# Changelog

All notable changes to SatQuery AI. Dates are ISO-8601. This changelog records **what changed and what
was verified**, and marks unresolved items honestly rather than omitting them. It follows the same
no-fabrication rule as the rest of the documentation set: an item that is `NOT RUN` stays `NOT RUN`, a
`REJECTED` verdict is never softened, and a negative result is reported as one.

**Status tags:** `RESOLVED` · `OPEN` · `CLOSED` · `DEFERRED` · `REJECTED` · `NOT RUN` · `BLOCKED` ·
`VERIFIED` · `MEASURED`.

**Conventions.** Versions are the project's own release line, not the plan's phase numbers (owner
decision D-5: the master plan's numbering is authoritative for *phases*; this changelog's numbering is
the *release* line). Where an earlier statement in a linked document was later found wrong, the original
is retained and a dated correction is appended — a retained-and-corrected record is more useful than a
silently overwritten one (`docs/OWNER_DECISIONS_2026-09-23.md`, convention).

**Related documents.** [`../MODEL_CARD.md`](../MODEL_CARD.md) · [`MODELS.md`](MODELS.md) ·
[`BENCHMARKS.md`](BENCHMARKS.md) · [`TRAINING.md`](TRAINING.md) · [`EVALUATION.md`](EVALUATION.md) ·
[`LIMITATIONS.md`](LIMITATIONS.md) · [`RESEARCH_NOTES.md`](RESEARCH_NOTES.md) ·
[`ARCHITECTURE.md`](architecture/README.md).

---

## [1.0.0] — 2026-09-25 — Final documentation, model release and archival freeze

The release that makes the project legible, citable and reproducible: repository documentation, a deep
research/engineering documentation set, model documentation, a Hugging Face model release, a curated
public repository, and a verified evidence archive.

### Added

**Repository documentation**

- **Repository README** — 27 sections, from motivation through citation: the deployed topology, the six
  supported tasks, routing, the eight execution events, measured results, live validation, and known
  limitations. **All 20 quoted metrics were verified against their source artifacts** at the README's
  printed precision (`release/tools/readme_metrics_report.txt` → `ALL CLAIMS VERIFIED`).
- **Eight real screenshots** captured from the post-fix live run
  (`release/repo/screenshots/`), each confirmed to show a live run (run id + config hash
  `78f1e3700da15aa1` + specialist evidence), not a mock.

**Deep architecture reference** (hub + 10 chapters)

| Chapter | Topic |
|---|---|
| `docs/architecture/README.md` | hub: thesis, sub-document index, cross-cutting principles |
| `docs/architecture/01-system-overview.md` | thesis, component inventory, frozen-backbone strategy |
| `docs/architecture/02-deployment-topology.md` | four tiers, gateway, tunnel, wake flow, cold start |
| `docs/architecture/03-request-lifecycle.md` | the nine-state controller, validation, tiling |
| `docs/architecture/04-router.md` | frozen MiniLM, the 5-head adapter, `interpret()` vs `chooseTask()`, lexical fallback |
| `docs/architecture/05-specialists.md` | all six tasks in depth |
| `docs/architecture/06-evidence-and-confidence.md` | evidence aggregation, temperature scaling, the eight events |
| `docs/architecture/07-configuration-freeze.md` | the registry, enforced invariants, the config hash |
| `docs/architecture/08-api-contract.md` | the four endpoints, envelopes, error codes |
| `docs/architecture/09-frontend.md` | the static pages, real-vs-preview, platform traps |
| `docs/architecture/10-observability-and-ops.md` | health, counters, traces, what is not observed |

**Research documentation** — `MODELS.md`, `BENCHMARKS.md`, `DATASETS.md`, `TRAINING.md`,
`EVALUATION.md`, `DEPLOYMENT.md`, `REPRODUCIBILITY.md`, `LIMITATIONS.md`, `RESEARCH_NOTES.md`,
`CHANGELOG.md`.

**Extended engineering-documentation set** (added in the same release, deep-reference depth):

`SECURITY.md`, `TESTING.md`, `GEOSPATIAL.md`, `DATA_PIPELINE.md`, `OPERATIONS.md`, `DEVELOPMENT.md`,
`PERFORMANCE.md`, `GLOSSARY.md`, `FRONTEND.md`, `SERVING.md`. See the
[complete documentation inventory](#documentation-inventory-shipped-in-100) below.

**Model documentation**

- **`MODEL_CARD.md`** — the Hugging Face model card, expanded to deep-reference depth in this pass: an
  overview; the six-artifact table with byte counts and full sha256; backbone dependencies with pinned
  revisions; intended and out-of-scope use; a per-artifact section (architecture, hyperparameters,
  training data, evaluation protocol, measured numbers, acceptance status, limitations) for each of the
  six artifacts; the full measured-performance table; the calibration negative result; the
  acceptance-status table; the evaluation gaps; the limitations; a training summary; provenance and
  verification; licence; and citation. Retains its YAML frontmatter (`license: other`, tags,
  `pipeline_tag: image-to-text`, `config_hash: 78f1e3700da15aa1`).
- **`models/manifest.json`** and **`models/checksums.sha256`** — **generated by reading the files**
  (`release/tools/generate_model_manifest.py`), never typed by hand. Six artifacts, each with
  `status: "PRESENT"` and the frozen config hash. Two digests (`change_vqa`, `vlm`) **agree exactly**
  with values recorded independently at promotion/fit time — an external cross-check, not a
  self-consistency claim.

