemini Flash Lite), product) | **0.858** | **+3.7 pts** | 0.757 (n=200) |
 
 Comparison set: best_single = highest recomputed-SPS non-ensemble, non-baseline leaderboard row (raw or product framework) evaluated on at least as many questions as the ensemble, after per-(model, framework, dataset) dedup; random_baseline = recomputed SPS of the random-baseline run for the same dataset (n in random_baseline_n).
 <!-- END GENERATED: ensemble -->
@@ -277,25 +283,25 @@ artifacts strip `human_distribution` (#308).
 
 | Constituent pair | Pearson r (per-question JSD) | n |
 |------------------|------------------------------|---|
-| SynthPanel (Haiku 4.5) ↔ SynthPanel (Gemini Flash Lite) | 0.419 | 100 |
-| SynthPanel (Haiku 4.5) ↔ SynthPanel (GPT-4o-mini) | 0.308 | 100 |
-| SynthPanel (Gemini Flash Lite) ↔ SynthPanel (GPT-4o-mini) | 0.493 | 100 |
+| Althing (Haiku 4.5) ↔ Althing (Gemini Flash Lite) | 0.419 | 100 |
+| Althing (Haiku 4.5) ↔ Althing (GPT-4o-mini) | 0.308 | 100 |
+| Althing (Gemini Flash Lite) ↔ Althing (GPT-4o-mini) | 0.493 | 100 |
 
 **opinionsqa** (n=684 common questions, mean pairwise r = 0.295):
 
 | Constituent pair | Pearson r (per-question JSD) | n |
 |------------------|------------------------------|---|
-| SynthPanel (Haiku 4.5) ↔ SynthPanel (Gemini Flash Lite) | 0.224 | 684 |
-| SynthPanel (Haiku 4.5) ↔ SynthPanel (GPT-4o-mini) | 0.355 | 684 |
-| SynthPanel (Gemini Flash Lite) ↔ SynthPanel (GPT-4o-mini) | 0.308 | 684 |
+| Althing (Haiku 4.5) ↔ Althing (Gemini Flash Lite) | 0.224 | 684 |
+| Althing (Haiku 4.5) ↔ Althing (GPT-4o-mini) | 0.355 | 684 |
+| Althing (Gemini Flash Lite) ↔ Althing (GPT-4o-mini) | 0.308 | 684 |
 
 **subpop** (n=200 common questions, mean pairwise r = 0.322):
 
 | Constituent pair | Pearson r (per-question JSD) | n |
 |------------------|------------------------------|---|
-| SynthPanel (Haiku 4.5) ↔ SynthPanel (Gemini Flash Lite) | 0.282 | 200 |
-| SynthPanel (Haiku 4.5) ↔ SynthPanel (GPT-4o-mini) | 0.260 | 200 |
-| SynthPanel (Gemini Flash Lite) ↔ SynthPanel (GPT-4o-mini) | 0.425 | 200 |
+| Althing (Haiku 4.5) ↔ Althing (Gemini Flash Lite) | 0.282 | 200 |
+| Althing (Haiku 4.5) ↔ Althing (GPT-4o-mini) | 0.260 | 200 |
+| Althing (Gemini Flash Lite) ↔ Althing (GPT-4o-mini) | 0.425 | 200 |
 
 The constituents' errors are **moderately positively correlated** (pairwise r 0.22–0.49 on per-question JSD; 0.27–0.44 on signed per-option residuals, see asserted constants) — not uncorrelated. The ensemble gain comes from partial, not full, independence of errors; the earlier "uncorrelated errors" framing overstated it.
 
