tokens-for-good 0.5.1 → 0.5.3

package/README.md

@@ -27,7 +27,7 @@ To change cadence later, run `npx tokens-for-good init` again.
 
 Each org takes about 5 minutes:
 
-1. **Research:** web search + 6-prompt methodology, then fill in a v3 EVIDENCE TABLE (8 rows of verbatim quotes + real URLs; blanks are honest when the evidence doesn't exist)
+1. **Research:** web search + 6-prompt methodology, then fill in a v3 EVIDENCE TABLE (8 rows of verbatim quotes + real URLs; blanks are honest when the evidence doesn't exist, and a genuinely evidence-free org is submitted with `no_evidence: true` rather than invented rows)
 2. **Verify:** every citation URL checked, hallucinations flagged and corrected
 3. **Humanize:** voice pass (no em dashes, no AI-tells, analyst voice)
 
@@ -56,10 +56,18 @@ Once installed, these are available to your AI via the MCP server:
 | `get_next_consolidation` | v3 consolidator: fetch your assignment + both source reports to merge |
 | `set_role_preference` | Prefer the low-fetch roles (validation/consolidation); best for local models |
 | `create_agent` / `list_agents` / `rotate_agent_key` / `revoke_agent` | Run several harnesses at once: each agent gets its own key + research slot |
-| `setup_automation` | Emits `/schedule` prompt (normally called by `/tfg-schedule` skill) |
+| `setup_automation` | Emits the self-contained `/schedule` prompt with the full methodology embedded (normally called by `/tfg-schedule` skill) |
 | `my_impact` / `research_status` | Your stats + the project leaderboard |
 | `snooze` | Quiet the session-start prompt for N days |
 
+## Security model
+
+Scheduled routines are **self-contained**: the full research methodology is embedded in the routine prompt when you set it up, so its instructions are frozen at install time — nothing remote can change what your standing agent does. At runtime the routine only exchanges JSON data with the TFG API (an org to research, a submission receipt, a version handshake at `/api/research/parameters`). Routines created before this format fetched instructions at runtime, which is why some harnesses showed a prompt-injection warning; upgrade in one command with `/tfg-schedule`.
+
+Server-side, every report passes deterministic citation verification (each EVIDENCE TABLE quote must appear on its cited page), an independent validator, and dual-research consolidation before scoring — and the scorer is code, never the model that wrote the report. API keys are SHA-256 hashed at rest and sent only via header.
+
+Have an idea to make TFG more secure or trustworthy? [Open an issue](https://github.com/Tokens-for-Good/tokens-for-good/issues/new/choose). For sensitive vulnerability reports, use a private support ticket from your [dashboard](https://tokensforgood.ai/support) instead of a public issue.
+
 ## Non-Claude-Code platforms
 
 - **OpenCode:** `init` writes `~/.config/opencode/opencode.json` and prints a cron line you can paste into `crontab -e`.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tokens-for-good",
-  "version": "0.5.1",
+  "version": "0.5.3",
   "type": "module",
   "description": "Contribute your spare AI tokens to research nonprofits for Fierce Philanthropy",
   "homepage": "https://tokensforgood.ai",

package/skills/tfg.md

@@ -30,3 +30,4 @@ The user wants to complete one Tokens for Good research cycle.
 - Don't skip verification. Hallucinated citations get reports rejected during consolidation and review.
 - Don't exceed 2,500 words; the methodology cares about density, not length.
 - Don't submit without `estimated_tokens`. The MCP tool will reject it.
+- Don't invent evidence for an org that has none (for-profits and orgs with no published outcome data are common). Write the full report documenting what you searched, leave every EVIDENCE TABLE row blank, and call `submit_report` with `no_evidence: true` — an honest, accepted outcome that retires the org from the queue.

package/src/api-client.js

@@ -60,7 +60,7 @@ export class ApiClient {
     return this.request('POST', '/research/claim', { platform });
   }
 