**Metric verification tooling**

- **`release/tools/verify_readme_metrics.py`** — walks every quoted metric to its source artifact and
  the exact key path inside it. Result: **`ALL CLAIMS VERIFIED`** (20/20), plus status assertions (VLM
  headline contains `ACCEPTANCE-REJECTED`; router `corpus_limited` with `n_val` 86; calibration
  `ece_improvement` negative). Output committed as `release/tools/readme_metrics_report.txt`.

**Release**

- **Hugging Face release** of the six trained artifacts to `thundercode/SatQuery` (public), with a
  verification record ([`../HF_RELEASE_VERIFICATION.md`](../HF_RELEASE_VERIFICATION.md)). Independent
  re-download over direct HTTPS confirmed **6/6 artifacts byte-identical**, 0 failed. Total released
  weight payload **131,130,325 bytes (~125 MiB)**; **no backbone weights are redistributed**.
- **Curated public repository** target `Anish-lab-blip/SatQuery-AI`, with
  [`../RELEASE_MANIFEST.md`](../RELEASE_MANIFEST.md) (42 files, 8,401,031 bytes, each with size and
  sha256, generated from disk).
- **Evidence archive** — scope set by owner decision G3 (`artifacts/` verbatim, ~3.7 GB, requiring
  ZIP64), with an integrity-verified extraction and a documented remote location (owner decision G2:
  Hugging Face Hub).

**This pass (documentation depth)**

- **`MODEL_CARD.md` expanded in place** to deep-reference depth (see Model documentation above).
- **`CHANGELOG.md` expanded in place** — the full 1.0.0 entry, the earlier version entries (0.9.0 →
  0.1.0), an Unreleased/Open section, and the **complete documentation inventory** below.

### Changed

- The monorepo `README.md` was **materially stale** and is superseded by this release's README. The
  stale version called the frontend "hermetic — no backend calls", put Render/Codespace as
  "in progress", described a `/v1/*` contract, and pointed at the untracked `deploy/` as the deployment
  source. **All four were wrong.**
- `hf/` documentation was **stale** — it asserted the project "does not own any model weights … ships no
  weights, no binaries, and no large artifacts" and "has no Hugging Face credentials". **Both were false
  at release time** and are corrected.
- **Documentation set expanded.** Ten new deep-reference docs were added after `RELEASE_MANIFEST.md` and
  the Hugging Face upload were generated. **Consequence, stated plainly:** the release manifest (42
  files) and the Hub's 26 markdown files **do not yet include** those ten documents. The manifest and the
  Hub listing are therefore **stale relative to the local release tree** and must be regenerated /
  re-uploaded before they can be treated as authoritative. See
  [the inventory](#documentation-inventory-shipped-in-100).

### Fixed

- **Router defect** — *"Where are the built-up areas in this image?"* collapsed to `vqa` and answered
  "River". Fixed in the dispatch path and validated across three independent live passes (8/8 each).
  `RESOLVED`. (Details in 0.9.0.)
- **Live-validation harness false positive** — synthetic CDP key events were silently dropped when the
  window lacked OS focus, so the harness dispatched the page's default query and recorded a false pass.
  Fixed with pre-dispatch form-state assertions (`q_ok`, `obs_ok`, `t0_ok`). `RESOLVED`.
- **Bundled EO change pair** — the demo pair differed in shape (726² vs 736²) and errored
  (`T1 and T2 must have the same shape`). Replaced with a same-shape pair (both 720×720). `RESOLVED`
  (supersedes `docs/FINAL_DELIVERY_REPORT.md` §6).
- **B-01 (Hugging Face header link)** — previously `BLOCKED`; the link is now live on all 11 pages.
  `CLOSED`.
- **Metric-verifier dotted-path bug** — the resolver split the artifact key `"0.50"` on its own dot,
  producing a false `NOKEY` for the two grounding recall claims. Fixed to prefer the **longest**
  matching key at each step. `RESOLVED`.

### Known issues at this release

- **B-07** — transient tunnel-agent gaps; a request can hang or return 504. Patch prepared, **NOT
  deployed**. `OPEN`.
- **B-02** — `/api/health` `codespace_name` carries a trailing `\n`. Cosmetic; the wake path strips it.
  `OPEN`.
- **No `LICENSE` file** exists in the source repository. `OPEN`.
- **No system-level end-to-end benchmark** exists. `NOT RUN`.
- **Router test split** never scored. `NOT RUN`.
- **Calibration** worsened ECE (0.013755 → 0.014929). Retained as part of the frozen config only.
- **VLM adapter** is `ACCEPTANCE-REJECTED` despite usable metrics (exact_match 0.963, F1 0.96432); the
  deployed path uses the unadapted model.
- **optical-SAR** and **change-VQA** rulings remain `OPEN`.
- **GitHub push** — at the time of `release/RELEASE_EXECUTION_CHECKLIST.md` Phase 5 the push to
  `Anish-lab-blip/SatQuery-AI` was `BLOCKED` (`403 Resource not accessible by personal access token`;
  the token was read-only for repository contents). Owner action.
- **Release manifest / Hub doc set** stale relative to the ten new deep docs (see *Changed*).