@@ -403,8 +409,8 @@ gated files are committed without `human_distribution`).
 | Provider | Framework | Template | Dataset | Mean model nonresponse | Mean human nonresponse | Mean abs gap | n |
 |----------|-----------|----------|---------|------------------------|------------------------|--------------|---|
 | Gemini 2.5 Flash | raw | default | gss | 0.106 | 0.056 | **0.126** | 75 |
-| SynthPanel (Haiku 4.5) | product | structured | gss | 0.064 | 0.056 | **0.095** | 75 |
-| SynthPanel (Haiku 4.5) | product | default | gss | 0.067 | 0.056 | **0.094** | 75 |
+| Althing (Haiku 4.5) | product | structured | gss | 0.064 | 0.056 | **0.095** | 75 |
+| Althing (Haiku 4.5) | product | default | gss | 0.067 | 0.056 | **0.094** | 75 |
 
 Comparison set: Per deduped non-baseline run whose committed per-question rows still carry human_distribution (full-tier datasets; gated files are stripped per #308): mean |model explicit-nonresponse mass (DK-style option mass + parsed refusal rate) - human's|, plus the five items with the largest model over-selection.
 

diff --git a/METHODOLOGY.md b/METHODOLOGY.md
@@ -33,7 +33,7 @@ For the full technical details, read on.
 
 ### 1.1 Analysis of the Two Frameworks
 
-**Proposed (synthpanel datascientist) — 4 sub-scores:**
+**Proposed (althing datascientist) — 4 sub-scores:**
 | Metric | Measures | Method |
 |--------|----------|--------|
 | P_dist | Distributional parity | Jensen-Shannon divergence |
@@ -144,7 +144,7 @@ default distribution. P_cond isolates the *conditioning mechanism itself*.
 
 **Conditioning protocol**: Each provider is tested with the persona spec it
 supports. For raw LLMs, we test all three OpinionsQA styles (QA, BIO, PORTRAY)
-and take the max. For dedicated providers (synthpanel, Ditto), we use their
+and take the max. For dedicated providers (althing, Ditto), we use their
 native persona interface.
 
 #### P_sub — Subgroup Consistency
@@ -269,7 +269,7 @@ class Distribution:
 @dataclass
 class ProviderMetadata:
     """Provider identification and capabilities."""
-    name: str                      # e.g. "synthpanel", "openai-gpt4o"
+    name: str                      # e.g. "althing", "openai-gpt4o"
     version: str                   # Semantic version or model identifier
     supports_logprobs: bool        # Can return token log-probabilities
     supports_distribution: bool    # Can return Distribution directly
@@ -329,7 +329,7 @@ class SynthRespondent(Protocol):
 
 | Provider | Adapter | Distribution Method | Notes |
 |----------|---------|-------------------|-------|
-| **synthpanel** | `SynthPanelAdapter` | Reported (native API returns distribution) | Our reference provider |
+| **althing** | `AlthingAdapter` | Reported (native API returns distribution) | Our reference provider |
 | **OpenAI (GPT-4o, etc.)** | `OpenAIAdapter` | Logprobs (top-k token logprobs) | `conditioning_style` mapped to system/user prompt |
 | **Anthropic (Claude)** | `AnthropicAdapter` | Sampling (no logprobs available) | Requires n_samples >= 30 for stable distributions |
 | **Ditto** | `DittoAdapter` | Reported (if API returns distribution) or Sampling | Pending API access to confirm capabilities |
@@ -372,7 +372,7 @@ class OpenAIAdapter:
 CLI invocation:
 ```bash
 synthbench run --provider openai --model gpt-4o --suite core
-synthbench run --provider synthpanel --suite full
+synthbench run --provider althing --suite full
 synthbench run --provider generic-http --config ./my-provider.yaml --suite core
 ```
 