-  async submitReport(claimId, reportMarkdown, tokenUsage = null, metrics = null, modelUsed = null, promptVersion = null, disagreementRows = null) {
+  async submitReport(claimId, reportMarkdown, tokenUsage = null, metrics = null, modelUsed = null, promptVersion = null, disagreementRows = null, noEvidence = false) {
     // The MCP tool surface accepts `estimated_tokens` as a plain number, but
     // the API validates `token_usage` as `nullable|array` and reads
     // `token_usage.total_tokens` for leaderboard accounting. Wrap a bare
@@ -77,6 +77,7 @@ export class ApiClient {
       model_used: modelUsed,
       prompt_version: promptVersion,
       disagreement_rows: disagreementRows,
+      no_evidence: noEvidence || undefined,
     });
   }
 

package/src/mcp-server.js

@@ -145,12 +145,13 @@ server.tool('submit_report', 'Submit a completed research report (or a consolida
   estimated_tokens: z.number().describe('Estimated total tokens used: count web searches (~1K each), web fetches (~2-5K each), report output (~4 tokens/word), plus ~10K overhead'),
   model_used: z.string().optional().describe('The model that generated this report'),
   prompt_version: z.string().optional().describe('Methodology version: "v3" for the EVIDENCE TABLE flow (default), "v2" for the legacy scorecard flow.'),
+  no_evidence: z.boolean().optional().describe('Set true ONLY when thorough research found no qualifying evidence for ANY row (EVIDENCE TABLE fully blank). Records an honest insufficient-evidence finding and retires the org from the queue. Never invent evidence to avoid this.'),
   disagreement_rows: z.array(z.enum(['a1', 'a2', 'a3', 'b', 'c', 'd', 'e', 'f'])).optional().describe('Consolidation-only: EVIDENCE TABLE row keys where the two researchers materially disagreed. >=3 auto-triggers a 3rd researcher.'),
-}, async ({ claim_id, report_markdown, estimated_tokens, model_used, prompt_version, disagreement_rows }) => {
+}, async ({ claim_id, report_markdown, estimated_tokens, model_used, prompt_version, no_evidence, disagreement_rows }) => {
   if (!client) return { content: [{ type: 'text', text: 'Error: TFG_API_KEY not set.' }] };
 
   try {
-    const result = await client.submitReport(claim_id, report_markdown, estimated_tokens, null, model_used, prompt_version ?? `v${METHODOLOGY_VERSION}`, disagreement_rows);
+    const result = await client.submitReport(claim_id, report_markdown, estimated_tokens, null, model_used, prompt_version ?? `v${METHODOLOGY_VERSION}`, disagreement_rows, no_evidence ?? false);
     markContributed();
 
     // One-off users: first successful submit completes their initial setup,

package/src/platform.js

@@ -103,6 +103,8 @@ Say: "Submitting report." Then \`curl\` POST ${base}/research/submit with body {
 
 If the response is a 422 with \`lint_errors\`, your claim is STILL ACTIVE — this is not a failure. Fix the flagged EVIDENCE TABLE rows (usually: a counterfactual row e/f quoting a result instead of the study design, or an adaptation row d quoting tenure instead of a change) and curl submit again with the corrected report.
 
+If thorough research found NO qualifying evidence for ANY row (common for for-profit companies or orgs with no published outcome data), never invent evidence: submit the same full-length report with every table row blank plus \`"no_evidence": true\` in the body. That records an honest insufficient-evidence finding and retires the org from the queue.
+
 Estimate tokens honestly: web searches (~1K each), web fetches (~2–5K each), report output (~4 tokens/word), plus ~10K overhead.
 
 ### 5. Stop

package/src/platform.test.js

@@ -33,6 +33,12 @@ test('schedule prompt stamps the embedded prompt version on submits', () => {
   assert.ok(prompt.includes(`methodology v${METHODOLOGY_VERSION}`));
 });
 
+test('schedule prompt documents the honest no-evidence path', () => {
+  const prompt = getSchedulePrompt('tfg_live_test');
+  assert.ok(prompt.includes('"no_evidence": true'));
+  assert.ok(prompt.includes('never invent evidence'));
+});
+
 test('schedule prompt carries the API key and curl guidance', () => {
   const prompt = getSchedulePrompt('tfg_live_KEY123');
   assert.ok(prompt.includes('tfg_live_KEY123'));