### Deployment revisions at this release

| Component | Revision |
|---|---|
| Frontend | `2d7ae53b482d` |
| Backend | `89d80eaddec5` |
| Inference | `5a0936ace491` |
| Public umbrella | `3dcabd32da41` |

**Live validation.** 3 passes × 8 cases, **8/8 each**, **24 live runs, 24 correct dispatches, 0 mock
nodes**, trace fill **94.4444 %**. Representative real run ids: vqa `run_ddf81a1c3871`, caption
`run_15d7b9ecf594`, grounding `run_d124d8b9adea` (3 regions, calibrated 0.52), change
`run_c6249437253d`, optical_sar `run_028b200fb710` (`docs/FINAL_DELIVERY_REPORT.md` §4).

### The six released artifacts at this release

Generated identities from [`../models/manifest.json`](../models/manifest.json); see
[`../MODEL_CARD.md`](../MODEL_CARD.md) §3 for the full table with source paths and cross-checks.

| Task | Kind | File | Bytes | sha256 (full) |
|---|---|---|---|---|
| `change` | trained head | `change/head.pt` | 63,231,009 | `c5ef31277b67aa01a593aec0eac503eeaccc6d674349fda20ca44c9cc6f8e9fa` |
| `change_vqa` | trained head | `change_vqa/head.pt` | 5,822,809 | `cfae5e43b97ca930f568dc5b8ae4f36b24e9ff717af226159802206ffd63a82a` |
| `optical_sar` | trained head | `optical_sar/head.pt` | 14,427,457 | `785815729a3a39fc34dc41894efaf00d8739365d970a3f830a326e68ae888dab` |
| `grounding` | trained head | `grounding/head.pt` | 12,639,041 | `93432f7034be91a8ffd9c1a84e3eeec00bed7832c043fe7f83d2be230284c6bb` |
| `router` | trained adapter | `router/adapter.pt` | 211,961 | `8527c3ed28a293e13293d48601d48e3ceafa137b9acabddaf5de31a58a509b5c` |
| `vlm` | LoRA adapter | `vlm/adapter_model.safetensors` | 34,798,048 | `07c76a75fa04624880ed7730590f5fdd7b145a8232e3c0af411c3c545a5adf5e` |

Every digest above is copied from [`../models/manifest.json`](../models/manifest.json); consult the
manifest, which wins over any hand-typed copy.

Total released weight payload: **131,130,325 bytes (~125 MiB)**. No backbone weights are redistributed.

### Measured results at this release

Every row names its source artifact; all 20 numeric claims are checked by
`release/tools/verify_readme_metrics.py` (`ALL CLAIMS VERIFIED`). The full annotated version is
[`../MODEL_CARD.md`](../MODEL_CARD.md) §8 and [`BENCHMARKS.md`](BENCHMARKS.md) §1.

| Capability | Metric | Value | Split / protocol | n | Status |
|---|---|---|---|---|---|
| Change | pooled IoU / macro IoU / pooled F1 | **0.8122 / 0.8457 / 0.8964** | LEVIR-CD-256 test, thr 0.50 | 2,048 | **VERIFIED** |
| Grounding | mean best IoU / recall@0.5 | **0.2838 / 0.2198** (canonical); **0.2566 / 0.1938** (matched6) | VRSBench | 16,159 | MEASURED (2 protocols) |
| Grounding | head-argmax / zero-shot IoU | **0.1215 / 0.0972** | canonical | 16,159 | MEASURED |
| Optical-SAR | accuracy / macro-F1 | **0.931 / 0.434161** | held-out test, 19 classes | 4,000 | MEASURED, ruling **OPEN** |
| Change-VQA | accuracy / macro-F1 | **0.697626 / 0.378373** (test); **0.651469 / 0.372309** (test2) | two test sets | 39,686 / 31,036 | MEASURED, ruling **OPEN** |
| VLM (adapted) | exact_match / F1 | **0.963 / 0.96432** | frozen 1,000-question subset | 1,000 | **ACCEPTANCE-REJECTED** |
| Router | overall **ungated** accuracy | **0.965116** | val, corpus-limited | 86 | MEASURED — **TEST NOT RUN** |
| Calibration | ECE before → after | **0.013755 → 0.014929** | val, T = 0.9773 | 16,441 | MEASURED — **worse** |
| System | end-to-end accuracy | — | — | — | **NOT RUN — none exists** |