@@ -617,9 +617,9 @@ Every SynthBench run produces a JSON artifact:
   "timestamp": "2026-04-10T12:00:00Z",
   "suite": "core",
   "provider": {
-    "name": "synthpanel",
+    "name": "althing",
     "version": "2.1.0",
-    "adapter": "SynthPanelAdapter",
+    "adapter": "AlthingAdapter",
     "config_hash": "sha256:..."
   },
   "scores": {
@@ -669,7 +669,7 @@ Every SynthBench run produces a JSON artifact:
 ================================================================
   SYNTHBENCH v1.0 SCORE CARD
 ================================================================
-  Provider:  synthpanel v2.1.0
+  Provider:  althing v2.1.0
   Suite:     SynthBench-Core (300 questions)
   Date:      2026-04-10
   Run ID:    run-20260410-a1b2c3
@@ -720,7 +720,7 @@ For multi-provider comparisons:
 ```
 | Provider        | Version | SPS  | P_dist | P_rank | P_cond | P_sub | P_refuse | Suite |
 |-----------------|---------|------|--------|--------|--------|-------|----------|-------|
-| synthpanel      | 2.1.0   | 0.74 | 0.81   | 0.77   | 0.68   | 0.72  | 0.71     | core  |
+| althing      | 2.1.0   | 0.74 | 0.81   | 0.77   | 0.68   | 0.72  | 0.71     | core  |
 | GPT-4o (portray)| 2026-03 | 0.61 | 0.67   | 0.63   | 0.55   | 0.58  | 0.62     | core  |
 | Claude (portray)| sonnet  | 0.59 | 0.65   | 0.61   | 0.52   | 0.60  | 0.58     | core  |
 | GPT-4o (raw)    | 2026-03 | 0.48 | 0.55   | 0.52   | 0.00   | 0.45  | 0.40     | core  |

diff --git a/README.md b/README.md
@@ -9,7 +9,7 @@ Open benchmark harness for synthetic survey respondent quality.
 
 **The MLPerf of synthetic UXR.**
 
-SynthBench measures how well synthetic respondent systems (like [synthpanel](https://github.com/DataViking-Tech/SynthPanel), Ditto, Synthetic Users, or raw ChatGPT prompting) reproduce real human survey response patterns against real Pew American Trends Panel and GlobalOpinionQA ground truth — so "it sounds plausible" gets replaced with a measurable similarity score.
+SynthBench measures how well synthetic respondent systems (like [althing](https://github.com/DataViking-Tech/Althing), Ditto, Synthetic Users, or raw ChatGPT prompting) reproduce real human survey response patterns against real Pew American Trends Panel and GlobalOpinionQA ground truth — so "it sounds plausible" gets replaced with a measurable similarity score.
 
 ## Quick Start
 
@@ -203,7 +203,7 @@ The leaderboard JSON carries per-row cost fields and a top-level
   runs listed in `config.ensemble_sources`.
 - `pricing_snapshot` records the per-model `input_per_1m` / `output_per_1m`
   rates used for this publish run, the `snapshot_date` anchor comment, and
-  the installed `synth_panel_version` that produced the rates.
+  the installed `althing_version` that produced the rates.
 
 This lets downstream consumers audit which pricing table produced which
 `cost_usd` and reconcile against provider-reported billing without guessing
@@ -220,7 +220,7 @@ both curves side-by-side.
 
 See [`docs/convergence-analysis.md`](docs/convergence-analysis.md) for the
 JSON schema, CLI flags, and the `load_convergence_baseline` integration
-surface that synthpanel's `--calibrate-against DATASET:QUESTION` flag
+surface that althing's `--calibrate-against DATASET:QUESTION` flag
 consumes.
 
 ## Citation

diff --git a/crew/architect/docs/accessibility-guidelines.md b/crew/architect/docs/accessibility-guidelines.md
@@ -228,7 +228,7 @@ Color-only encoding fails for ~8% of male users (colorblindness). For small stac
 ### Badges
 
 - `bg-accent/15 text-accent` — acceptable with the darker accent token (verify per-instance).
-- Provider tags (`bg-provider-synthpanel/10 text-provider-synthpanel`) — provider colors are hex literals; check each one against the surfaces they sit on.
+- Provider tags (`bg-provider-althing/10 text-provider-althing`) — provider colors are hex literals; check each one against the surfaces they sit on.
 
 ---