### Status of every claim at this release

The facts most likely to be misstated, and their correct form. These are the same statements the release
README, the model card and the benchmark doc carry; they are repeated here so a reader of the changelog
alone cannot misread the release.

| Fact | Correct statement | Never say |
|---|---|---|
| Grounding | two protocols **and** two decode variants: canonical 0.2838/0.2198, matched6 0.2566/0.1938, head_argmax 0.1215, zero-shot 0.0972 | "grounding is 0.28" |
| Calibration | ECE **worsened** 0.013755 → 0.014929; retained only because it is in the frozen config | "calibration improved confidence" |
| VLM adapter | metrics usable, status **ACCEPTANCE-REJECTED**; deployed path uses the unadapted model | "the VLM is accepted" / "the VLM is broken" |
| Optical-SAR | accuracy **0.931 always with macro-F1 0.434161**; ruling `OPEN`; single-label subset only | "0.931 accuracy" alone |
| Change-VQA | **two** test sets, both quoted; ruling `OPEN` | "change-VQA is 0.70" |
| Router | **0.965116 is validation, ungated, n = 86**; test `NOT RUN` | "router test accuracy" |
| End-to-end | no system-level benchmark exists; none is claimed | any end-to-end accuracy |
| Change | pooled IoU 0.8122 / macro IoU 0.8457 / pooled F1 0.8964 — the only `VERIFIED` headline | treating it as system-level |

### Verification and reproduction at this release

```bash
# 1. Verify every quoted metric against its source artifact  (cheap, no GPU)
python release/tools/verify_readme_metrics.py
#    -> STATUS ... MATCH ... / RESULT: ALL CLAIMS VERIFIED

# 2. Verify the released artifact identities
sha256sum -c models/checksums.sha256
#    -> change/head.pt: OK  ... vlm/adapter_model.safetensors: OK

# 3. Re-download and re-hash the Hugging Face release (independent of the upload step)
python release/tools/hf_verify.py
#    -> artifacts verified : 6 / artifacts failed : 0

# 4. Regenerate the artifact manifest from disk (must not change any value)
python release/tools/generate_model_manifest.py
```

`release/tools/readme_metrics_report.txt` and `release/tools/hf_verify_report.txt` are the committed
outputs of steps 1 and 3.

### Compatibility and migration notes

- **Frozen config hash.** `78f1e3700da15aa1` is the sha256 of `configs/base.yaml` truncated to 16
  characters. Every artifact records it. Editing the registry moves the hash and detaches every
  published number from its configuration — new defaults therefore live in **code** (e.g.
  `DEFAULT_HEAD_PATH`), not in `base.yaml` (owner decision D-4). A test
  (`test_the_frozen_config_hash_has_not_moved`) pins this.
- **Grounding default semantics changed.** `head_path = None` **no longer means "no head"** — it now
  means *"use the shipped head if present."* An explicit `head_path` still wins. Any caller that relied
  on `None` meaning "disable the head" must now pass a path that does not resolve (D-4).
- **VLM adapter opt-in.** The Phase 6 adapter is attached only when the `SATQUERY_VLM_ADAPTER`
  environment variable (or an explicit `adapter_path` argument) is set. It is **not** the deployed
  default.
- **`optical_sar` capability reporting.** `/api/capabilities` keeps `available = false` for
  `optical_sar` even though the capability is locally executable — the under-report is a statement about
  *deployment certification*, not about code, and two tests pin it (owner decision D-7).
- **Phase numbering.** The master plan's numbering is authoritative for phases; the repository's
  historical numbering is retained as history with a mapping (`docs/PHASE_NUMBERING_MAP.md`). This
  changelog's version numbers are the *release* line, independent of both (D-5).

---

## [0.9.0] — 2026-09-25 — Router-defect sprint

### The defect

Two functions with different information:

- **`interpret()`** — produces the console's *reading*; **asset-count-blind** (text only).
- **`chooseTask()`** — performs *dispatch*; **asset-count-aware**.

*"Where are the built-up areas in this image?"* collapsed to **`vqa`** and answered **"River"** instead
of routing to `grounding`; *"Where is the new airport?"* behaved the same way. The defect was in the
dispatch path's handling of spatial/lexical cues, so region queries fell through to the generic VQA
specialist.

### The fix

- The temporal slot in `interpret()` now derives from `changeStem` / `pairWording` / `newAsChange`.
- `built` was dropped as a change marker.
- The quantifier `area` was tightened to `\barea\b`.

### Verification

Three commits deployed to the frontend; **3 independent live validation passes**, 8/8 each — **24 live
runs, 24 correct dispatches, 0 mock nodes, trace fill 94.4444 %**.

| Pass | Deployed HEAD | Result |
|---|---|---|
| 1 | `ff46eba42b18` + `d413d3672311` | 8/8 |
| 2 | `2d7ae53b482d` | 8/8 |
| 3 | `2d7ae53b482d` | 8/8 |

Both defect queries now dispatch to `grounding`:

| Query | Run id | Dispatched |
|---|---|---|
| Where are the built-up areas in this image? | `run_467ffa406f22` | `grounding` |
| Where is the new airport? | `run_46980ba55c62` | `grounding` |

### Harness false positive — caught before it could lie

An earlier live-validation harness typed queries with **synthetic CDP key events**, which Chrome
**silently drops when the window lacks OS focus**. The harness therefore dispatched the page's *default*
query and still recorded a "result" — a **false pass**. Fixed by asserting form state before dispatch
(`q_ok`, `obs_ok`, `t0_ok`) and using deterministic query entry (a `js()` value-set plus `type_text()`
via CDP `Input.insertText`). The earlier 8/8 run was re-examined and confirmed **not** infected — its
answers were query-specific and the query text was embedded in the answers.

### Also in this sprint

- Wrote the delivery report and the handoff.
- Prepared (but did **not** deploy) the B-07 backend patch: it adds `forward_unavailable` (503) and
  `upstream_timeout` (504) codes plus the `codespace_name` `.strip()` fix; verified `py_compile` clean.
  **Status: `OPEN` — prepared, NOT deployed.**

`RESOLVED` for the router defect and the harness bug. Root cause and lessons recorded in
[`RESEARCH_NOTES.md`](RESEARCH_NOTES.md) §3–§4.

---

## [0.8.0] — 2026-09-24/25 — Deployment hardening

### Topology settled

```
Browser → Cloudflare Pages (satquery.pages.dev)
        → Render hub (satquery-orchestrator)
        → outbound TUNNEL (/tunnel/agent long-poll)
        → GitHub Codespace (FastAPI inference, CPU)
        → models (change · change_vqa · optical_sar fusion · grounding · VLM/CROMA)
        → real ResultEnvelope → Render → browser
```

The tunnel exists because a forwarded Codespace port on a **private** repo returns `302`. Deploy sources
are three private repositories; the local `deploy/` is stale/untracked and is **not** the deployed
source.

### The five historical backend blockers — closed

1. ASGI stack present in `requirements`;
2. `$PORT` read;
3. `OPTIONS` preflight handled;
4. fail-fast app construction;
5. adapter integrity compared against an expected digest.

### Live topology verification (probed, not asserted)

| Probe | Result |
|---|---|
| `GET /api/health` | 200, `agent_connected: true` |
| `GET /api/capabilities` | 200, six tasks each `available: true` |
| `POST /api/infer {}` | 422 `invalid_request` |
| response header | `x-satquery-transport: tunnel` |

### Behaviour recorded (not fixed, because it is not a defect)

- **Cold start** is tens of seconds; the Render free tier sleeps and the Codespace may be stopped.
- **`transport_mode: auto` fallthrough** — in `auto` mode a tunnel timeout falls through to the forward
  path, burning `wake_timeout_s` (120 s) on a `302`; worst case ≈ **249 s** (150 + 120). Measured. This
  is the root shape of B-07 (`OPEN`).

`RESOLVED` for the five blockers; the `auto` fallthrough is documented and carried forward.

---

## [0.7.0] — Grounding and change heads

### Grounding — resolution frozen at 224

Trainable head over frozen RemoteCLIP ViT-B/32. **224 chosen, 448 REJECTED** by a pre-registered paired
test over all 16,159 VRSBench eval records: mean best IoU −0.0147, recall@0.5 −0.0022, recall@0.10
−0.0699, recall@0.25 −0.0243, at **1.59×** the latency; paired 95 % CI [−0.0160, −0.0134],
**t = −22.63**; 448 better on 8.5 %, worse on **20.9 %** of records. The pre-registered rule and the
paired test **agree** on 224, and the artifact records
`rule_changed_since_preregistration: false`. `RESOLVED`. Evidence:
`docs/PHASE7_RESOLUTION_DECISION.md`.

Head design: per-cell feature `concat([patch, text, patch*text, global_pool]) = 4 × 512 = 2048`;
cell-relative (YOLO-style) boxes; exactly one positive cell per target (the one containing the
ground-truth box centre); `positive_confidence_weight: 20.0` to stop the 1-of-49 objectness collapse.

### Change — the only VERIFIED headline

STANet-style Siamese detector on LEVIR-CD-256 (`train 7,120 / val 1,024 / test 2,048`), ResNet-18 + PAM,
tile 256, threshold 0.50. Measured on the held-out test split (n = 2048): **pooled IoU 0.8122 / macro
IoU 0.8457 / pooled F1 0.8964** — the only headline in the project carrying the `VERIFIED` tag.

### Optical-SAR — measured, ruling OPEN

CROMA-base fusion head (`2318 → 512 → 19`), mandatory channel/band dropout, availability mask consumed
by the head (not CROMA). Sweep: **10 runs × 2 arms × seeds 100–104**. Arm decision on
`best_val_accuracy`: A 0.837100 vs B 0.839100, floor 0.0285 → **Arm A retained** (owner ruling R-14).
Pre-registered 115-class metric on the held-out split (n = 4,000): **accuracy 0.931 / macro-F1
0.434161**, ruling `OPEN`. The 224 localisation floor for zero-shot grounding is **1/7 of image width
per token**, documented as a known limitation.

---

## [0.6.0] — VLM adaptation

### Training

SmolVLM-500M LoRA adapter (PEFT, `r = 16`, `alpha = 32`, `dropout = 0.05`, fp16; **224** target modules
across `q/k/v/o/gate/up/down_proj`, all inside `model.text_model`; **8,683,520** trainable params =
**1.6823 %** of the 516,165,824-param base; seed 42). The **vision tower was left untouched**
(`trainable_subtrees = {"model.text_model": 8683520}`; the 86,433,024-param vision model and the
11,796,480-param connector are in `frozen_params`).

### Metrics usable, artifact REJECTED

Adapted exact_match **0.963**, F1 **0.96432**, aggregate test delta **+49.50 pp** (46.80 → 96.30). But
the pre-registered rule **v002** passes V1 and fails its **V2** per-class guardrail on the independent
test split: class *Mixed forest* (`n = 33`) drops **12.1212 pp** with `lost_questions 4`, `z 2.1335` —
failing **both** halves of v002. Per item V, a complete run that fails V2 is `REJECTED`.

Both rejection records are preserved (Run 1's own manifest under v001 on val; the test-split
adjudication under v002 on test); the val-split `ACCEPTED` under v002 is recorded **for completeness
only** and is not final acceptance — it decides on the same val subset that motivated v002.

### Findings

- **F5-1** — `AutoModelForVision2Seq` **does not exist** in transformers 5.17.0 (not merely deprecated);
  the loader class is resolved by feature detection over `("AutoModelForImageTextToText",
  "AutoModelForVision2Seq", "AutoModelForMultimodalLM")`.
- **F5-2** — the processor's default `longest_edge` is 2048, which upscales a 512-px tile 4× and then
  splits it into **17** sub-images and 1142 prompt tokens. The plan estimated a 4× cost overrun; the real
  figure is **~17×**. `processor_longest_edge` is pinned to 512 and enforced at config load.
- **F5-3** — SmolVLM requires one `<image>` token per image; hand-written prompt strings raise
  `ValueError`, so prompts are always built through `processor.apply_chat_template()`.

### Closure

Phase 6 closed **without retraining or modifying** the adapter. Evidence:
`artifacts/vlm/phase6_closure.json`, `docs/PHASE6_CLOSURE.md`. Two traps recorded: `adapter_sha256` names
a **tree hash** (`5c6b8631…`) and a **file** sha256 (`07c76a75…`), which are not interchangeable; and the
promoted adapter is **not** `checkpoint-2000`.

---

## [0.5.0] — Router

### Architecture

A 50,822-parameter five-head adapter over frozen MiniLM embeddings (`router/adapter.py`): `LayerNorm →
Linear(384 → 128) → GELU → Dropout(0.10)`, then `task_head (→6)`, `modality_head (→4)` and three binary
heads `temporal`/`spatial`/`language` (each `→1` logit). Heads use small-std init (`std 0.02`, zero
bias) so the initial sigmoid sits near 0.5. The adapter does **not** back-propagate into MiniLM.

### Training

Encoder frozen ⇒ embeddings cached ⇒ **CPU-only training**. Measured: **20 epochs over 4,096 × 384
vectors in 0.28 s**. Config: 60 epochs, batch 64, lr 0.001, weight decay 0.01; loss weights task 1.0 /
modality 0.3 / binary 0.5; `val_ratio` 0.15; `hard_negatives_to_test: true`.

### Data and splits

A synthetic query corpus: **576** queries in **54** groups, split `train 410 / val 86 / test 80`. Splits
are **by group** (template / hard-negative family), never by example, so template variants cannot leak
across the boundary (finding F4-3). Hard-negative families are placed in the **test** split so their
accuracy measures generalisation, not memorisation.

### Validation-only threshold sweep

50 thresholds `0.50 … 0.99`, `select_by: covered_accuracy`. Shipped threshold **0.70** (coverage
0.848837, covered accuracy 0.972603, fallback 0.151163, n_covered 73); the sweep's selected row is
threshold 0.76 (coverage 0.790698, covered accuracy 1.0, n_covered 68). Overall **ungated** accuracy
**0.965116** on **val n = 86** — corpus-limited (plan minimum 500 val queries, 100 hard negatives; val
carries 0 hard negatives by design). **Test split NOT RUN** (`test_split_touched: false`).

### Deterministic fallback

A lexical fallback with ordered precedence and a `self_check()` against curated hard negatives, so the
router degrades to a deterministic rule rather than to an arbitrary class when the adapter is
low-confidence.

---

## [0.1.0] — Foundation

- Monorepo; frozen configuration registry (`configs/base.yaml`); **config-hash freeze**
  (`78f1e3700da15aa1`); **nine-state controller** (`RECEIVE → PARSE → VALIDATE → PLAN → PREPROCESS →
  EXECUTE → AGGREGATE → VERIFY → RESPOND`); **four-endpoint** inference contract; evidence schema
  (`Region` / `ChangeRegion` / `Evidence`, `max_items` 32, coordinate system `normalized_0_1`); **eight
  frontend execution events**.
- Enforced invariants established at load time: fusion width `3 × 768 + 12 + 2 = 2318`; grounding head
  feature `4 × 512 = 2048`; CROMA `image_resolution % 8 == 0` (native 120 → 225 patches); router
  `max_length ≤ 256`; VLM `processor_longest_edge ≤ image.tile_size`; change tile 256; VLM
  `prompt_must_use_chat_template: true`.
- Input geometry and normalisation frozen: `image.max_pixels` 25,000,000; `tile_size` 512;
  `tile_overlap` 128; `max_tiles` 64; `top_k_tiles` 4. Optical percentile 2–98; SAR dB clip [−30, 5].
- The rule that shaped every later decision: **"no magic numbers anywhere in Python; everything tunable
  lives here"** — and, from the outset, the no-fabrication rule that a missing dependency, corpus or
  checkpoint produces a diagnosable error, never a plausible number.

---

## Unreleased / open

| Item | State |
|---|---|
| Deploy the B-07 tunnel-fallthrough fix | prepared, **NOT deployed** |
| Score the router test split | **NOT RUN** |
| Build a system-level end-to-end benchmark | **NOT RUN** |
| Select a licence | **OPEN** |
| Resolve the optical-SAR and change-VQA rulings | **OPEN** |
| Plot the calibrated reliability curve | **NOT RUN** |
| Replace the illustrative benchmark adapter numbers with real runs | **NOT RUN** |
| Regenerate `RELEASE_MANIFEST.md` and re-upload the ten new deep docs to the Hub | **NOT RUN** |
| Push the curated release to `Anish-lab-blip/SatQuery-AI` (token scope) | **BLOCKED** — owner action |
| Fix B-02 (`codespace_name` trailing newline) | **OPEN** (cosmetic) |
| Cross-dataset, human, robustness and adversarial evaluation | **NOT RUN** |

---

## Documentation inventory shipped in 1.0.0

**Every documentation file in the release tree.** Sizes are the measured byte counts in
`release/repo/`, as of the 2026-09-26 documentation pass. The **"in manifest / Hub"** column records
whether the file was included in `RELEASE_MANIFEST.md` and the Hugging Face upload — the ten newest deep
docs were written **after** both were generated, so they are present locally but **not yet** in the
manifest or on the Hub.

### Root-level documents

| File | Purpose | Bytes | In manifest / Hub |
|---|---|---|---|
| `README.md` | repository front page (27 sections) | 86,336 | yes |
| `MODEL_CARD.md` | model card for the six trained artifacts (Hugging Face card) | 57,166 | yes (previous revision; **expanded in this pass**) |
| `RELEASE_MANIFEST.md` | every released file with size + sha256 | 10,402 | yes |
| `HF_RELEASE_VERIFICATION.md` | independent HF re-download verification (6/6 MATCH) | 6,443 | yes |

### Research and engineering documents (`docs/`)

| File | Purpose | Bytes | In manifest / Hub |
|---|---|---|---|
| `docs/architecture/README.md` | architecture hub: thesis, sub-document index, cross-cutting principles | 7,672 | yes |
| `docs/MODELS.md` | per-model detail: hyperparameters, acceptance status, measured params | 103,145 | yes |
| `docs/BENCHMARKS.md` | every measured metric with its artifact and protocol | 55,999 | yes |
| `docs/DATASETS.md` | dataset provenance, splits, leakage controls | 67,723 | yes |
| `docs/TRAINING.md` | training procedures and hyperparameters | 90,697 | yes |
| `docs/EVALUATION.md` | evaluation protocols and honesty rules | 111,672 | yes |
| `docs/DEPLOYMENT.md` | live topology, env vars, mechanics, traps | 16,041 | yes |
| `docs/REPRODUCIBILITY.md` | what can and cannot be reproduced | 65,200 | yes |
| `docs/LIMITATIONS.md` | the full limitation catalogue | 8,356 | yes |
| `docs/RESEARCH_NOTES.md` | engineering findings and negative results | 11,513 | yes |
| `docs/CHANGELOG.md` | this file | *(this file)* | yes (previous revision; **expanded in this pass**) |
| `docs/MASTER_ARCHITECTURE_PLAN.md` | the original master plan, included as a historical document | 65,304 | yes |
| `docs/SECURITY.md` | security posture, trust boundaries, rate limiting, secrets hygiene | 114,342 | **no — new** |
| `docs/TESTING.md` | the test suites, guards, and what is not covered | 96,897 | **no — new** |
| `docs/GEOSPATIAL.md` | CRS, transforms, alignment, coordinate systems | 92,888 | **no — new** |
| `docs/DATA_PIPELINE.md` | ingestion, extraction, feature caches, split policy | 66,285 | **no — new** |
| `docs/OPERATIONS.md` | running the system: health, counters, traces, runbooks | 62,277 | **no — new** |
| `docs/DEVELOPMENT.md` | local development, invariants, contribution mechanics | 69,367 | **no — new** |
| `docs/PERFORMANCE.md` | latency, memory, throughput, cost notes | 52,946 | **no — new** |
| `docs/GLOSSARY.md` | terms, abbreviations, status vocabulary | 40,998 | **no — new** |
| `docs/FRONTEND.md` | the static frontend: pages, real-vs-preview, platform traps | 73,926 | **no — new** |
| `docs/SERVING.md` | the serving path: endpoints, envelopes, error codes, wiring | 75,797 | **no — new** |

### Architecture reference (`docs/architecture/`)

| File | Topic | Bytes | In manifest / Hub |
|---|---|---|---|
| `docs/architecture/01-system-overview.md` | thesis, component inventory, frozen-backbone strategy | 8,376 | yes |
| `docs/architecture/02-deployment-topology.md` | four tiers, gateway, tunnel, wake flow, cold start | 85,979 | yes |
| `docs/architecture/03-request-lifecycle.md` | the nine-state controller, validation, tiling | 130,175 | yes |
| `docs/architecture/04-router.md` | frozen MiniLM, the 5-head adapter, `interpret()` vs `chooseTask()` | 262,652 | yes |
| `docs/architecture/05-specialists.md` | all six tasks in depth | 252,171 | yes |
| `docs/architecture/06-evidence-and-confidence.md` | evidence aggregation, temperature scaling, the eight events | 108,044 | yes |
| `docs/architecture/07-configuration-freeze.md` | the registry, enforced invariants, the config hash | 106,023 | yes |
| `docs/architecture/08-api-contract.md` | the four endpoints, envelopes, error codes | 127,341 | yes |
| `docs/architecture/09-frontend.md` | the static pages, real-vs-preview, platform traps | 78,287 | yes |
| `docs/architecture/10-observability-and-ops.md` | health, counters, traces, what is not observed | 104,695 | yes |

### Generated artifacts and tooling

| File | Purpose | Bytes | In manifest / Hub |
|---|---|---|---|
| `models/manifest.json` | generated manifest of the six trained artifacts (bytes + sha256) | 4,458 | yes |
| `models/checksums.sha256` | generated checksums, keyed by repository path | 642 | yes |
| `tools/verify_readme_metrics.py` | walks every quoted metric to its source artifact | 6,911 | yes |
| `tools/readme_metrics_report.txt` | its output (`ALL CLAIMS VERIFIED`) | 4,027 | yes |
| `tools/generate_model_manifest.py` | generates `models/manifest.json` + `checksums.sha256` from disk | 6,868 | yes |
| `tools/generate_release_manifest.py` | generates `RELEASE_MANIFEST.md` from disk | 4,100 | yes |
| `tools/hf_verify.py` | re-downloads and re-hashes the released artifacts | 3,878 | yes |
| `tools/verify_archive.py` | verifies the evidence archive extraction | 4,946 | yes |
| `tools/build_archive.py` | builds the ZIP64 evidence archive | 8,351 | yes |

**Inventory totals.** 4 root-level documents · 22 documents in `docs/` · 10 architecture chapters ·
9 generated-artifact / tooling files. **45 files.** The ten newest deep docs (`SECURITY.md`,
`TESTING.md`, `GEOSPATIAL.md`, `DATA_PIPELINE.md`, `OPERATIONS.md`, `DEVELOPMENT.md`, `PERFORMANCE.md`,
`GLOSSARY.md`, `FRONTEND.md`, `SERVING.md`) are the only ones not yet reflected in
`RELEASE_MANIFEST.md` / the Hub; regenerating the manifest and re-uploading is an `Unreleased` item
above.

---

## Where the evidence lives

| Topic | Evidence |
|---|---|
| All measured metrics | `artifacts/**/*.json`, verified by `release/tools/verify_readme_metrics.py` |
| Release identity and byte counts | `models/manifest.json`, `models/checksums.sha256`, `RELEASE_MANIFEST.md` |
| Hugging Face release integrity | [`../HF_RELEASE_VERIFICATION.md`](../HF_RELEASE_VERIFICATION.md), `release/tools/hf_verify_report.txt` |
| Router defect + 3 live passes | [`RESEARCH_NOTES.md`](RESEARCH_NOTES.md) §3–§4 |
| Grounding resolution rejection | `docs/PHASE7_RESOLUTION_DECISION.md` |
| VLM rejection / Phase 6 closure | `artifacts/vlm/phase6_closure.json`, `docs/PHASE6_CLOSURE.md` |
| Optical-SAR metric | `docs/PHASE12_115_METRIC_COMPUTED.md`, `artifacts/optical_sar/fusion_head_production_v001/pre_registered_115_metric.json` |
| Owner decisions | `docs/OWNER_DECISIONS_2026-09-23.md` |
| Documentation reconciliation | `docs/DOCUMENTATION_RECONCILIATION.md` |
| Release execution state | `release/RELEASE_EXECUTION_CHECKLIST.md`, `release/CURRENT_RELEASE_STATE.md` |